Skip to content

A Civilization-style strategy game about governing Israel from 15 May 1948. Real ministers, real map, every fact sourced. Unity + deterministic C# sim.

331 KB
Raw Blame

Visual v1: measurements, budgets and checkpoints

Plan: docs/superpowers/plans/2026-10-11-statecraft-visual-v1.md. Promo targets: promo/final/03-gameplay-1936.png, promo/final/02-map-diorama.png. Poses: PromoPose.Gameplay03 (34.86 E, 32.00 N, 60 km, yaw 10) and PromoPose.Diorama02 (35.05 E, 31.55 N, 250 km, yaw 0). Since #174 the camera keeps its view on the map, and centres it east to west wherever the view is wider than the map. In the player, which runs that fit every frame, diorama02 can shift by up to a few hundred metres from its pose. The controller accepted this (owner-delegated). The other poses do not move.

Art direction (owner ruling, 14 October 2026)

The owner ruled "this is the art direction" on four references (Slack, #visual-design-art-techart, 14 October 2026):

  • demo2 shot T29, "26 May 1948: one army": the coast-to-Jordan diorama, units on white ring bases, Arab armour with flags (promo/demo2/out/demo2-final-16x9.mp4, about 1:11);
  • demo2 shot L14, "Latrun is captured by the enemy": a low-angle close-up of the fortress (same file, about 0:32);
  • demo2 shot D15, "20 May: Degania stops the Syrian tanks": a close-up at the kibbutz gate (same file, about 0:36);
  • promo/final/03-gameplay-1936.png.

The shot ids are those of promo/demo2/SHOTLIST-v2.md and shots-v2.py; demo2's files are local to the promo work and not in the repository. Two rules from that thread: lock the visual language before more models are made ("it's a question of visual language, not quality"), and the next step is a vertical slice of a few terrain hexes with the models on them. New models and visual PRs are judged against these four references.

How to measure

  • Frame time: ?vprofile (WebGL) or -vprofile (desktop player, quits when done), always under uv run tools/machine_lock.py --exclusive --wait-unity -- ... (no other Unity job on the machine). Per pose, every configuration (all, each toggle off alone as -<name>, none) runs 3 times in interleaved orders, 300 frames after 60 warm-up frames (and, before the first pose, one discarded warm-up pass of every configuration); Chrome runs with --disable-frame-rate-limit --disable-gpu-vsync and the anti-throttling flags (--disable-backgrounding-occluded-windows --disable-renderer-backgrounding --disable-background-timer-throttling), headless through web_check.py (add --stall 120), headed through headed.py profile. Both fix the page's viewport to exactly 1920x1080 CSS pixels at the DPR over CDP (Emulation.setDeviceMetricsOverride, before the page loads), so headed and headless runs draw the same pixels even when the laptop screen clamps the headed window (its final composite is then scaled; the backbuffer is not). The runtime quality fallback (issue #44, "Runtime quality fallback") is off under ?vprofile, ?profile and ?shot, so these see the tier's own scale and tilt-shift. Every line states res= and the render scale=, and both tools FAIL on a res= other than 1920x1080 times the DPR (and on a stall: no console line for 120 s). Read the PROFILE summary lines: a feature costs median(all) - median(-name). The WebGL frame clock is quantised to about 1 ms: averages are fine, but every p95 is a whole millisecond, so a p95 budget is only good to about 1 ms.
  • Boot: uv run tools/headed.py boot --runs 3 [--baseline MS] [--baseline-first-frame MS] [--cpu R] (three cold loads in a kept-awake headed window with a 1920x1080 viewport). Two spans, each the median of the runs: map_build_ms (heightmap decoded to map built) and first_frame_ms (navigation start to the second animation frame after GAME OK: download, scene load and texture uploads, map build and the first draw with its shader compiles). --cpu 4 adds CPU throttling as a cheap low-end proxy; every line states cpu=.
  • Headed frame time: uv run tools/headed.py profile --dpr 1 (--dpr 2 is the fill-rate stress proxy). --wait-unity also waits for an interactive Unity editor that has any worktree open (it holds unity/Temp/UnityLockfile): close it, or the profile waits.
  • Screenshots: tools/art_checkpoint.sh [--no-build] [--units] [--borderproof] <name> (WebGL ?shot + editor renders of both tiers + side-by-side images and their deuteranopia/protanopia copies under captures/visual/<name>/, git-ignored; rows in the main checkout's captures/visual/art-log.tsv with the branch, short commit and the image's absolute path). It refuses to run while QualitySettings.asset (and, when it builds, unity/Assets/Settings or ProjectSettings.asset) has unstaged changes, restores them on every exit, and takes the pose list from PromoPose.cs. Each checkpoint switches the worktree's build target WebGL -> OSXUniversal (editor renders) and back at the next build, a reimport each way: budget a few minutes per checkpoint. --units adds the units (docs/units.md): &units on the WebGL shots and -units on the editor renders. That shows both the unit art slice and the units in play as a new game starts. --borderproof adds the region borders' synthetic proof pattern (see "Region borders") to the editor renders only (a release WebGL build ignores &borderproof); its art-log rows are named <name>-borderproof. Normal play draws the units in play, but a ?shot run switches them off unless &units is given (UnitLayer.HiddenInCaptures), and the editor draws them only with -units. So plain checkpoints never show units, and the --units rows are named <name>-units, so they are never compared with plain ones. The kits do change for the units, though: from units-in-play's 943184c they keep a clearing where a unit stands in a town (Tel Aviv, West Jerusalem; MapScene.UnitClearings). That is a kit baseline break in the art log.

Budgets (proposed; ratified in Task 10: see "Tier decisions")

Measured on the development Mac (Apple silicon, integrated GPU), 1920x1080 at DPR 1, Chrome for Testing:

  • headless, all features on: median p95 <= 16.7 ms at both poses, and median(all) - median(none) <= 8 ms;
  • each feature: median(all) - median(-feature) <= 2.5 ms (tilt-shift <= 1.5 ms, labels <= 1 ms); a cost counts as over only when it exceeds the budget plus the larger of the two spreads;
  • headed numbers are recorded, plus a DPR-2 headed run in Task 10 as a fill-rate stress proxy for weaker GPUs; the 30 fps floor for an integrated-GPU laptop (research notes 2.3) is 33 ms p95;
  • boot: each span (map_build_ms and first_frame_ms, median of 3, at --cpu 1 and at --cpu 4) grows by at most max(300 ms, 50% of its Task 1 baseline): the tool's default when --budget is not given (controller ruling 3).

GLES3 check: the require allowlist (Task 3 Step 0)

ShaderChecks.GlslErrors fails any #extension X : require not in ShaderChecks.AllowedRequires (default-deny). An entry is added only after a WebGL 2 compileShader of a shader carrying that line succeeds in headless Chrome.

  • Report (-executeMethod Gles3Report.Run, 2026-10-10): URP Lit, URP Terrain Lit and Statecraft/HexSelection show 0 errors and no require line, so AllowedRequires stays empty. Running the report exposed a harness bug, fixed here: Gles3Errors looped over SubshaderCount but indexed GetSerializedSubshader. URP Lit has 2 subshaders and 1 serialized one, so the report threw. The loop now uses SerializedSubshaderCount.
  • Probe (corrected in Task 3 fix round 1). In WebGL 2, #extension GL_X : require compiles only after the page has called gl.getExtension("<WebGL name>"). The first probe never called it, so it could only ever say FAILS. The procedure (page and command: see "Probe page and command" below):
    1. Run Playwright's chrome-headless-shell (Chrome for Testing 149.0.7827.55) with --use-angle=swiftshader --enable-unsafe-swiftshader --dump-dom under a timeout. Full Chrome's --headless=new --dump-dom never exits on this Mac. Match the output with grep -E, because BSD grep ignores \|.
    2. Log gl.getSupportedExtensions().
    3. Positive control. For supported extensions that have a GLSL directive, compile a require shader before and after gl.getExtension. It must FAIL before and print OK after. If no control passes, the probe is broken, not the extension.
    4. Enable every supported extension, as Unity's loader does, then compile the require under test (vertex and fragment).
    5. Unity's loader. The WebGL framework (*.framework.js, emscripten) calls GLctx.getExtension(ext) for every entry of getSupportedExtensions() except names containing lose_context or debug. So step 4 matches the player, and a GLSL extension whose WebGL extension is supported is also enabled in game.
  • Result (2026-10-10, SwiftShader). Supported: EXT_clip_control, EXT_color_buffer_float, EXT_color_buffer_half_float, EXT_conservative_depth, EXT_depth_clamp, EXT_disjoint_timer_query_webgl2, EXT_float_blend, EXT_polygon_offset_clamp, EXT_texture_compression_bptc/rgtc, EXT_texture_filter_anisotropic, EXT_texture_mirror_clamp_to_edge, NV_shader_noperspective_interpolation, OES_draw_buffers_indexed, OES_sample_variables, OES_shader_multisample_interpolation, OES_texture_float_linear, OVR_multiview2, WEBGL_clip_cull_distance, WEBGL_compressed_texture_*, WEBGL_multi_draw, WEBGL_polygon_mode, WEBGL_stencil_texturing (plus the debug and lose_context ones).
    • Positive controls: GL_OVR_multiview2, GL_EXT_conservative_depth, GL_NV_shader_noperspective_interpolation, GL_OES_sample_variables, GL_OES_shader_multisample_interpolation and GL_ANGLE_multi_draw each FAIL before getExtension and are OK after it, in both stages.
    • With WEBGL_clip_cull_distance enabled, GL_ANGLE_clip_cull_distance is OK but GL_EXT_clip_cull_distance FAILS. WebGL's GLSL names can differ from GLES's, so the probe must use the exact name that Unity's GLSL requires.
    • Targets: GL_EXT_shader_non_constant_global_initializers and GL_EXT_texture_buffer still FAIL ("extension is not supported") with every supported extension enabled. Neither is a WebGL extension, so the negatives stand.
    • AllowedRequires stays empty, because none of our shaders requires anything. An extension is added only when one of our shaders needs it and its probe prints OK. Note that SwiftShader's list is not every GPU's list: a desktop-GPU run can list more.

Probe page and command

Save the page as $S/probe2/probe.html, outside the repo. Pass the GLSL names under test in the URL hash, comma-separated, and take each name verbatim from the shader report's #extension ... : require line (or the Gles3 error text), never from the WebGL extension name. To map a WebGL name to its GLSL name when you add a target to GLSL: OVR_multiview2 -> GL_OVR_multiview2, WEBGL_multi_draw -> GL_ANGLE_multi_draw, WEBGL_clip_cull_distance -> GL_ANGLE_clip_cull_distance. A target FAIL for a name that is not in the page's GLSL table only counts if a mapped positive control passed in the same run.

<html><body><script>
// WebGL extension name -> the GLSL #extension name it unlocks (null: a WebGL-only extension with no GLSL directive)
const GLSL = { OVR_multiview2: 'GL_OVR_multiview2', WEBGL_clip_cull_distance: 'GL_ANGLE_clip_cull_distance',
  WEBGL_multi_draw: 'GL_ANGLE_multi_draw', EXT_conservative_depth: 'GL_EXT_conservative_depth',
  WEBGL_draw_instanced_base_vertex_base_instance: 'GL_ANGLE_base_vertex_base_instance_shader_builtin',
  WEBGL_shader_pixel_local_storage: 'GL_ANGLE_shader_pixel_local_storage', NV_shader_noperspective_interpolation: 'GL_NV_shader_noperspective_interpolation',
  OES_sample_variables: 'GL_OES_sample_variables', OES_shader_multisample_interpolation: 'GL_OES_shader_multisample_interpolation',
  EXT_texture_shadow_lod: 'GL_EXT_texture_shadow_lod', WEBGL_blend_func_extended: 'GL_EXT_blend_func_extended' };
const gl = document.createElement('canvas').getContext('webgl2');
const lines = [];
function compile(ext, stage) {
  const sh = gl.createShader(stage === 'v' ? gl.VERTEX_SHADER : gl.FRAGMENT_SHADER);
  const body = stage === 'v' ? 'void main() { gl_Position = vec4(0.0); }' : 'precision highp float;\nout vec4 o;\nvoid main() { o = vec4(1.0); }';
  gl.shaderSource(sh, '#version 300 es\n#extension ' + ext + ' : require\n' + body);
  gl.compileShader(sh);
  const ok = gl.getShaderParameter(sh, gl.COMPILE_STATUS);
  return ok ? 'OK' : 'FAILS(' + (gl.getShaderInfoLog(sh) || '').trim().split('\n')[0] + ')';
}
if (!gl) lines.push('NO WEBGL2'); else {
  const sup = gl.getSupportedExtensions();
  lines.push('SUPPORTED ' + sup.join(' '));
  const targets = (location.hash.slice(1) || 'GL_EXT_shader_non_constant_global_initializers,GL_EXT_texture_buffer').split(',');
  // before getExtension: everything must fail (control that the require gate is real)
  for (const w of sup) if (GLSL[w]) lines.push('BEFORE ' + w + ' ' + GLSL[w] + ' v=' + compile(GLSL[w], 'v') + ' f=' + compile(GLSL[w], 'f'));
  for (const w of sup) if (GLSL[w]) { gl.getExtension(w); lines.push('AFTER ' + w + ' ' + GLSL[w] + ' v=' + compile(GLSL[w], 'v') + ' f=' + compile(GLSL[w], 'f')); }
  // the GLSL names under test: enable every supported WebGL extension first, then compile
  for (const w of sup) gl.getExtension(w);
  for (const t of targets) lines.push('TARGET ' + t + ' v=' + compile(t, 'v') + ' f=' + compile(t, 'f'));
}
document.body.textContent = lines.join('\n');
</script></body></html>
S=<scratch dir outside the repo>
HS=$(find ~/Library/Caches/ms-playwright/chromium_headless_shell-1228 -name chrome-headless-shell -type f | head -1)
perl -e 'alarm 60; exec @ARGV' "$HS" --user-data-dir=$S/probe2/prof --use-angle=swiftshader --enable-unsafe-swiftshader \
  --dump-dom "file://$S/probe2/probe.html#GL_EXT_foo,GL_EXT_bar" 2>/dev/null | sed -e 's/<[^>]*>//g' \
  | grep -E "^(SUPPORTED|BEFORE|AFTER|TARGET|NO)"
grep -oE 'getSupportedExtensions\(\)\|\|\[\];exts.forEach[^}]*' unity/Builds/Web-disabled/Build/*.framework.js   # the loader's enable-all

Read the output as follows:

  • At least one AFTER ... v=OK f=OK line (the control) is required; without it, the probe is broken.
  • TARGET <name> v=OK f=OK means the extension may be added to ShaderChecks.AllowedRequires, with the comment // <name>: verified to work in WebGL 2 in headless Chrome (Chrome for Testing 149.0.7827.55, <date>), but only if one of our shaders actually requires it.
  • TARGET ... FAILS means that shader cannot run on the web tier.
  • Re-check the loader grep after a Unity upgrade.

Baseline (Task 1, before any visual change; re-recorded in fix round 1)

Viewport fixed to 1920x1080 at DPR 1 over CDP; Chrome kept awake; build from 0b6058e (harness only).

Run Lines
headless PROFILE summary visual=all pose=gameplay03 runs=3 median_avg_ms=2.29 spread_ms=0.17 median_p95_ms=4.00
PROFILE summary visual=none pose=gameplay03 runs=3 median_avg_ms=2.35 spread_ms=0.14 median_p95_ms=4.00
PROFILE summary visual=all pose=diorama02 runs=3 median_avg_ms=6.84 spread_ms=0.06 median_p95_ms=7.00
PROFILE summary visual=none pose=diorama02 runs=3 median_avg_ms=6.92 spread_ms=0.05 median_p95_ms=8.00
run line: PROFILE visual=all pose=gameplay03 res=1920x1080 scale=0.80 rep=0 frames=300 avg_ms=2.43 p95_ms=4.00
headed (DPR 1) PROFILE summary visual=all pose=gameplay03 runs=3 median_avg_ms=2.31 spread_ms=0.57 median_p95_ms=4.00
PROFILE summary visual=none pose=gameplay03 runs=3 median_avg_ms=1.83 spread_ms=0.76 median_p95_ms=3.00
PROFILE summary visual=all pose=diorama02 runs=3 median_avg_ms=7.18 spread_ms=0.29 median_p95_ms=8.00
PROFILE summary visual=none pose=diorama02 runs=3 median_avg_ms=7.15 spread_ms=0.40 median_p95_ms=8.00
run line: PROFILE visual=all pose=gameplay03 res=1920x1080 scale=0.80 rep=0 frames=300 avg_ms=1.78 p95_ms=3.00
boot, --cpu 1 BOOT map_build_ms median=443 runs=3 cpu=1
BOOT first_frame_ms median=3547 runs=3 cpu=1
boot, --cpu 4 BOOT map_build_ms median=1813 runs=3 cpu=4
BOOT first_frame_ms median=6963 runs=3 cpu=4

Boot budgets from these (ruling 3, baseline + max(300 ms, 50%)): --cpu 1 map build 443 -> 743 ms, first frame 3547 -> 5321 ms; --cpu 4 map build 1813 -> 2720 ms, first frame 6963 -> 10445 ms.

Notes:

  • The headed window is clamped by the laptop screen (1728 px wide); the emulated 1920x1080 viewport still gives a 1920x1080 backbuffer (res=1920x1080), only the final composite to the window is scaled.
  • No toggles exist yet, so all and none are the same scene: their difference and the spreads are the noise floor. Headless spreads are 0.05-0.17 ms; headed gameplay03 is noisier (spreads 0.57-0.76 ms, all vs none 0.48 ms apart).
  • Superseded first recording (commit 33fb072, before the CDP viewport): headless ran at res=1920x937 and headed at res=1728x860, with a first-pose spread of 0.68 ms (headless) and one headed run that stalled; do not compare with it.

Task 2: sun and shadows

Headless, viewport 1920x1080 at DPR 1 (res=1920x1080 scale=0.80), web tier (shadow map 2048, one cascade, cascade border 0.1).

Shadow fit (fix round 1, review I1). URP fades shadows by straight-line distance from the camera, from (1 - cascadeBorder) x shadowDistance out (URP 17.3 Shadows.hlsl, ShadowUtils.GetScaleAndBiasForLinearDistanceFade). The first delivery fitted 1.75 x the rig distance from view depth, which left the top corners partly or wholly faded (about 76% faded on the web tier at a 0.08 top inset). ShadowFit now takes, on every rig Apply, the farthest straight-line distance from the camera to the ground hits of the four viewport corners (the current aspect and top inset included; ground plane y = 0, the terrain's lowest point, so the fit is never short) and divides it by (1 - cascadeBorder). It clamps to 3 km .. 1,000 km (the far clip). LightingTests.TheShadowDistanceCoversTheWholeView checks 6/45/120/250 km x 16:9, 21:9, 4:3 x top insets 0, 0.08, 0.3, 0.5 x borders 0.1 and 0.2.

Decision: web-tier cascade border 0.2 -> 0.1 (Mobile_RPAsset m_CascadeBorder). With one cascade, the border is the fade band at the edge of the fitted range. A tenth keeps the fade off-screen while spending less of the 2048 map past the view (the fitted distance is 1/0.9 rather than 1/0.8 of the farthest corner). PC keeps 0.108 (4 cascades).

Run Lines
headless, first delivery (8b20507, 1.75 x D view-depth fit, border 0.2) PROFILE summary visual=all pose=gameplay03 runs=3 median_avg_ms=2.11 spread_ms=0.10 median_p95_ms=3.00
PROFILE summary visual=-shadows pose=gameplay03 runs=3 median_avg_ms=2.43 spread_ms=0.11 median_p95_ms=4.00
PROFILE summary visual=all pose=diorama02 runs=3 median_avg_ms=10.12 spread_ms=0.43 median_p95_ms=11.00
PROFILE summary visual=-shadows pose=diorama02 runs=3 median_avg_ms=7.39 spread_ms=0.17 median_p95_ms=8.00
headless, fix round 1 (corner fit, border 0.1) PROFILE summary visual=all pose=gameplay03 runs=3 median_avg_ms=2.03 spread_ms=0.09 median_p95_ms=3.00
PROFILE summary visual=-shadows pose=gameplay03 runs=3 median_avg_ms=2.42 spread_ms=0.08 median_p95_ms=4.00
PROFILE summary visual=all pose=diorama02 runs=3 median_avg_ms=9.87 spread_ms=0.46 median_p95_ms=11.00
PROFILE summary visual=-shadows pose=diorama02 runs=3 median_avg_ms=7.12 spread_ms=0.11 median_p95_ms=8.00

Budget check, fix round 1 (median(all) - median(-shadows) <= 2.5 ms + the larger spread):

  • gameplay03: -0.39 ms against 2.59 ms: within.
  • diorama02: 2.75 ms against 2.96 ms (2.5 + 0.46): within, with 0.21 ms to spare.

For Task 10 (review M5). At gameplay03, all is faster than -shadows by more than both spreads, in both deliveries. Three runs understate the run-to-run noise, so the 0.2 ms headroom at diorama02 means little. Re-measure diorama02 shadows with more repeats (at least 7) and headed. If the tier review needs headroom, the levers are:

  • terrain shadow LOD or pixel error (the whole terrain is drawn into the shadow map at about 400 km);
  • a cheaper fit at the far zoom (fit to the map's extent rather than the frustum once the frustum is wider than the map);
  • a web-tier shadow map of 1024 at the far zoom only.

Rendered coverage (review I2, I3). ReliefCastsShadowsAtMapScale runs at both tiers over the Dead Sea escarpment at 60 km. Every other registered visual except water is off and the settlement markers are hidden, so only relief casts. Results: shadowed share 3.06% (web) and 3.12% (PC); the control at the asset's own 50 m is 0.00% at both. Hiding the markers moved the web share from 3.03% to 3.06%, which is noise, so the cubes were never a material part of it. The 2% floor stays: it allows a third of the measured share to be lost to the merge (Task 3's tilt-shift softening edges, Task 4's grade compressing darks, Track B's haze lifting distant shadow), and the control's 0.00% shows the separation is clean. If the merge squeezes it further, switch to a relative test (lit < flat * 0.9), not a lower floor.

Ambient (review M4). Superseded by Task 4: the explicit probe described here was never read. In Trilight mode Unity replaces it with its own linear conversion of the colours, and Task 4 removed MapLighting.AmbientProbe. The evaluated values below are that conversion, and they match the constants once converted to sRGB; see "Task 4: grade". The original note follows. MapLighting.AmbientProbe builds the trilight from AddAmbientLight plus two AddDirectionalLight lobes. The lobes add constant and linear terms in every direction, so the probe evaluates darker than its constants, which are inputs, not the evaluated colours. Evaluated values, pinned in LightingTests: up (0.216, 0.255, 0.326), horizon (0.184, 0.179, 0.171), down (0.119, 0.103, 0.080); the constants are sky (0.52, 0.60, 0.72), equator (0.48, 0.46, 0.42), ground (0.26, 0.22, 0.17). The images were judged on the evaluated values, and Track B's thresholds were derived with them, so a least-squares refit would be a look change for Task 4, not a fix here.

Task 3: tilt-shift

Two bands, blurred by a half-resolution render-graph pass (TiltShiftFeature on both renderers, Statecraft/TiltShift):

  • the blur is a separable 9-tap Gaussian whose radius is TiltShift.Weight(y) * MaxRadiusUv;
  • the focus sits at the middle of the map area below the top bar;
  • the blur weakens to 35% between 60 km and the 250 km overview.
  • since issue #390, the far (top) band is lighter at the gameplay zoom (see "Issue #390" below).
  • since issue #458 (the owner's ruling of 2026-10-14), the near (bottom) band is lighter at the gameplay zoom and about gone at the overview (see "Issue #458" below).

URP's far-only Gaussian DoF is retired. Toggle: tiltshift; the URL flag notiltshift (desktop: -notiltshift) boots with it off.

Constants (fix round 1, review I2). HalfWidth 0.13, Feather 0.17 and MaxRadiusUv 14/1080 (was 0.17, 0.18 and 10/1080). Full blur now starts 0.30 from the focus. That is the bottom and top 20% with no bar; under the web tier's top bar (8.8% of the screen) it is the bottom 15.5%, and 15.5% between the band and the bar. The promo has about 15% at the bottom and 20% at the top.

Edit-mode ratios (gradient energy shifted / sharp, gameplay pose, 960x540, no bar):

Tier top 0.90–0.98 bottom 0.02–0.10 middle top 0.81–0.87 bottom 0.13–0.19 top at 1920x1080
Mobile 0.16 0.24 1.00 0.17 0.22 0.14
PC 0.11 0.18 1.00 0.12 0.15 0.11

TopInset 0.3: focus band 1.00, upper band 0.13. (Before fix round 1: Mobile top 0.22 and bottom 0.31; PC 0.16 and 0.24.)

WebGL band check, with the real top bar (review m3). I compared the checkpoint's gameplay03.png with a ?shot&notiltshift shot from the same build, using columns 30–80% (clear of the info panel and minimap) and rows measured from the top. The bar ends at 0.088.

  • Ratios by row (from the top):

    Rows 0.10–0.275 0.30 0.325 0.35 0.375 0.425–0.675 0.70 0.725 0.75 0.775 0.80–1.00
    Ratio 0.01 0.13 0.32 0.60 0.86 1.00 0.74 0.45 0.22 0.07 0.01
  • The sharp band is centred at about 0.54 from the top. The focus is at (1 − 0.088)/2 = 0.456 from the bottom, so this is the right way up; a flipped composite would centre at 0.456 from the top.

  • The review's check: rows 0.20–0.30 from the top (nearer the bar) keep 0.02 of their detail, and rows 0.70–0.80 keep 0.37.

  • Coverage. Fully blurred (ratio <= 0.05) from the bar to 0.30 (21% of the screen at the top) and from 0.80 to the bottom (20% at the bottom). The promo has about 20% at the top and 15% at the bottom. By the design weight (weight = 1), the coverage is 15.6% at each end.

Profile (1920x1080, DPR 1, render scale 0.80, 3 interleaved runs per configuration), fix round 1 build:

headless (web_check.py)
PROFILE summary visual=all pose=gameplay03 runs=3 median_avg_ms=1.96 spread_ms=0.13 median_p95_ms=3.00
PROFILE summary visual=-tiltshift pose=gameplay03 runs=3 median_avg_ms=2.21 spread_ms=0.79 median_p95_ms=4.00
PROFILE summary visual=all pose=diorama02 runs=3 median_avg_ms=9.65 spread_ms=0.20 median_p95_ms=10.00
PROFILE summary visual=-tiltshift pose=diorama02 runs=3 median_avg_ms=9.72 spread_ms=0.26 median_p95_ms=11.00
headed (headed.py profile --dpr 1)
PROFILE summary visual=all pose=gameplay03 runs=3 median_avg_ms=2.01 spread_ms=0.39 median_p95_ms=3.00
PROFILE summary visual=-tiltshift pose=gameplay03 runs=3 median_avg_ms=1.89 spread_ms=0.35 median_p95_ms=3.00
PROFILE summary visual=all pose=diorama02 runs=3 median_avg_ms=10.11 spread_ms=0.54 median_p95_ms=11.00
PROFILE summary visual=-tiltshift pose=diorama02 runs=3 median_avg_ms=10.55 spread_ms=0.59 median_p95_ms=11.00

Cost median(all) - median(-tiltshift), all within the 1.5 ms budget and within the spreads:

gameplay03 diorama02
headless −0.25 ms −0.07 ms
headed +0.12 ms −0.44 ms
first build: headless −0.08 ms +0.48 ms
first build: headed +0.01 ms +0.07 ms

Task 4: grade

MapGrade adds the following to the rig's runtime VolumeProfile, all behind the grade toggle, with LDR grading on both tiers (values as of Task 4; issue #38 retuned them, see "Issue #38: grade retune"):

  • a Neutral tonemapper;
  • post exposure +0.15, contrast 14, saturation 16 (superseded at the Merge point: MapGrade.Saturation is 0, see "Merge point");
  • white balance temperature 5 (the plan says 9; see the merge arithmetic below) and tint 2;
  • split toning: teal shadows (0.36, 0.52, 0.60), amber highlights (0.96, 0.78, 0.55), balance −15;
  • a dark-navy vignette: 0.26, smoothness 0.45.

What was there before (measured). The pipeline assets' own volume profile, Settings/SampleSceneProfile (the URP template's), already applied the following to every frame:

  • Neutral tonemapping;
  • bloom 0.25 above 1 (high quality);
  • a 0.2 black vignette.

So the build was not "without a tonemapper". The grade toggle off falls back to that profile, not to a raw image. Two consequences:

  • The control in TheSunlitDesertDoesNotClip sets the grade's tonemapper to None. That overrides the template's Neutral, and the graded Negev then clips 2.3% of its pixels, against 0.00% with Neutral.
  • URP strips the tonemapper's uber variants unless some VolumeProfile asset has a Tonemapping component. The runtime profile is not an asset, so it does not count. TheGradeIsWarmLandTealShadowsAndAGentleVignette checks that such an asset exists. Today two assets carry one: the template's SampleSceneProfile and the global settings' DefaultVolumeProfile. The check fails only if both lose their Tonemapping component. If Task 10 drops the template profile (the bloom decision), it should add a project-owned VolumeProfile with Neutral, so the guarantee no longer rests on URP template leftovers.

Bloom is the template's, not this plan's; Task 10 decided: bloom ships off on both tiers (see "Tier decisions").

Readability (review I9), from edit mode at 960x540, with the tilt-shift off for these detail checks.

  • ΔE76 between the selection band and the land 1 km inside the hex:

    Pose Graded Ungraded
    gameplay03 32.3 24.5
    diorama02 48.6 36.6
  • Colour-blind check (Machado 2009, gameplay pose):

    Simulation sea vs land selection vs land selection vs sea
    Deuteranopia 68.6 31.8 98.6
    Protanopia 67.8 31.5 97.6
  • Saturation rises from 0.438 to 0.564.

  • The vignette's corner/centre ratio drops from 0.603 to 0.483.

  • In the checkpoint's deuteranopia and protanopia copies, the coast (blue against yellow) and the hex lines stay clear.

Ambient (review M4, closed by measurement).

  • How the probe actually works. MapLighting.AmbientProbe() was never what the shaders read. In Trilight mode Unity replaces any assigned probe with its own conversion of the three colours, at once. It does this in the editor, and in the WebGL player too: a diagnostic build logged the getter at Apply, frame 120 and frame 600.

    • Editor renders with our probe, a least-squares refit and a black probe are identical to 3 decimals. Changing the trilight colours changes the image.
  • Why it looked like 40%. Unity converts the sRGB colours to linear. The "40% of nominal" (horizon 0.18 against 0.48) compared linear values with sRGB constants. In sRGB the horizon is (0.466, 0.460, 0.450) against the equator's (0.48, 0.46, 0.42).

  • Decision: keep the look. The dead explicit probe is removed, which changes nothing rendered. LightingTests now asserts the horizon is within 0.05 of the equator in sRGB.

  • Why no brighter refit. A refit that brightens the ambient towards the constants (trilight colours raised so the probe's linear values reach the constants) does not help. Measured with the grade, in editor renders:

    Pose Warmth Saturation Luma Contrast
    gameplay03 0.142 → 0.082 0.572 → 0.518 +0.03 unchanged
    diorama02 0.154 → 0.090 0.526 → 0.458 +0.03 unchanged

    That works against the merge's warmth target. It would also shift every Track B threshold, which were derived with the current ambient. Nothing Track B measures moves.

Grade terms at diorama02 (editor, Mobile tier, no UI; each line takes one term out of the full grade). With the full grade: warmth 0.154, saturation 0.526, contrast 0.110. With the grade off: warmth 0.120, saturation 0.419, contrast 0.088. The temperature here was 9.

Term Warmth Saturation Other
Saturation 16 +0.027 +0.084
White balance 9 +0.030 +0.014
Split toning −0.039 (the teal shadows) −0.041
Vignette −0.018
Contrast +0.011 contrast

Merge arithmetic (diorama02 warmth, WebGL frame).

  • Track A alone at Temperature 9: t04 is 0.118 (t03 was 0.088).
  • Track B's contribution: t06-terrain-r1 0.0384 − t01 0.0171 = +0.021 before the grade.
  • How the grade scales it: the grade's saturation (×1.16), contrast and exposure scale chroma by about 1.0–1.3, so Track B's +0.021 becomes +0.021 to +0.027.
  • At Temperature 9: 0.118 + 0.021…0.027 = 0.139–0.145. That is already at the promo's 0.145, before Track B's Task 7 haze and palette work.
  • Lowering the temperature: one unit is about 0.003 of frame warmth: 0.030 for 9 units in the editor, × 0.88 for the WebGL frame's UI share. So Temperature 5 gives 0.118 − 4 × 0.003 = 0.106, and the checkpoint measured 0.106.
  • Expected merged diorama02 warmth: 0.106 + 0.021…0.027 = 0.127–0.133. That is at least 0.087 by 0.04, and leaves 0.012–0.018 for Track B's later warming.
  • Contrast: t03 0.153 → t04 0.170, plus Track B's +0.014 (t06-r1 0.1731 − t01 0.1588), gives about 0.18–0.19 after the merge, moving towards the promo's 0.200.
WebGL Saturation Contrast Warmth Promo (sat / contrast / warmth)
diorama02, t03 0.422 0.153 0.088 0.425 / 0.200 / 0.145
diorama02, t04 (temp 9, 9603700) 0.52 0.169 0.118
diorama02, t04 (temp 5, 2ac7b6e) 0.51 0.170 0.106
gameplay03, t03 0.444 0.155 0.082 0.382 / 0.203 / 0.064
gameplay03, t04 (temp 5) 0.55 0.174 0.107

Over the promo (owner's call).

  • Saturation is now about 0.09 above the promo at diorama02 and 0.17 above it at gameplay03. The lever is Saturation (16 → 8 is about −0.04).
  • Gameplay03 warmth is already above its promo (0.064) on Track A alone, and Track B adds about +0.03. The promo's gameplay frame has green fields and orchards, which is the art track's vegetation, and a global grade cannot fix that without cooling the diorama.

Seam test. TheSeaBeyondTheEdgeMatchesTheTerrainSea measures the seam twice:

  • Before the grade, against 0.012. This isolates the grade, which is a per-pixel LUT plus a vignette that is flat this close to the centre, so it cannot open a seam.
  • On the shipped, graded frame, against 0.024. The grade scales the sea's chroma by about 2: the sub-step mismatch between the terrain sea and the skirt shader goes from 0.008 to 0.016.

At the merge, Track B's patch-mean version should keep this graded/ungraded loop.

Profile (headless, 1920x1080, DPR 1, scale 0.80, 3 interleaved runs; build 9603700. Temperature only changes LUT values, so the cost is the same at 2ac7b6e):

PROFILE summary visual=all pose=gameplay03 runs=3 median_avg_ms=1.95 spread_ms=0.39 median_p95_ms=3.00
PROFILE summary visual=-grade pose=gameplay03 runs=3 median_avg_ms=1.93 spread_ms=0.37 median_p95_ms=3.00
PROFILE summary visual=all pose=diorama02 runs=3 median_avg_ms=9.74 spread_ms=0.06 median_p95_ms=11.00
PROFILE summary visual=-grade pose=diorama02 runs=3 median_avg_ms=9.74 spread_ms=0.02 median_p95_ms=11.00

The grade costs +0.02 ms at gameplay03 and 0.00 ms at diorama02, inside the noise. The tonemapper and the vignette were already in the uber pass (the template profile), and the LUT pass ran before too.

Baked textures (Task 5)

tools/terrain_bake.py regenerates unity/Assets/Textures/Map/terrain_bake.png, coast.png and ao.png from the committed heightmap in about 10 s. The output is deterministic per toolchain: re-running it with the pinned numpy 2.5.3 and pillow 12.3.0 on the same machine gives byte-identical PNGs. Another platform may differ in its zlib build or in 1-ulp sin/arctan results. Compare decoded pixels (np.array_equal), not file bytes; R and B may then differ by one level here and there.

Task 6: terrain

The Statecraft Terrain Lit Shader Graph (tools/terrain_graph.py + StatecraftTerrain.hlsl): one terrain layer (one pass) instead of seven, no basemap fallback. WebGL build (-compression disabled): BUILD Succeeded WebGL -> Builds/Web-disabled: 73.4 MB on disk, 0 errors (Task 1: 70.9 MB; +2.5 MB, the two baked textures), no Shader error. GLES3 check: ShaderChecks.Gles3Errors is empty for Statecraft/StatecraftTerrain (no #extension ... : require; AllowedRequires not needed).

Run Lines
headless PROFILE summary visual=all pose=gameplay03 runs=3 median_avg_ms=1.58 spread_ms=0.93 median_p95_ms=2.00
PROFILE summary visual=-haze pose=gameplay03 runs=3 median_avg_ms=1.88 spread_ms=0.87 median_p95_ms=4.00
PROFILE summary visual=none pose=gameplay03 runs=3 median_avg_ms=1.67 spread_ms=0.40 median_p95_ms=3.00
PROFILE summary visual=all pose=diorama02 runs=3 median_avg_ms=7.25 spread_ms=0.47 median_p95_ms=8.00
PROFILE summary visual=-haze pose=diorama02 runs=3 median_avg_ms=7.13 spread_ms=0.11 median_p95_ms=8.00
PROFILE summary visual=none pose=diorama02 runs=3 median_avg_ms=7.16 spread_ms=0.20 median_p95_ms=8.00
boot, --cpu 1 BOOT map_build_ms median=509 runs=3 cpu=1
BOOT first_frame_ms median=3851 runs=3 cpu=1 (budgets 743 / 5321 ms: within)

Against the Task 1 baseline (all): gameplay03 2.29 -> 1.58 ms, diorama02 6.84 -> 7.25 ms (+0.41 ms, inside 2.5 ms plus the spread). Haze costs nothing measurable (all - -haze: -0.30 / +0.12 ms, both inside the spreads). The boot phase biome texture baked takes 94-106 ms (it replaces Task 1's biome colours blended).

Task 6 fix round 1

Haze retuned (review I1: start 200 km, scale 250 km, max 0.45, rift boost 0.6, colour sRGB (0.84, 0.74, 0.58)), ground palette x1.107 in sRGB, curvature gain 0.40, graph and nodes at Single precision, HDRP target dropped, trilinear map textures (bake aniso 2), the fine noise octave faded by its pixel footprint. The AO is its own uncompressed R8 texture (ao.png): measured against the source PNG, BC1 stepped it by 2+ levels on 404 of 681 broad Judean texels (59%); R8 gives 0 of 681 (Judean slopes) and 0 of 6206 (Negev plateau). WebGL build: 76.9 MB (+3.5 MB for ao.png), no Shader error. GLES3 report (with vis-light e8b1c48's serialized-subshader loop): 0 errors and no require for Statecraft/StatecraftTerrain, Statecraft/HexSelection, URP Lit (the sea on this branch) and URP Terrain Lit.

Run Lines
headless PROFILE summary visual=all pose=gameplay03 runs=3 median_avg_ms=2.41 spread_ms=0.88 median_p95_ms=4.00
PROFILE summary visual=-haze pose=gameplay03 runs=3 median_avg_ms=2.34 spread_ms=0.48 median_p95_ms=4.00
PROFILE summary visual=none pose=gameplay03 runs=3 median_avg_ms=2.46 spread_ms=0.07 median_p95_ms=4.00
PROFILE summary visual=all pose=diorama02 runs=3 median_avg_ms=7.29 spread_ms=0.07 median_p95_ms=8.00
PROFILE summary visual=-haze pose=diorama02 runs=3 median_avg_ms=7.31 spread_ms=0.10 median_p95_ms=8.00
PROFILE summary visual=none pose=diorama02 runs=3 median_avg_ms=7.34 spread_ms=0.09 median_p95_ms=8.00
boot, --cpu 1 BOOT map_build_ms median=452 runs=3 cpu=1
BOOT first_frame_ms median=3559 runs=3 cpu=1 (budgets 743 / 5321 ms: within)

Art (WebGL, whole frame; vis-ground has no Task 2 light, so compare with t01): diorama02 warmth 0.017 (t01) / 0.007 (t06) -> 0.038, saturation 0.36 -> 0.41, contrast 0.171 -> 0.173, palette distance 0.85 -> 0.78; map land alone (column x 0.45-0.6) R-B 0.158 -> 0.229, saturation 0.254 -> 0.363, luma 0.576 -> 0.572. gameplay03 warmth 0.009 (t01) / 0.017 -> 0.040, contrast 0.163 -> 0.173, palette distance 0.69 -> 0.66. The checkpoint also shoots the 6 km hero zoom over the Judean slopes and the Negev (PromoPose.CloseChecks, haze off): no AO steps.

Task 7: sea

One water function (StatecraftMap.hlsl ScWaterShade) on both sides of the map's edge: the terrain graph outputs it as emission on a black, fully metallic, fully occluded surface, and the border skirt's Sea.mat (Statecraft/Sea, passes Sea, DepthOnly, DepthNormals) returns it directly. (Since issue #195 the terrain graph outputs only the black surface; MapScene's TerrainWater mesh adds the water on top with Sea.mat in its _ScTerrainWater mode. See "Issue #195".) It is driven by Task 5's coast distance: a turquoise -> teal -> navy ramp over 12 km, foam within 350 m, the Dead Sea pale turquoise with a white salt rim. It samples the main light's shadows itself, takes the cooler water haze (ScHazeWater), and falls back to the flat sea colour under the water toggle. WebGL build: 77.1 MB, no Shader error. GLES3: ShaderChecks.Gles3Errors is empty for Statecraft/Sea and the terrain (no require).

Deviations from the plan:

  • Glint 0.06, not 0.25. With Map.unity's default 50-degree sun, the camera's 50-degree pitch at 120 km meets the reflection at yaw 180. There the 0.25 lobe lit a warm patch rising 0.08 in sRGB red over 6 km of open sea, and the seam test (samples 5 km apart) read it as a 0.031 seam. A profile across the edge showed no step at the edge itself, and the difference scaled with the glint. At 0.06, every seam line is <= 0.004.
  • The held skirt is averaged along the edge. Holding each edge vertex's own height for 20 km (SkirtFall = {0, 0.05, 0.15, 0.45, 1}) drew the edge's ridges out as long stripes across the Jordan skirt (checkpoint t07-sea, first run). Rings 1-4 now hold the edge height averaged over +-8 km (MapBuilder.SkirtHoldSmoothingM). Ring 0 still meets the terrain exactly. New test TheHeldSkirtCarriesThePlateauNotTheEdgesRidges: the mean spoke-to-spoke step 20 km out is 8.5 m against 36.7 m at the edge (31.2 m before).
Run Lines
headless PROFILE summary visual=all pose=gameplay03 runs=3 median_avg_ms=2.38 spread_ms=0.42 median_p95_ms=4.00
PROFILE summary visual=-haze pose=gameplay03 runs=3 median_avg_ms=2.35 spread_ms=0.08 median_p95_ms=4.00
PROFILE summary visual=-water pose=gameplay03 runs=3 median_avg_ms=1.70 spread_ms=0.61 median_p95_ms=3.00
PROFILE summary visual=none pose=gameplay03 runs=3 median_avg_ms=1.72 spread_ms=0.27 median_p95_ms=2.00
PROFILE summary visual=all pose=diorama02 runs=3 median_avg_ms=7.64 spread_ms=1.56 median_p95_ms=9.00
PROFILE summary visual=-haze pose=diorama02 runs=3 median_avg_ms=7.67 spread_ms=0.29 median_p95_ms=8.00
PROFILE summary visual=-water pose=diorama02 runs=3 median_avg_ms=7.78 spread_ms=0.35 median_p95_ms=9.00
PROFILE summary visual=none pose=diorama02 runs=3 median_avg_ms=7.48 spread_ms=0.02 median_p95_ms=8.00
boot, --cpu 1 BOOT map_build_ms median=472 runs=3 cpu=1
BOOT first_frame_ms median=3592 runs=3 cpu=1 (budgets 743 / 5321 ms: within)

Against Task 6 fix round 1 (all): gameplay03 2.41 -> 2.38 ms, diorama02 7.29 -> 7.64 ms (+0.35 ms, inside 2.5 ms plus the spread). Water costs 0.68 ms at gameplay03 (all - -water, against a 0.61 ms spread) and nothing measurable at diorama02 (-0.14 ms), well inside 2.5 ms. The profile ran on the build before the skirt averaging, which changes no shader. The boot lines are from the final build.

Art (WebGL): diorama02 saturation 0.41 -> 0.39, contrast 0.173 -> 0.186, warmth 0.038 -> 0.058, palette distance 0.78 -> 0.76. gameplay03 contrast 0.173 -> 0.194, palette distance 0.66 -> 0.65. The sea grades from turquoise shallows with a foam line to navy, and the Dead Sea is pale turquoise with white salt rims. Coast streaks: none visible. A 12x high-pass of the gameplay03 coast shows only the foam band's waves and single-level output steps in the deep water, so the bake's distance needs no smoothing (Task 5 carry). Plateau edge (diorama02): the Jordanian plateau now continues past the east edge at the edge's broad height instead of sagging at once, as a flat muted band level with the map's eastern rim. A faint low-contrast banding remains there at grazing angles.

Task 7 fix round 1

  • Haze at full width (review I1, the Task 6 carry).
    • At diorama02, the haze toggle moves the far band 0.047 and the near band 0.002 over the full width. Task 6 measured 0.023 / 0.026, with the sea and skirt unhazed. Restricted to the land column it was 0.069 / 0.026.
    • So TheHazeThickensWithDistance and TheHazeStartsBeyondTheNearestDioramaLandAndGrowsWithDistance sample x 0..1 again. By distance: near 0.0019, mid 0.0212, far 0.0473.
    • TheOpenSeaStaysBlueUnderTheHaze now asserts that the water haze applies. The open sea at about 300 km moves 0.059 in luma with the haze on, and 0.000 with ScHazeWater dropped on both sides (the red check).
  • Seam test close pair (M1).
    • The pair sits at +-600 m across the edge. Each side is the mean of five 7x7 patches, 2 km apart along the edge, because a single 1 km patch carries one wave cell or glint sparkle on its own (up to 0.012).
    • A step of 5% on the skirt sea only within 1.5 km of the edge passes the +-2.5 km pairs (0.000) and fails the close pair (0.015). All 16 lines are <= 0.004.
  • Skirt hold (M4).
    • The hold is now a running mean over the edge's own samples (prefix sums, +-8 km). It replaces the 17 taps 1 km apart.
    • East edge, spoke-to-spoke step 20 km out: 8.5 -> 5.2 m. What remains is the plateau's own trend along the edge.
    • The test bound is under 0.18 x the edge's step.
    • The faint grazing-angle banding in diorama02 is gone.
  • Land skips the water shading (M3). [branch] on water > 0.
    • gameplay03 all minus -water: 0.03 ms before, 0.01 ms after. Both are inside the spreads, so there is no measurable change on this GPU.
    • Before, on the build at 6757168: PROFILE summary visual=all pose=gameplay03 runs=3 median_avg_ms=2.49 spread_ms=0.17 and visual=-water ... median_avg_ms=2.46 spread_ms=0.06.
    • After: visual=all ... median_avg_ms=2.55 spread_ms=0.07, visual=-water ... median_avg_ms=2.54 spread_ms=0.02; diorama02 all 7.34 ms (before 7.38 ms).
    • The earlier 0.68 ms water cost was run-to-run noise.
  • Build and boot. WebGL 77.1 MB, no Shader error. Boot --cpu 1: map_build_ms 469, first_frame_ms 3598 (budgets 743 / 5321).

Task 8: grid

The hex grid is drawn by the terrain graph and by the skirt's Sea.shader from one distance field (ScHexEdgeDistance, tools/hexgrid.py's layout; HexGrid is the C# twin), anti-aliased with fwidth, 1.5 px wide, light on land (0.25) and water (0.35), fading to 0.6 of that from 40 km to 240 km camera distance (superseded: land 0.35, water 0.10, fade to 0.35 plus a footprint fade; see "Task 8 fix round 1" and "Merge point"). The MeshTopology.Lines overlay, its material, OverlayClearanceM and its boot phase are gone; the selected-hex band sits 30 m above the ground (was 140 m). Toggle grid, registered in MapShaderGlobals.Apply through assets.OnTeardown. All numbers here are pre-grade (Track A's grade and tilt-shift are not in this worktree).

  • Rendered tests (EditMode). Edge luma change 0.059 (hills, 30 km). Grid coverage: 3.08% at 20 km, 18.15% at 250 km (bound 0.05%-25%). Rift hex lines: Beit She'an 0.062, Jericho 0.058, Dead Sea shore / Arava 0.016 (light lines on the bright desert are faint; that test's bound is 0.012, the plan's 0.03 holds only off the desert).
  • Seam. Task 7's pairs and close strip, grid off: 16 lines, max 0.004. Grid on: 12 asserted lines, max 0.008; the close strip at yaw 90/270 is logged only (0.014 at yaw 270: the x = 0 hex line falls unequally into the two foreshortened patches). The plan's hex-centre pairs at x = +-4330 m were dropped: the middle-row pair differs by 0.027 at yaw 180 with the grid off too (a warm sun glint on the terrain side), a pre-existing water effect.
  • Profile (WebGL, 1920x1080).
    • PROFILE summary visual=all pose=gameplay03 runs=3 median_avg_ms=2.42 spread_ms=0.74 median_p95_ms=4.00
    • PROFILE summary visual=-grid pose=gameplay03 runs=3 median_avg_ms=2.41 spread_ms=0.22 median_p95_ms=4.00
    • PROFILE summary visual=all pose=diorama02 runs=3 median_avg_ms=7.30 spread_ms=0.12 median_p95_ms=8.00
    • PROFILE summary visual=-grid pose=diorama02 runs=3 median_avg_ms=7.28 spread_ms=0.34 median_p95_ms=8.00
    • The grid costs 0.01-0.02 ms, inside the spreads; with the line mesh's draw call gone, gameplay03 all is 2.42 ms (Task 7 fix round 1: 2.55 ms).
  • Build and boot. WebGL 77 MB, no Shader error; no GLES derivative error (no #pragma require derivatives needed). Boot --cpu 1: BOOT map_build_ms median=419, BOOT first_frame_ms median=3442 (Task 7: 469 / 3598).
  • Look (t08-grid). The sea lattice runs on past the map's edge with no step. At the 250 km overview the sea lattice is dense and reads strongly; land lines are faint, nearly invisible on the desert.

Task 8 fix round 1

  • Line anti-aliasing (review I1). fwidth of the V-shaped edge distance collapsed on the line itself. The fix projects the derivatives of the continuous grid coordinate on the nearest edge's normal (metres per pixel across the edge) and box-filters: saturate(0.5 * widthPx + 0.5 - d / aa).
    • New rendered test GridLinesAreUnbrokenAndAboutTheirWidthAtGameplay03 (1920x1080, opacities forced to 1, linear luma).
      • Red at b409955: sea peak ratio 0.41 along one edge, FWHM 2.74 px.
      • Green: sea ratio 0.85, FWHM 1.66 px (1.51-1.95); land FWHM 1.62 px (1.37-1.97).
  • Weights (review I2). GridWaterOpacity 0.35 -> 0.10. The distance fade is now lerp(1, 0.35). Lines also fade out as a hex shrinks from 30 to 10 px on screen (circumradius over the pixel footprint). Land stays 0.25 until the grade merges.
    • Sea line in gameplay03 (x 120-170, y 480-560): background 49, peak 83, +34, ratio 1.69, 2 px wide. Before: +86, 2.75, 4 px. Promo (x 470-520, y 550-580): background 109, peak 151, +43, ratio 1.39. Our ratio is higher only because our sea is darker (pre-grade).
    • diorama02 near sea line: +6 over 73 (before +50). Grid coverage at 1920x1080: 0.74% at 20 km, 0.76% at 250 km (was 3.1% / 18.2% at 960x540).
  • Seam (M1). The close strips sit on even hex rows (x = 0 a hex centre); all 32 lines (grid off and on, four yaws) are asserted, max 0.004.
  • Rift. Arava line change 0.015 (bound 0.012, land opacity unchanged).
  • Profile (machine load average about 7, spreads up to 1 ms):
    • PROFILE summary visual=all pose=gameplay03 runs=3 median_avg_ms=1.71 spread_ms=0.97 median_p95_ms=3.00
    • PROFILE summary visual=-grid pose=gameplay03 runs=3 median_avg_ms=2.22 spread_ms=0.99 median_p95_ms=4.00
    • PROFILE summary visual=all pose=diorama02 runs=3 median_avg_ms=7.23 spread_ms=0.67 median_p95_ms=8.00
    • PROFILE summary visual=-grid pose=diorama02 runs=3 median_avg_ms=7.22 spread_ms=0.54 median_p95_ms=8.00
  • Build. WebGL, no Shader error; golden values OK in the checkpoint run.

Checkpoints

The images live in the owner's local captures/visual/ folder (git-ignored); owner verdicts are filled in as the checkpoints are reviewed.

Task Name Images Owner verdict
1 t01-before captures/visual/t01-before/-compare.png (owner)
2 t02-light captures/visual/t02-light/-compare.png (owner) Note: at the new 6 km hero zoom the view is filled by the 600-1,800 m placeholder marker cubes, so judge the zoom range, not the close-up art, until the settlement art track replaces them.
3 t03-tiltshift captures/visual/t03-tiltshift/-compare.png (re-shot in fix round 1: bands 0.13/0.17, radius 14/1080) (owner) Note: the map labels are still world-space text, so they blur with the map in the top and bottom bands until Task 9 moves them to UI Toolkit.
4 t04-grade captures/visual/t04-grade/-compare.png (with -deuteranopia / -protanopia copies) (owner) Note: saturation is above the promo (0.51/0.43 diorama, 0.55/0.38 gameplay); MapGrade.Saturation is the lever.
6 t06-terrain captures/visual/t06-terrain/-compare.png, editor/-compare.png (owner)
6 (fix 1) t06-terrain-r1 captures/visual/t06-terrain-r1/-compare.png, hero06judea.png, hero06negev.png, editor/-compare.png (owner)
7 t07-sea captures/visual/t07-sea/-compare.png, hero06judea.png, hero06negev.png, editor/-compare.png (owner)
7 (fix 1) t07-sea-r1 captures/visual/t07-sea-r1/-compare.png, hero06judea.png, hero06negev.png, editor/-compare.png (owner)
8 t08-grid captures/visual/t08-grid/-compare.png, hero06judea.png, hero06negev.png, editor/-compare.png (owner)
8 (fix 1) t08-grid-r1 captures/visual/t08-grid-r1/-compare.png, hero06judea.png, hero06negev.png, editor/-compare.png (owner)
9 t09-labels captures/visual/t09-labels/-compare.png (owner)
borders promo-borders-before, promo-borders-noproof, promo-borders-r2 captures/visual/promo-borders-before/-compare.png (main 807e300), captures/visual/promo-borders-noproof/-compare.png (the branch as played: no data, no border), captures/visual/promo-borders-r2/editor/-compare.png (the synthetic --borderproof checker, labelled on screen; editor renders only), all at d1973db (owner) Note: no real region data yet, so play shows no border (see "Region borders"). The earlier promo-borders-proof / -after rows ran on uncommitted work over 807e300 (old ring pattern).
merge t-merge captures/visual/t-merge/-compare.png, hero06judea.png, hero06negev.png, editor/-compare.png (before the grade re-tune: t-merge-pre, t-merge-r1) (owner) Note: the land is one dry-grass gold with no fields, orchards or cities yet; that gap is content (art track), not the grade.
merge (fix 1) t-merge-r3 captures/visual/t-merge-r3/-compare.png (t-merge-r2 = with the rejected border colour) (owner) Note: olive coastal plain and valleys; the selected hex is now the promo's ivory ring colour, gold stays the UI chrome's.
10 t10-tiers captures/visual/t10-tiers/-compare.png, hero06judea.png, hero06negev.png, editor/-compare.png; weak GPU: captures/visual/t10-swiftshader (owner) Note: render scale 1.0 on the web (sharper lines and pills); bloom, SSAO and the depth/opaque copies off; the colour is unchanged within 0.001.

Task 9: labels

Settlement names are UI Toolkit pills (MapLabelsPresenter, EB Garamond SemiBold on navy with a gold edge) in the click-through map-labels layer, placed in GameRoot.LateUpdate with an inline pixel translate and hidden while the game loads. The world-space TextMesh labels are gone.

Frame time, headless, DPR 1, res=1920x1080 scale=0.80 (?vprofile, machine_lock.py --exclusive --wait-unity):

Pose all -labels none all - (-labels) budget
gameplay03 2.30 (spread 0.15) 2.36 (0.06) 2.34 (0.04) -0.06 ms <= 1 + 0.15 ms: within
diorama02 6.82 (0.06) 6.55 (0.02) 6.61 (0.14) +0.27 ms <= 1 + 0.06 ms: within

The profile holds the camera still, so after the first frames Refresh returns early (unchanged view-projection, panel size and zoom); the cost measured is the drawing of the pills, not the per-frame placement while the camera moves.

Deviations from the plan text, each forced by a test:

  • Projection: RuntimePanelUtils.CameraTransformWorldToPanel flips y with Screen.height, which is wrong when the camera draws to a texture of another size (edit-mode tests: 600 px off). The presenter maps the camera's viewport point through camera.rect onto the panel's own size instead.
  • Declutter at the overview: the Tel Aviv - Lod cluster's markers sit 10-60 px apart there, and a town's pill could not avoid a neighbour's marker with any anchor (Rishon LeZion's pill ended 10 px under Jaffa's marker). A pill other than a city's is now also hidden when it comes within NeighbourClearPx (2 x LiftPx + MarginPx = 16 px) of another settlement's marker top. At 90 km this hides nothing (11 of 25 shown, as before); at the overview 10 of 25 are shown.
  • City pills: 20 px text and 8 px side padding at first. Fix round 1 restores the plan's 22 px and 10 px (see below).
  • ThePillIsEbGaramondOnNavyWithAGoldEdgeAndHiddenWhileLoading uses a 1920x1080 panel instead of UiTestKit's: the batch-mode screen gives that panel about 0.39 px per point, and UI Toolkit rounds the 1 px border up to one physical pixel (2.57 points).

Task 9 fix round 1 (review I1-I5, M1-M7)

  • Panels: pills are culled against the cached rects of the top bar, the info panel, the minimap and the event card, marked dirty by GeometryChangedEvent; a pill must lie wholly on screen, and a marker under a panel is no one's neighbour. A container resize re-places the pills before the repaint (M4).
  • Anchors: the USS anchor class is tried first, then lower right, lower left, lower centre and centred below the marker, as a pixel shift on the slot's translate. The legibility rule (more than 2 x own gap + MarginPx from every other visible marker top) applies to every pill; a city whose anchors all crowd a neighbour keeps the first that is clear of panels and pills. A pill hidden in the last refresh needs HysteresisPx = 2 more (M7).
  • Shown sets (MapLabelTests): 90 km: tel_aviv jaffa lod ramla petah_tikva rishon_lezion rehovot netanya (every visible marker); overview: jerusalem tel_aviv lod gaza beersheba petah_tikva netanya; diorama02 with the top-bar inset also labels Hadera (checkpoint).
  • City pills are back to the plan's 22 px and 10 px padding (towns 19 px), so cities read larger (M5). At the overview Tel Aviv's pill takes the lower-right anchor, over the sea.
  • Refresh allocates nothing (Is.Not.AllocatingGCMemory). Editor Mono, 240 refreshes while turning and zooming: LABELS refresh moving avg_us=27 p95_us=38.
  • ?vprofile&vmoving (FrameProfile, additive): every run restarts at its pose and turns 0.5 degrees per frame. Headless, DPR 1, res=1920x1080 scale=0.80:
Pose all -labels none all - (-labels) budget
gameplay03-moving 2.24 (spread 0.65) 2.28 (0.78) 2.44 (0.63) -0.04 ms <= 1 + 0.78 ms: within
diorama02-moving 10.21 (0.10) 10.20 (0.16) 10.10 (0.34) +0.01 ms <= 1 + 0.16 ms: within
gameplay03 (still) 2.47 (0.51) 2.44 (0.04) 2.35 (0.34) +0.03 ms within
diorama02 (still) 6.76 (0.38) 6.53 (0.29) 6.58 (0.41) +0.23 ms within

diorama02 costs about 3.4 ms more while turning than still, with or without labels (the rest of the scene).

Task 9 fix round 2 (re-review N1-N3, shown-set robustness)

  • A GeometryChangedEvent on a panel (event card, info panel, minimap, top bar) re-places the pills at once, inside the layout pass, so a card that opens never covers a pill for a frame (N1). The pills' summed sizes are part of the early-out key, so a pill that grows with the camera still is re-placed (N2). LABELS refresh moving avg_us=28 p95_us=38.
  • N3: WebGL ?smoke=5&navlog (SMOKE OK 5 turns 1948-06-19 d3fdd459...) shows no layout or style warning in the console; the only non-game lines are the existing WebGL: INVALID_ENUM: getInternalformatParameter driver notices.
  • ShownPillsNeverOverlapAndEachReadsAsItsOwn pins only unambiguous names (90 km: Tel Aviv, Petah Tikva; overview: Tel Aviv, Jerusalem, Gaza, Beersheba), at least 8 and 7 pills, Hadera or Netanya at the overview, and completeness: every visible marker not decluttered by zoom with no other visible marker within 2 pill heights + MarginPx is labelled. Disabling the anchor fallback turns it red (overview: shown pills Expected: greater than or equal to 7 But was: 6).

Merge point (Tracks A, B, C and main on visual-v1)

Merged in order vis-light, vis-labels, main, then vis-ground (d5037af). Conflicts: SettlementTests.cs (the seam test now runs Track A's graded/ungraded loop around Track B's pairs and even-row strips, grid off and on, four yaws: 64 asserted checks, max 0.006 ungraded against 0.012 and 0.005 graded against 0.024), docs/visual-v1.md, tools/README.md. Carries closed: the stale _label filter (Task 9 M8) is gone; Track B's colour and detail render tests (TerrainShaderTests, SeaTests, the seam and rift tests) run with the tilt-shift off (task-3 carry m6).

Counts. Unity edit mode: base 0a2a391 346; Track A 5fa3fa9 382 (+36); Track B ee555ac 371 (+25); Track C ebfdb73 361 (+15); main since the base +15 (UnityMcpGuardTests 10, GameInputsTests 5); minus 1 for VisualHarnessTests.AnAllowedRequireIsAcceptedAndAnUnknownOneFailsWithHowToAddIt, which both A and B carry (the Task 3 Step 0 cherry-pick). 346 + 36 + 25 + 15 + 15 - 1 = 436, measured total="436" passed="436" failed="0", no CS warning from Assets/(Scripts|Editor|Tests). Tools 143 + 1 (main) + 29 (Track B) = 173 passed, 1 skipped. Sim 625 (git diff 0a2a391 -- sim content shows only main's owner sign-off 25d2ef6; no visual track touches either).

Cross-track interactions found by the render tests.

  • GradeTests.TheSelectedHexStillStandsOutOnGradedLand failed at gameplay03 ungraded: deltaE76 18.2 (< 20; graded 23.1; Track A alone 24.5 / 32.3). Track B's coastal plain is dry-grass gold, the hue of the gold selection band. The selection is now warm ivory (1, 0.95, 0.80): 31.9 / 34.1 at gameplay03, 30.7 / 32.2 at diorama02.
  • Shadows (ReliefCastsShadowsAtMapScale, water on): 4.53% web, 6.99% PC (floor 2%; Track A alone 3.05 / 3.23). Haze tests: near 0.004, mid 0.028, far 0.054. Zoom extremes at MinDistance 6 km: grid covers 0.22% (1.57% at 250 km). Rift line change (bound 0.012): Beit She'an 0.096, Jericho 0.085, Arava 0.026.

Grade re-tune (Task 4 review I1). WebGL headless, 1920x1080 at DPR 1:

Shot Saturation Contrast Warmth
promo 02 / promo 03 0.425 / 0.382 0.200 / 0.203 0.145 / 0.064
diorama02, Saturation 16 (t-merge-pre) 0.5345 0.2058 0.1405
gameplay03, Saturation 16 0.6629 0.2147 0.1509
diorama02, Saturation 4 (t-merge-r1) 0.4860 0.2070 0.1181
diorama02, Saturation 0 (t-merge) 0.4729 0.2074 0.1099
gameplay03, Saturation 0 0.5822 0.2168 0.1078

Diorama02 now meets both bars (warmth in [0.087, 0.145], saturation <= 0.48). TheGradeLiftsSaturation: 0.568 -> 0.598 (+5.3%, bar 5%; replaced in fix round 1). Gameplay03 stays 0.20 over its promo in saturation and 0.04 in warmth: the promo's frame is green fields and orchards, which no global grade can supply.

Grid (Task 8 N2). Land opacity 0.25 -> 0.35: at 0.25 the graded land lines were invisible on the plain at both poses. At 0.35 gameplay03 shows a faint land lattice and diorama02 a faint near-field one, as in the promos. Fade thresholds unchanged.

Web. BUILD Succeeded WebGL -> Builds/Web-disabled: 77.1 MB on disk, 0 errors; ?golden&smoke=5 PASS with all three values from docs/determinism.md. Boot (headed, machine_lock --exclusive, median of 3): --cpu 1 map_build 456 / first_frame 3585 ms (budgets 743 / 5321); --cpu 4 1863 / 6952 ms (budgets 2720 / 10445). Both within.

Merge point fix round 1 (merge review)

  • I1. TheGradeLiftsSaturation is replaced by TheGradeKeepsSaturationNearTheUngradedFrame: graded/ungraded saturation at gameplay03 in [0.98, 1.12], and MapGrade.Saturation in [-10, 4]. Measured x1.042.
  • Palette (art gap a). coastal_plain Lit(0.52, 0.54, 0.33), valley Lit(0.50, 0.53, 0.33). WebGL:
Shot Saturation Contrast Warmth
promo 02 / promo 03 0.425 / 0.382 0.200 / 0.203 0.145 / 0.064
diorama02, t-merge -> t-merge-r3 0.4729 -> 0.4694 0.2074 -> 0.2034 0.1099 -> 0.1045
gameplay03, t-merge -> t-merge-r3 0.5822 -> 0.5646 0.2168 -> 0.1959 0.1078 -> 0.0774

EditMode 436/436 with no threshold change: selection vs land 33.4 / 35.5 (gameplay03), colour-blind selection-land 34.5 / 36.4, rift lines 0.121 / 0.104 / 0.026.

  • Border land colour (art gap b): not taken. BorderLandColour (0.76, 0.66, 0.47) (t-merge-r2) rendered the south-west slab at (0.69, 0.58, 0.37), against (0.56, 0.50, 0.37) before. The terrain next to it renders (0.82, 0.69, 0.43), so the step shrank. But the skirt is 17% of diorama02, and its saturation rose 0.34 -> 0.47. That added +0.021 to the frame, which reached 0.488 against the 0.48 bar (warmth 0.129). Reverted. A lower-chroma colour of the same luma is NOT the next step: the final ruling is in "Open items for the art track" (fix the table/frame and the skirt in the terrain's own look, not a colour patch).

Tier decisions (Task 10)

Measured under machine_lock.py --exclusive --wait-unity, each run started at a 1-minute load average of 3.0 or less (the machine idles at 3-4: other sessions; every run's before/after load is in the Task 10 report). Step 1 is the merged build at render scale 0.8 (8d45a93, which adds the bloom toggle); the candidates are that build with one URP value changed; "final" is the build with the decisions applied. A cost is median(all) - median(-name); over budget only beyond budget + the larger spread.

Setting Web Desktop Evidence
Path Forward Forward+ research notes 2.1 (Forward+ on WebGL reported unstable)
Depth / opaque copies off / off off / off (were on) no visual v1 effect samples them; the Gaussian DoF that did is retired (Task 3)
Render scale 1.0 (was 0.8) 1.0 all three rules hold at 1.0 (headless p95 <= 16.7 ms, headed <= 20 ms, DPR 2 <= 33 ms): headless PROFILE summary visual=all pose=gameplay03 runs=3 median_avg_ms=2.58 spread_ms=0.18 median_p95_ms=5.00
PROFILE summary visual=all pose=diorama02 runs=3 median_avg_ms=9.96 spread_ms=0.05 median_p95_ms=11.00; headed PROFILE summary visual=all pose=gameplay03 runs=3 median_avg_ms=2.39 spread_ms=0.10 median_p95_ms=5.00
PROFILE summary visual=all pose=diorama02 runs=3 median_avg_ms=10.14 spread_ms=0.20 median_p95_ms=11.00; DPR 2 PROFILE summary visual=all pose=gameplay03 runs=3 median_avg_ms=4.01 spread_ms=0.00 median_p95_ms=5.00
PROFILE summary visual=all pose=diorama02 runs=3 median_avg_ms=18.59 spread_ms=0.36 median_p95_ms=19.00
Fill-rate stress (DPR 2, headed, final at 1.0) PROFILE summary visual=all pose=gameplay03 runs=3 median_avg_ms=3.78 spread_ms=0.05 median_p95_ms=5.00
PROFILE summary visual=all pose=diorama02 runs=3 median_avg_ms=18.08 spread_ms=0.06 median_p95_ms=20.00
— a stand-in for a GPU with a quarter of the fill rate; under the 33 ms floor with 13 ms to spare
Boot BOOT map_build_ms median=432 runs=3 cpu=1
BOOT first_frame_ms median=3563 runs=3 cpu=1
BOOT map_build_ms median=1751 runs=3 cpu=4
BOOT first_frame_ms median=6797 runs=3 cpu=4
BOOT map_build_ms median=426 runs=3 cpu=1
BOOT first_frame_ms median=3477 runs=3 cpu=1
BOOT map_build_ms median=1746 runs=3 cpu=4
BOOT first_frame_ms median=6696 runs=3 cpu=4 (Step 1, then final)
— each within max(300 ms, 50%) of its Task 1 baseline (443 / 3547; 1813 / 6963 ms); both runs exit 0
Shadows 1 cascade, 2048, hard 4 cascades, 2048, soft -shadows at diorama02: Step 1 headless 10.32 - 7.36 = 2.96 ms (spread 0.32: over 2.82), headed 2.90 (0.41: within), DPR 2 4.99 (0.23); final headless 2.89 (0.41: within 2.91), headed 2.92 (0.29: over 2.79), DPR 2 4.80. A 1024 build measured 2.84 (headless, spread 0.09), 2.94 (headed) and 5.03 (DPR 2): (1024 lines, headless: PROFILE summary visual=all pose=diorama02 runs=3 median_avg_ms=10.08 spread_ms=0.09 median_p95_ms=11.00
PROFILE summary visual=-shadows pose=diorama02 runs=3 median_avg_ms=7.24 spread_ms=0.03 median_p95_ms=9.00
headed: PROFILE summary visual=all pose=diorama02 runs=3 median_avg_ms=10.20 spread_ms=0.28 median_p95_ms=11.00
PROFILE summary visual=-shadows pose=diorama02 runs=3 median_avg_ms=7.26 spread_ms=0.13 median_p95_ms=9.00
DPR 2: PROFILE summary visual=all pose=diorama02 runs=3 median_avg_ms=18.34 spread_ms=0.21 median_p95_ms=19.00
PROFILE summary visual=-shadows pose=diorama02 runs=3 median_avg_ms=13.31 spread_ms=0.10 median_p95_ms=14.00): the map's resolution is not the cost (the caster pass redraws the whole map's geometry at the overview), so 1024 would blur the relief for nothing. 2048 kept; the overrun (about 0.1-0.4 ms beyond budget + spread at the 250 km overview only) is named here. gameplay03: 0.33 ms final
Tilt-shift downsample 2 downsample 2 -tiltshift cost: Step 1 headless 0.09 / 0.09 ms (gameplay03 / diorama02), final headless 0.03 / -0.04, DPR 2 0.29 / -0.04: far under 1.5 ms
Tilt-shift bottom band unchanged (HalfWidth 0.13 / Feather 0.17) same final gameplay03 shot against ?notiltshift, gradient energy per 1% row band (columns 30-80%): fully blurred (ratio <= 0.05-0.09) from under the bar to about 26% and from about 78-83% to the bottom, so the bottom band is 17-22% against the promo's ~15%. The 2-7 point gap is the one Task 3 offered the owner (bottom-only 0.16 / 0.19). Owner's ruling (2026-10-13, issue #390): the bands stay as they are, except that the far band is lighter at the gameplay zoom ("Issue #390"); the bottom band is not changed. Owner's ruling (2026-10-14, issue #458): the bottom band's height stays, but it is lighter at the gameplay zoom and about gone at the overview ("Issue #458")
Bloom off (was the URP template's 0.25 above 1) off rendered on both tiers at both poses it changed at most 2/255 of any pixel (mean 0.25/255 at diorama02, warmth +0.0003, measured with a temporary bloom toggle, commit 8d45a93, removed in fix round 1 once bloom shipped off); -bloom cost Step 1 headless 0.37 ms (diorama02), headed 0.18, macOS 0.17 / 0.13. Off in the pipeline profile (Settings/SampleSceneProfile), which keeps its Neutral tonemapper (the stripping guarantee, GradeTests)
SSAO none off (was on) metre radius cannot reach km relief; baked AO replaces it (Tasks 5-6)
Grade, vignette, haze, grid, water, labels on on Step 1 headless costs (gameplay03 / diorama02, ms): grade -0.02 / 0.27, haze 0.08 / 0.11, grid 0.07 / -0.14, water 0.04 / 0.14, labels -0.14 / -0.02; final headless: grade 0.00 / -0.03, haze -0.03 / -0.10, grid 0.00 / -0.01, water -0.05 / -0.06, labels -0.01 / 0.01; all within budget. The grid, unmeasured on a quiet machine since Task 8, costs nothing measurable
Baked shading texture block-compressed block-compressed review minor 9: recorded, not changed (mobile browsers are outside the 1920x1080 target)
WebGPU, GPU Resident Drawer, compute not used not used research notes 2.1 (an M5 evaluation)
Weak GPU SwiftShader run passes golden and shots — captures/visual/t10-swiftshader (both GOLDEN OK lines and SHOT done; gameplay03 saturation / contrast / warmth 0.5656 / 0.1960 / 0.0773 against the GPU's 0.5655 / 0.1960 / 0.0771)
Runtime quality fallback (issue #44) on off see "Runtime quality fallback" below: when the frame-time p95 stays over the 30 fps floor, render scale 1.0 -> 0.9 -> 0.8, then the tilt-shift off, then the town building models' far-off copies without shadows at every zoom; back up in reverse

Colour after the decisions (WebGL ?shot, t10-tiers against t-merge-r3): diorama02 saturation 0.4694 -> 0.4686, contrast 0.2034 -> 0.2032, warmth 0.1045 -> 0.1038 (bars: warmth 0.087-0.145, saturation <= 0.48); gameplay03 saturation 0.5646 -> 0.5655 (+0.0009, away from the promo's 0.382 by a rounding-level step: the 1:1 render), contrast 0.1959 -> 0.1960, warmth 0.0774 -> 0.0771. The sea, palette and grade are untouched.

The built macOS player (unity/Builds/Mac/Statecraft.app) does not run in the background (runInBackground is off): its -vprofile waits at GAME OK until its window is in front. Bring that player forward (Task 10 used open -a unity/Builds/Mac/Statecraft.app); never script, open -a or AppleScript the Unity editor (team rule).

The -bloom lines below come from the temporary toggle (8d45a93); in the final runs bloom already ships off, so -bloom equals all there. At gameplay03 in Step 1 (scale 0.8), -bloom, -shadows and none all read about 0.4 ms above all (headless 2.51 / 2.40 / 2.40 against 2.03; headed 2.34 / 2.37 / 2.24 against 1.85). none cannot really be slower than all; the pose runs at about 2 ms, at the WebGL clock's 1 ms resolution, and the readings look bimodal (2.0 / ~2.4 ms). Unexplained; it changes no decision (every gameplay03 cost is far inside its budget).

Runtime quality fallback (issue #44)

The Deferred row "Runtime quality fallback" (review I1), sized by the DPR-2 stress numbers above. QualityFallback (pure logic) and QualityFallbackDriver (feeds it Time.unscaledDeltaTime, applies its step), attached by CameraRig.Init through the map's teardown like the visual toggles.

  • Steps: level 0 is the tier (scale 1.0, tilt-shift on); 1: scale 0.9; 2: scale 0.8 (the research's 0.8-1.0 dynamic range, notes 2.3); 3: scale 0.8 and the tilt-shift off; 4 (v0.3.7): as 3, and the town building models draw their far-off copies without shadows at every zoom (TownModelLayer.LowTier), the 30 km opening too. Since v0.3.5 the opening sits below the close switch (40 km), so the originals and their shadows draw from the first frame; the nightly's start-screen p95 on its software GPU went from 35-88 ms (v0.3.4) to 320-690 ms (v0.3.5, v0.3.6). The opening and the switch stay; only a device still over the floor after step 3 loses the originals. One step per decision, up in reverse. The scale is the active URP asset's renderScale (never above the tier's own); the tilt-shift uses its own TiltShift.FallbackOff, so stepping up never turns on a tilt-shift the player (?notiltshift) or a profile turned off. The map's teardown puts all three back, and a fixed quality choice clears the last two.
  • Window: 3 s, back to back, one decision each. The p95 of a 3 s window is over only when at least 5% of its frames are (5 of 90 at 30 fps), so a one-off hitch (End Turn's redraw, a GC, a first shader compile) never trips it and a real shortfall does within 3 s. Nothing is measured in the first 3 s after the map is built, or 0.5 s after each change; a frame over 1 s (a hidden tab coming back) drops the window; a window needs at least 3 frames, which a full 3 s window of counted frames (each at most 1 s) always has, so the slowest devices (1-3 fps) step down too (fix round 1, review M2; a minimum of 10 had left them with no decision).
  • Down: p95 > 34 ms. The floor is 33 ms (30 fps). The WebGL frame clock is quantised to about 1 ms, so exactly 30 fps (a 30 Hz cap from a display or a battery saver, or vsync halving 60 Hz) reads as 33 and 34 ms frames: that meets the floor, and a lower quality could not raise a capped rate.
  • **Up: 3 windows in a row (9 s) with p95 < 20 ms; hysteresis.** One step up draws at most 1.27 times the pixels (0.8 -> 0.9, 0.9 -> 1.0), so a fill-bound 20 ms frame lands near 25 ms, under the floor; the tilt-shift costs under 1.5 ms (-tiltshift above: 0.29 ms at DPR 2). Windows between 20 and 34 ms change nothing and restart the count. A step up undone within two windows doubles the windows the next one needs (3, 6, 12 ... at most 48, 2.4 minutes), so a device on the edge settles instead of flipping.
  • Web tier only. QualityFallback.Wanted is true only in the WebGL player: the desktop tier is the development and capture target (its renders and -vprofile must show the tier's own look), and in the editor a scale change would change the URP asset. Even on WebGL it stays off under ?vprofile, ?profile and ?shot (instruments that must see the tier as configured) and with ?nofallback; every change logs QUALITY level=<n> scale=<s> tiltshift=<on|off> towns=<close|far> p95_ms=<ms>.
  • The player's quality choice (#40). Automatic in Settings is this behaviour. High (the PC tier at render scale 1.0) and Low (the Mobile tier at 0.8) fix the quality: the fallback holds while one is set (QualityFallbackDriver.Feed), and every switch starts it again from the top (GraphicsQuality.Changed).
  • Presentation only: it reads no sim state and writes only the render scale, the tilt-shift flag and the towns' low tier, so turns, hashes and saves cannot change. The goldens and the 5-turn smoke match with the original #44 fallback in that build: uv run tools/web_check.py unity/Builds/Web-disabled --query "golden&smoke=5" with the three --expect values from docs/determinism.md: PASS. This is #44's evidence; the v0.3.7 step changes no sim code.
  • Tests (QualityFallbackTests, EditMode, synthetic frame times): the step order and range, one step per window and not before a full window, nothing during start-up, hitches, a pause and 33/34 ms frames do not trigger, recovery in reverse only after 3 good windows, the dead band holds and restarts the count, a failed step up doubles the wait (capped); integration: no driver in the editor; with the web tier's answer forced, MapScene.Build attaches one, its steps reach the URP asset, TiltShift.FallbackOff (not TiltShift.Enabled) and TownModelLayer.LowTier, and a rebuild or teardown removes it and restores the scale and the town originals. TownModelResidencyTests checks that the low tier at the opening draws Tel Aviv's far copies with no shadow casters, and that clearing it puts the originals back.
  • Original #44 evidence in a real browser (predates the towns= field and level 4): headed Chrome for Testing, ANGLE Metal, 1920x1080 at DPR 1, under machine_lock.py --exclusive --wait-unity, a scratch CDP script on headed.Browser: 30 s idle at --cpu 1, no QUALITY line. With CPU throttling at 20x from t = 30 s: QUALITY level=1 scale=0.90 tiltshift=on p95_ms=97.00, level=2 scale=0.80 tiltshift=on p95_ms=97.00, level=3 scale=0.80 tiltshift=off p95_ms=107.00 (level 3 at t = 39.6 s). Throttling lifted at t = 49.6 s: level=2 ... tiltshift=on p95_ms=17.00, level=1 scale=0.90, level=0 scale=1.00 (t = 77.4 s, about 9.5 s apart), then no line in the next 15 s. At DPR 2 (the fill-rate stress proxy, frame rate uncapped) 60 s idle: no QUALITY line, as the DPR-2 p95 (20 ms at diorama02) is under the floor.
  • Boot with it (headed.py boot --runs 3, on top of the packed heightmap): --cpu 1 BOOT map_build_ms median=472, BOOT first_frame_ms median=3609; --cpu 4 map_build_ms median=1983, first_frame_ms median=7158: within the boot budget (exit 0).

Final measurements

Step 1 headless (scale 0.80):

  • PROFILE summary visual=all pose=gameplay03 runs=3 median_avg_ms=2.03 spread_ms=0.16 median_p95_ms=3.00
  • PROFILE summary visual=-bloom pose=gameplay03 runs=3 median_avg_ms=2.51 spread_ms=0.58 median_p95_ms=5.00
  • PROFILE summary visual=-grade pose=gameplay03 runs=3 median_avg_ms=2.05 spread_ms=0.25 median_p95_ms=3.00
  • PROFILE summary visual=-grid pose=gameplay03 runs=3 median_avg_ms=1.96 spread_ms=0.27 median_p95_ms=3.00
  • PROFILE summary visual=-haze pose=gameplay03 runs=3 median_avg_ms=1.95 spread_ms=0.53 median_p95_ms=3.00
  • PROFILE summary visual=-labels pose=gameplay03 runs=3 median_avg_ms=2.17 spread_ms=0.31 median_p95_ms=4.00
  • PROFILE summary visual=-shadows pose=gameplay03 runs=3 median_avg_ms=2.40 spread_ms=0.03 median_p95_ms=4.00
  • PROFILE summary visual=-tiltshift pose=gameplay03 runs=3 median_avg_ms=1.94 spread_ms=0.21 median_p95_ms=3.00
  • PROFILE summary visual=-water pose=gameplay03 runs=3 median_avg_ms=1.99 spread_ms=0.09 median_p95_ms=3.00
  • PROFILE summary visual=none pose=gameplay03 runs=3 median_avg_ms=2.40 spread_ms=0.09 median_p95_ms=4.00
  • PROFILE summary visual=all pose=diorama02 runs=3 median_avg_ms=10.32 spread_ms=0.19 median_p95_ms=11.00
  • PROFILE summary visual=-bloom pose=diorama02 runs=3 median_avg_ms=9.95 spread_ms=0.02 median_p95_ms=11.00
  • PROFILE summary visual=-grade pose=diorama02 runs=3 median_avg_ms=10.05 spread_ms=0.51 median_p95_ms=11.00
  • PROFILE summary visual=-grid pose=diorama02 runs=3 median_avg_ms=10.46 spread_ms=0.84 median_p95_ms=12.00
  • PROFILE summary visual=-haze pose=diorama02 runs=3 median_avg_ms=10.21 spread_ms=0.06 median_p95_ms=11.00
  • PROFILE summary visual=-labels pose=diorama02 runs=3 median_avg_ms=10.34 spread_ms=0.47 median_p95_ms=12.00
  • PROFILE summary visual=-shadows pose=diorama02 runs=3 median_avg_ms=7.36 spread_ms=0.32 median_p95_ms=10.00
  • PROFILE summary visual=-tiltshift pose=diorama02 runs=3 median_avg_ms=10.23 spread_ms=0.21 median_p95_ms=11.00
  • PROFILE summary visual=-water pose=diorama02 runs=3 median_avg_ms=10.18 spread_ms=0.08 median_p95_ms=11.00
  • PROFILE summary visual=none pose=diorama02 runs=3 median_avg_ms=7.16 spread_ms=0.33 median_p95_ms=9.00

Run line: PROFILE visual=all pose=gameplay03 res=1920x1080 scale=0.80 rep=0 frames=300 avg_ms=2.05 p95_ms=3.00

Step 1 headed, DPR 1 (scale 0.80):

  • PROFILE summary visual=all pose=gameplay03 runs=3 median_avg_ms=1.85 spread_ms=0.08 median_p95_ms=3.00
  • PROFILE summary visual=-bloom pose=gameplay03 runs=3 median_avg_ms=2.34 spread_ms=0.48 median_p95_ms=4.00
  • PROFILE summary visual=-grade pose=gameplay03 runs=3 median_avg_ms=2.00 spread_ms=0.17 median_p95_ms=3.00
  • PROFILE summary visual=-grid pose=gameplay03 runs=3 median_avg_ms=1.85 spread_ms=0.04 median_p95_ms=3.00
  • PROFILE summary visual=-haze pose=gameplay03 runs=3 median_avg_ms=1.86 spread_ms=0.02 median_p95_ms=3.00
  • PROFILE summary visual=-labels pose=gameplay03 runs=3 median_avg_ms=1.84 spread_ms=0.35 median_p95_ms=3.00
  • PROFILE summary visual=-shadows pose=gameplay03 runs=3 median_avg_ms=2.37 spread_ms=0.02 median_p95_ms=4.00
  • PROFILE summary visual=-tiltshift pose=gameplay03 runs=3 median_avg_ms=1.85 spread_ms=0.86 median_p95_ms=3.00
  • PROFILE summary visual=-water pose=gameplay03 runs=3 median_avg_ms=1.86 spread_ms=0.11 median_p95_ms=3.00
  • PROFILE summary visual=none pose=gameplay03 runs=3 median_avg_ms=2.24 spread_ms=0.04 median_p95_ms=4.00
  • PROFILE summary visual=all pose=diorama02 runs=3 median_avg_ms=10.27 spread_ms=0.41 median_p95_ms=11.00
  • PROFILE summary visual=-bloom pose=diorama02 runs=3 median_avg_ms=10.09 spread_ms=0.47 median_p95_ms=11.00
  • PROFILE summary visual=-grade pose=diorama02 runs=3 median_avg_ms=10.18 spread_ms=0.09 median_p95_ms=11.00
  • PROFILE summary visual=-grid pose=diorama02 runs=3 median_avg_ms=10.70 spread_ms=0.64 median_p95_ms=12.00
  • PROFILE summary visual=-haze pose=diorama02 runs=3 median_avg_ms=10.51 spread_ms=0.52 median_p95_ms=12.00
  • PROFILE summary visual=-labels pose=diorama02 runs=3 median_avg_ms=10.60 spread_ms=0.49 median_p95_ms=12.00
  • PROFILE summary visual=-shadows pose=diorama02 runs=3 median_avg_ms=7.37 spread_ms=0.09 median_p95_ms=10.00
  • PROFILE summary visual=-tiltshift pose=diorama02 runs=3 median_avg_ms=10.52 spread_ms=0.33 median_p95_ms=12.00
  • PROFILE summary visual=-water pose=diorama02 runs=3 median_avg_ms=10.55 spread_ms=0.22 median_p95_ms=12.00
  • PROFILE summary visual=none pose=diorama02 runs=3 median_avg_ms=7.15 spread_ms=0.11 median_p95_ms=9.00

Step 1 headed, DPR 2 (scale 0.80):

  • PROFILE summary visual=all pose=gameplay03 runs=3 median_avg_ms=3.49 spread_ms=0.15 median_p95_ms=5.00
  • PROFILE summary visual=-bloom pose=gameplay03 runs=3 median_avg_ms=3.28 spread_ms=0.02 median_p95_ms=5.00
  • PROFILE summary visual=-grade pose=gameplay03 runs=3 median_avg_ms=3.41 spread_ms=0.09 median_p95_ms=5.00
  • PROFILE summary visual=-grid pose=gameplay03 runs=3 median_avg_ms=3.36 spread_ms=0.21 median_p95_ms=5.00
  • PROFILE summary visual=-haze pose=gameplay03 runs=3 median_avg_ms=3.49 spread_ms=0.23 median_p95_ms=5.00
  • PROFILE summary visual=-labels pose=gameplay03 runs=3 median_avg_ms=3.47 spread_ms=0.01 median_p95_ms=5.00
  • PROFILE summary visual=-shadows pose=gameplay03 runs=3 median_avg_ms=3.11 spread_ms=0.07 median_p95_ms=5.00
  • PROFILE summary visual=-tiltshift pose=gameplay03 runs=3 median_avg_ms=3.36 spread_ms=0.09 median_p95_ms=5.00
  • PROFILE summary visual=-water pose=gameplay03 runs=3 median_avg_ms=3.48 spread_ms=0.05 median_p95_ms=5.00
  • PROFILE summary visual=none pose=gameplay03 runs=3 median_avg_ms=2.57 spread_ms=0.21 median_p95_ms=4.00
  • PROFILE summary visual=all pose=diorama02 runs=3 median_avg_ms=18.24 spread_ms=0.23 median_p95_ms=19.00
  • PROFILE summary visual=-bloom pose=diorama02 runs=3 median_avg_ms=18.12 spread_ms=1.06 median_p95_ms=20.00
  • PROFILE summary visual=-grade pose=diorama02 runs=3 median_avg_ms=18.35 spread_ms=0.27 median_p95_ms=19.00
  • PROFILE summary visual=-grid pose=diorama02 runs=3 median_avg_ms=18.27 spread_ms=0.11 median_p95_ms=19.00
  • PROFILE summary visual=-haze pose=diorama02 runs=3 median_avg_ms=18.24 spread_ms=0.11 median_p95_ms=19.00
  • PROFILE summary visual=-labels pose=diorama02 runs=3 median_avg_ms=18.23 spread_ms=0.11 median_p95_ms=19.00
  • PROFILE summary visual=-shadows pose=diorama02 runs=3 median_avg_ms=13.25 spread_ms=0.16 median_p95_ms=14.00
  • PROFILE summary visual=-tiltshift pose=diorama02 runs=3 median_avg_ms=18.23 spread_ms=0.19 median_p95_ms=19.00
  • PROFILE summary visual=-water pose=diorama02 runs=3 median_avg_ms=18.28 spread_ms=0.17 median_p95_ms=19.00
  • PROFILE summary visual=none pose=diorama02 runs=3 median_avg_ms=13.05 spread_ms=0.07 median_p95_ms=14.00

Step 1 macOS IL2CPP player (PC tier as merged: SSAO, depth and opaque copies, bloom on):

  • PROFILE summary visual=all pose=gameplay03 runs=3 median_avg_ms=1.45 spread_ms=0.01 median_p95_ms=1.48
  • PROFILE summary visual=-bloom pose=gameplay03 runs=3 median_avg_ms=1.32 spread_ms=0.03 median_p95_ms=1.34
  • PROFILE summary visual=-grade pose=gameplay03 runs=3 median_avg_ms=1.45 spread_ms=0.00 median_p95_ms=1.47
  • PROFILE summary visual=-grid pose=gameplay03 runs=3 median_avg_ms=1.45 spread_ms=0.00 median_p95_ms=1.47
  • PROFILE summary visual=-haze pose=gameplay03 runs=3 median_avg_ms=1.45 spread_ms=0.00 median_p95_ms=1.48
  • PROFILE summary visual=-labels pose=gameplay03 runs=3 median_avg_ms=1.45 spread_ms=0.00 median_p95_ms=1.47
  • PROFILE summary visual=-shadows pose=gameplay03 runs=3 median_avg_ms=1.16 spread_ms=0.00 median_p95_ms=1.19
  • PROFILE summary visual=-tiltshift pose=gameplay03 runs=3 median_avg_ms=1.38 spread_ms=0.00 median_p95_ms=1.40
  • PROFILE summary visual=-water pose=gameplay03 runs=3 median_avg_ms=1.35 spread_ms=0.00 median_p95_ms=1.37
  • PROFILE summary visual=none pose=gameplay03 runs=3 median_avg_ms=0.88 spread_ms=0.00 median_p95_ms=0.90
  • PROFILE summary visual=all pose=diorama02 runs=3 median_avg_ms=3.73 spread_ms=0.01 median_p95_ms=3.98
  • PROFILE summary visual=-bloom pose=diorama02 runs=3 median_avg_ms=3.56 spread_ms=0.05 median_p95_ms=3.70
  • PROFILE summary visual=-grade pose=diorama02 runs=3 median_avg_ms=3.71 spread_ms=0.06 median_p95_ms=3.82
  • PROFILE summary visual=-grid pose=diorama02 runs=3 median_avg_ms=3.70 spread_ms=0.04 median_p95_ms=3.82
  • PROFILE summary visual=-haze pose=diorama02 runs=3 median_avg_ms=3.72 spread_ms=0.04 median_p95_ms=3.90
  • PROFILE summary visual=-labels pose=diorama02 runs=3 median_avg_ms=3.69 spread_ms=0.03 median_p95_ms=3.81
  • PROFILE summary visual=-shadows pose=diorama02 runs=3 median_avg_ms=2.34 spread_ms=0.02 median_p95_ms=2.41
  • PROFILE summary visual=-tiltshift pose=diorama02 runs=3 median_avg_ms=3.66 spread_ms=0.02 median_p95_ms=3.79
  • PROFILE summary visual=-water pose=diorama02 runs=3 median_avg_ms=3.68 spread_ms=0.04 median_p95_ms=3.81
  • PROFILE summary visual=none pose=diorama02 runs=3 median_avg_ms=2.18 spread_ms=0.02 median_p95_ms=2.30

Final headless (scale 1.00):

  • PROFILE summary visual=all pose=gameplay03 runs=3 median_avg_ms=2.70 spread_ms=0.02 median_p95_ms=5.00
  • PROFILE summary visual=-bloom pose=gameplay03 runs=3 median_avg_ms=2.72 spread_ms=0.03 median_p95_ms=5.00
  • PROFILE summary visual=-grade pose=gameplay03 runs=3 median_avg_ms=2.70 spread_ms=0.10 median_p95_ms=5.00
  • PROFILE summary visual=-grid pose=gameplay03 runs=3 median_avg_ms=2.70 spread_ms=0.07 median_p95_ms=5.00
  • PROFILE summary visual=-haze pose=gameplay03 runs=3 median_avg_ms=2.73 spread_ms=0.11 median_p95_ms=5.00
  • PROFILE summary visual=-labels pose=gameplay03 runs=3 median_avg_ms=2.71 spread_ms=0.04 median_p95_ms=5.00
  • PROFILE summary visual=-shadows pose=gameplay03 runs=3 median_avg_ms=2.37 spread_ms=0.02 median_p95_ms=4.00
  • PROFILE summary visual=-tiltshift pose=gameplay03 runs=3 median_avg_ms=2.67 spread_ms=0.05 median_p95_ms=5.00
  • PROFILE summary visual=-water pose=gameplay03 runs=3 median_avg_ms=2.75 spread_ms=0.05 median_p95_ms=5.00
  • PROFILE summary visual=none pose=gameplay03 runs=3 median_avg_ms=2.34 spread_ms=0.05 median_p95_ms=4.00
  • PROFILE summary visual=all pose=diorama02 runs=3 median_avg_ms=10.06 spread_ms=0.41 median_p95_ms=11.00
  • PROFILE summary visual=-bloom pose=diorama02 runs=3 median_avg_ms=10.05 spread_ms=0.34 median_p95_ms=11.00
  • PROFILE summary visual=-grade pose=diorama02 runs=3 median_avg_ms=10.09 spread_ms=0.27 median_p95_ms=11.00
  • PROFILE summary visual=-grid pose=diorama02 runs=3 median_avg_ms=10.07 spread_ms=0.25 median_p95_ms=11.00
  • PROFILE summary visual=-haze pose=diorama02 runs=3 median_avg_ms=10.16 spread_ms=0.07 median_p95_ms=11.00
  • PROFILE summary visual=-labels pose=diorama02 runs=3 median_avg_ms=10.05 spread_ms=0.07 median_p95_ms=11.00
  • PROFILE summary visual=-shadows pose=diorama02 runs=3 median_avg_ms=7.17 spread_ms=0.01 median_p95_ms=8.00
  • PROFILE summary visual=-tiltshift pose=diorama02 runs=3 median_avg_ms=10.10 spread_ms=0.04 median_p95_ms=11.00
  • PROFILE summary visual=-water pose=diorama02 runs=3 median_avg_ms=10.12 spread_ms=0.07 median_p95_ms=11.00
  • PROFILE summary visual=none pose=diorama02 runs=3 median_avg_ms=7.17 spread_ms=0.05 median_p95_ms=9.00

Run line: PROFILE visual=all pose=gameplay03 res=1920x1080 scale=1.00 rep=0 frames=300 avg_ms=2.71 p95_ms=5.00

Final headed, DPR 1:

  • PROFILE summary visual=all pose=gameplay03 runs=3 median_avg_ms=2.72 spread_ms=0.06 median_p95_ms=5.00
  • PROFILE summary visual=-bloom pose=gameplay03 runs=3 median_avg_ms=2.66 spread_ms=0.05 median_p95_ms=5.00
  • PROFILE summary visual=-grade pose=gameplay03 runs=3 median_avg_ms=2.68 spread_ms=0.01 median_p95_ms=5.00
  • PROFILE summary visual=-grid pose=gameplay03 runs=3 median_avg_ms=2.70 spread_ms=0.01 median_p95_ms=5.00
  • PROFILE summary visual=-haze pose=gameplay03 runs=3 median_avg_ms=2.71 spread_ms=0.07 median_p95_ms=5.00
  • PROFILE summary visual=-labels pose=gameplay03 runs=3 median_avg_ms=2.67 spread_ms=0.03 median_p95_ms=5.00
  • PROFILE summary visual=-shadows pose=gameplay03 runs=3 median_avg_ms=2.37 spread_ms=0.01 median_p95_ms=4.00
  • PROFILE summary visual=-tiltshift pose=gameplay03 runs=3 median_avg_ms=2.69 spread_ms=0.09 median_p95_ms=5.00
  • PROFILE summary visual=-water pose=gameplay03 runs=3 median_avg_ms=2.67 spread_ms=0.10 median_p95_ms=5.00
  • PROFILE summary visual=none pose=gameplay03 runs=3 median_avg_ms=2.31 spread_ms=0.01 median_p95_ms=4.00
  • PROFILE summary visual=all pose=diorama02 runs=3 median_avg_ms=9.98 spread_ms=0.29 median_p95_ms=10.00
  • PROFILE summary visual=-bloom pose=diorama02 runs=3 median_avg_ms=10.03 spread_ms=0.33 median_p95_ms=11.00
  • PROFILE summary visual=-grade pose=diorama02 runs=3 median_avg_ms=9.95 spread_ms=0.31 median_p95_ms=11.00
  • PROFILE summary visual=-grid pose=diorama02 runs=3 median_avg_ms=9.97 spread_ms=0.28 median_p95_ms=10.00
  • PROFILE summary visual=-haze pose=diorama02 runs=3 median_avg_ms=10.03 spread_ms=0.07 median_p95_ms=11.00
  • PROFILE summary visual=-labels pose=diorama02 runs=3 median_avg_ms=10.00 spread_ms=0.04 median_p95_ms=11.00
  • PROFILE summary visual=-shadows pose=diorama02 runs=3 median_avg_ms=7.06 spread_ms=0.04 median_p95_ms=8.00
  • PROFILE summary visual=-tiltshift pose=diorama02 runs=3 median_avg_ms=10.01 spread_ms=0.04 median_p95_ms=11.00
  • PROFILE summary visual=-water pose=diorama02 runs=3 median_avg_ms=10.06 spread_ms=0.06 median_p95_ms=11.00
  • PROFILE summary visual=none pose=diorama02 runs=3 median_avg_ms=7.01 spread_ms=0.02 median_p95_ms=9.00

Final headed, DPR 2:

  • PROFILE summary visual=all pose=gameplay03 runs=3 median_avg_ms=3.78 spread_ms=0.05 median_p95_ms=5.00
  • PROFILE summary visual=-bloom pose=gameplay03 runs=3 median_avg_ms=3.78 spread_ms=0.06 median_p95_ms=5.00
  • PROFILE summary visual=-grade pose=gameplay03 runs=3 median_avg_ms=3.75 spread_ms=0.04 median_p95_ms=5.00
  • PROFILE summary visual=-grid pose=gameplay03 runs=3 median_avg_ms=3.75 spread_ms=0.02 median_p95_ms=5.00
  • PROFILE summary visual=-haze pose=gameplay03 runs=3 median_avg_ms=3.76 spread_ms=0.05 median_p95_ms=5.00
  • PROFILE summary visual=-labels pose=gameplay03 runs=3 median_avg_ms=3.75 spread_ms=0.05 median_p95_ms=5.00
  • PROFILE summary visual=-shadows pose=gameplay03 runs=3 median_avg_ms=3.50 spread_ms=0.02 median_p95_ms=5.00
  • PROFILE summary visual=-tiltshift pose=gameplay03 runs=3 median_avg_ms=3.49 spread_ms=0.05 median_p95_ms=5.00
  • PROFILE summary visual=-water pose=gameplay03 runs=3 median_avg_ms=3.54 spread_ms=0.06 median_p95_ms=5.00
  • PROFILE summary visual=none pose=gameplay03 runs=3 median_avg_ms=3.08 spread_ms=0.01 median_p95_ms=4.00
  • PROFILE summary visual=all pose=diorama02 runs=3 median_avg_ms=18.08 spread_ms=0.06 median_p95_ms=20.00
  • PROFILE summary visual=-bloom pose=diorama02 runs=3 median_avg_ms=18.11 spread_ms=0.15 median_p95_ms=19.00
  • PROFILE summary visual=-grade pose=diorama02 runs=3 median_avg_ms=18.13 spread_ms=0.10 median_p95_ms=19.00
  • PROFILE summary visual=-grid pose=diorama02 runs=3 median_avg_ms=18.12 spread_ms=0.10 median_p95_ms=19.00
  • PROFILE summary visual=-haze pose=diorama02 runs=3 median_avg_ms=18.13 spread_ms=0.07 median_p95_ms=19.00
  • PROFILE summary visual=-labels pose=diorama02 runs=3 median_avg_ms=18.14 spread_ms=0.14 median_p95_ms=19.00
  • PROFILE summary visual=-shadows pose=diorama02 runs=3 median_avg_ms=13.28 spread_ms=0.09 median_p95_ms=14.00
  • PROFILE summary visual=-tiltshift pose=diorama02 runs=3 median_avg_ms=18.12 spread_ms=0.08 median_p95_ms=19.00
  • PROFILE summary visual=-water pose=diorama02 runs=3 median_avg_ms=18.12 spread_ms=0.19 median_p95_ms=20.00
  • PROFILE summary visual=none pose=diorama02 runs=3 median_avg_ms=13.21 spread_ms=0.08 median_p95_ms=14.00

Final macOS player (SSAO, copies and bloom off):

  • PROFILE summary visual=all pose=gameplay03 runs=3 median_avg_ms=1.02 spread_ms=0.02 median_p95_ms=1.05
  • PROFILE summary visual=-bloom pose=gameplay03 runs=3 median_avg_ms=1.03 spread_ms=0.00 median_p95_ms=1.05
  • PROFILE summary visual=-grade pose=gameplay03 runs=3 median_avg_ms=1.02 spread_ms=0.00 median_p95_ms=1.05
  • PROFILE summary visual=-grid pose=gameplay03 runs=3 median_avg_ms=1.02 spread_ms=0.00 median_p95_ms=1.05
  • PROFILE summary visual=-haze pose=gameplay03 runs=3 median_avg_ms=1.03 spread_ms=0.00 median_p95_ms=1.05
  • PROFILE summary visual=-labels pose=gameplay03 runs=3 median_avg_ms=1.02 spread_ms=0.00 median_p95_ms=1.05
  • PROFILE summary visual=-shadows pose=gameplay03 runs=3 median_avg_ms=0.73 spread_ms=0.00 median_p95_ms=0.74
  • PROFILE summary visual=-tiltshift pose=gameplay03 runs=3 median_avg_ms=0.95 spread_ms=0.00 median_p95_ms=0.98
  • PROFILE summary visual=-water pose=gameplay03 runs=3 median_avg_ms=0.91 spread_ms=0.00 median_p95_ms=0.94
  • PROFILE summary visual=none pose=gameplay03 runs=3 median_avg_ms=0.59 spread_ms=0.00 median_p95_ms=0.61
  • PROFILE summary visual=all pose=diorama02 runs=3 median_avg_ms=2.72 spread_ms=0.01 median_p95_ms=2.81
  • PROFILE summary visual=-bloom pose=diorama02 runs=3 median_avg_ms=2.73 spread_ms=0.01 median_p95_ms=2.81
  • PROFILE summary visual=-grade pose=diorama02 runs=3 median_avg_ms=2.72 spread_ms=0.05 median_p95_ms=2.81
  • PROFILE summary visual=-grid pose=diorama02 runs=3 median_avg_ms=2.72 spread_ms=0.00 median_p95_ms=2.81
  • PROFILE summary visual=-haze pose=diorama02 runs=3 median_avg_ms=2.73 spread_ms=0.03 median_p95_ms=2.81
  • PROFILE summary visual=-labels pose=diorama02 runs=3 median_avg_ms=2.72 spread_ms=0.04 median_p95_ms=2.81
  • PROFILE summary visual=-shadows pose=diorama02 runs=3 median_avg_ms=1.41 spread_ms=0.01 median_p95_ms=1.51
  • PROFILE summary visual=-tiltshift pose=diorama02 runs=3 median_avg_ms=2.70 spread_ms=0.03 median_p95_ms=2.81
  • PROFILE summary visual=-water pose=diorama02 runs=3 median_avg_ms=2.73 spread_ms=0.01 median_p95_ms=2.81
  • PROFILE summary visual=none pose=diorama02 runs=3 median_avg_ms=1.38 spread_ms=0.01 median_p95_ms=1.41

macOS -golden, both builds: GOLDEN OK 1950 71e9904d0ee48083db9bbba281f3f046136482ba83ee056e48d57b62e76a5b07, GOLDEN OK 1952 73590e324f19e9959a455a4da79e93d7089dfcf283c13aac02af6b0a624b9767.

Region borders (promo-borders, art track gap 5)

Promo 03 outlines regions with coloured bands along hex edges, over the faint hex lattice; promo 02 draws none (its era has no such layer), so diorama02 is checked only for staying thin. Measured on promo/final/03-gameplay-1936.png (2752 px wide, row and column profiles across the lines):

Promo 03 band From the region's edge inwards Widths (2752 px, mid-depth) sRGB
Blue blue stripe, white stripe, a grey-blue wash fading inwards 6 + 6 px, wash about 10 px 125,140,182 / 251,251,248
Red-orange red-orange stripe, cream stripe, a faint warm wash 4 + 3 px, wash about 8 px 176,93,61 / 246,188,124

Each band lies inside its region, as Civ draws territory, so two outlined regions meet as a double band (blue+white next to red+cream: promo 03 near Tel Aviv). The bands are world-space: they widen towards the camera.

The default look: A2 (owner ruling, 27 Oct 2026, refs #196). The owner picked "A2: stronger bands (Recommended)" from the Civ-like prototype's sheet (captures/pr-1794, civ-borders/sheet.png), so the map draws A2 unless told otherwise. Next to main's look, A2's bands keep their weight when the camera pulls out (no far fade); the colour stripe is never under 6 px, the light stripe 1 px and the wash 12 px, measured across the edge so they hold under the tilt. Both sides wash alike at 0.45, falling as 1 - x squared, so the wash holds well into the hex and reaches a quarter of its inradius (about 953 m on 4.4 km hexes); a light tint of the region's colour covers the territory, 0.12 up close and rising to 0.25 at the overview. Since #2431 the stronger tier's two sides also glow ("Issue #2431", below). Band plus wash never pass half the inradius, so a far hex is not filled by its outline. The neutral grey keeps its 0.14 wash and its dashes, never shorter than 8 px. The other variants stay selectable for comparison: ?bordervariant=legacy in the WebGL URL (or -bordervariant legacy on desktop and in the editor) draws main's earlier look exactly, and a, b, c, b2 and c2 draw the prototype's other variants; anything else draws A2 (MapBorders.VariantRequested). The paragraphs below describe the drawing every variant shares; their widths, wash strengths and the 2 px floor are Legacy's.

What is drawn. MapBorders (C#) bakes one texel per hex (RegionHexes, RGBA8 linear, point-sampled, 54 x 66 on the current hexes.json: r + 256 g the region id, b the style; since issue #195, a's low six bits say which edges border another region) and sets the globals; StatecraftBorders.hlsl's ScBorderApply, called last in the terrain function, draws the band along each edge whose neighbour is in another region. A pixel farther from its hex's edges than the widest band plus wash can reach fetches nothing (exact: a segment is never nearer than its line); one nearer fetches its own hex and only the neighbours within reach (one, or two near a corner); since issue #195 it fetches only its own hex, whose edge bits replace the neighbours' fetches. The first version fetched all seven texels on every land pixel and cost 3.3 ms at gameplay03 (visual=all 14.97 against -borders 11.64, a loaded machine): over the 2.5 ms budget. The distance is to the edge segment (corners rounded, so bands join without notches), in pixels across the edge (Task 8's anisotropic filter), box-filtered. Stripes and wash are unlit (they replace BaseColor and go out as Emission), so they read the same in sun and shadow, as the promo's glowing lines do. A hex in no region (id 0) gets no band and stops its neighbours' bands: the coast is never outlined. Bands stay on land: where a boundary between two coastal land hexes crosses the shore, the band fades with the water mask (ScWater) instead of running up to 4 km out to the hex edge (review M7; BandsStayOffTheWater). Styles (MapBorders.Primary / Secondary): band 230 m / 200 m (colour share 0.5 / 0.6), wash 0.9 / 1.0 of the band at 0.40 / 0.18. The band is at least 2 px wide unless that would make band plus wash wider than half a hex's inradius, which wins: below about 8 px of inradius on screen (the far overview) it goes under 2 px (review M2). A style with no wash is safe (the wash ramp's width is floored, review M3). The light stripes are set above white (sRGB 1.08 / 1.05 red) so the graded result reaches the promo's. Toggle borders. _ScBorders.z (test seam MapBorders.NoEarlyOut) turns the early-out off; TheEarlyOutIsExact renders both promo poses both ways and finds 0 pixels apart (review M6). Visual only: no sim state read or written, goldens unchanged.

The no-data path (a map built with no layer: MapScene.Build alone; play now draws the 1948 sub-districts, below). Owner ruling (2026-10-11): no invented political boundaries; borders come from real data only (the 1945-48 districts and sub-districts, the 1949 armistice lines when dated, fronts). The repository has none of these: hexes.json holds elevation, land and biome; settlements_1948.yaml points; the armistice events (content/events/armistice_*.yaml) dates and text but no geometry; the sim has no fronts (war mode, #21 / M4 N4). So MapBorders.Apply runs on every build with no layer. Against main (review I1, branch merged with main 632edbd, both editor renders from the same Library):

  • MapCapture.SurfFrames (gameplay03, web tier, the water's clock pinned, 60 frames): 0 of 124,416,000 pixels differ. The renders are kept in captures/visual/promo-borders-i1/ (borders-maincheck/ main, promo-borders/ the branch, main2/ main's second Promo run).
  • MapCapture.Promo (both poses, both tiers; the water's clock is not pinned there): main against the branch differs in 1,334 / 15,089 / 116,761 / 19,904 pixels (Mobile-diorama02, Mobile-gameplay03, PC-diorama02, PC-gameplay03; largest step 26/255 in PC-diorama02, 2/255 elsewhere), and main against a second main run in 513 / 1,737 / 116,542 / 10,498 pixels with the same largest steps: the water's clock, not the branch.
  • Inside the branch, WithoutRegionDataTheToggleIsANoOp only shows the toggle does nothing without data.
  • The cost of the no-data path (the hoisted ddx/ddy and one uniform branch) against main's build is not measured: the machine ran at load 22 with 13 Unity processes from other tracks, so a cross-build difference would be noise. The in-build all against -borders difference (-0.02 / +0.05 ms) is the noise floor, not that cost.

BorderProof is a synthetic pattern for tests and captures only (review I3): a checker of 4 x 4 axial hex blocks (parallelograms in a regular lattice over the whole map, unrelated to any town or line), alternately Primary and Secondary, land only. While it is drawn the camera carries the label "SYNTHETIC TEST PATTERN — not real boundaries", so every capture labels itself: a legacy TextMesh (bold, white with a dark drop copy, 3.3% of the frame high) that renders under URP, placed before each render from the camera's aspect, 35% down the frame inside the tilt-shift's sharp band (legible in captures/visual/promo-borders-r2/editor/*-gameplay03.png and *-diorama02.png). TheProofLabelRendersAtTheTopLeft renders it shown and hidden at 1920x1080 and 1440x1080: about 12,800 pixels change in the upper left, none elsewhere, and its left end is on screen; TheProofIsLabelledOnScreenWhileDrawn checks it goes with the toggle and the layer. Only the editor and development builds honour its flag (Application.isEditor || Debug.isDebugBuild; AReleaseBuildIgnoresTheProofFlag): -borderproof in the editor's promo renders and tests, tools/art_checkpoint.sh --borderproof (editor renders; the release WebGL shots ignore it, so the WebGL frame profile can no longer draw it), and ?borderproof / -borderproof in a development player.

Data: the 1948 sub-districts (promo-borderdata). The sixteen sub-districts of six districts in force on 15 May 1948, the game's start: the Administrative Divisions (Amendment) Proclamation, 1947 (Palestine Gazette No. 1552, Supplement No. 2, 30 Jan 1947, pp. 125-133, in force 1 Jan 1947) revoked and replaced the whole Schedule, so it alone defines them; the 1942 (No. 1163) and 1945 (No. 1415) changes are superseded (cited, not read). No later amendment was found (two web searches, 2026-10-11; then the Gazette's 1947 and 1948 indexes, read 2026-10-16: research_notes/Map borders/subdistrict-names-and-amendments-1947-48.md section 3); the check is closed (owner ruling, 2026-10-23, #196), and the note's section 3.3 gaps are recorded in the file.

  • True lines: content/map/subdistricts_1948.geojson (status sensitivity-reviewed, checked 2026-10-11; names as in the Gazette, Hebrew from the Gazette's 1945 Hebrew schedule (owner ruling, 2026-10-23, #196), Arabic from Wikipedia, the source noted per row), written by tools/borders_digitise.py from the Survey of Palestine map "Administrative Divisions" (1:750,000, drawn April 1946, red boundary overprint May 1947; map 6 of the "Maps of Palestine" prepared for UNSCOP, June 1947), two Commons scans of different prints (National Library of Israel), SHA-256 checked. Method: the graticule (35-36 E x 30-33 N) georeferences each scan (affine, rms 5-6 px, about 0.2 km); the red overprint plus the black dots between its dashes, with the district lettering and the name underlines erased, is closed and flood-labelled from each sub-district's chief town (a leak fails the run); a fine closing then gives the narrow salients the coarse one filled back to their own side, and Gazette-listed Metulla (Safad, p. 127), which the 1:750,000 map draws on the frontier's tip, is kept in Safad; six connectors where a line runs under a black label or stops short of the coast, and the Metulla tip, are recorded as uncertain segments; water (sea, the Dead Sea at its 1948 level) from the game's DEM, so the coast is the game's coast. Accuracy about +-1 km: areas within -4.1% to +3.7% of Village Statistics 1945 (which follows the 1945 Schedule; Beersheba's rounded figure makes its check nominal); the second print agrees to a mean 0.11-0.23 km, at most 1.23 km (per feature in the file); 41 independent Gazette-listed places (none a digitiser seed) plus Metulla (kept by its patch) fall in their sub-district (Ra'anana 0.1 km over the Jaffa/Tulkarm line, within the accuracy). The file is written one vertex per line and is byte-identical on a rerun.
  • Display: tools/borders_bake.py gives each land hex the sub-district containing its centre (a shore hex whose centre is in the water: the side of its nearest land, so the hexes off Taba, west of Rafah's frontier and by the Lisan in the Dead Sea are outside), the hex that holds a Gazette-listed place that place's sub-district, never outside (controller ruling, owner-delegated, 2026-10-11; GAZETTE_PLACES / GAZETTE_SETTLEMENTS, placed as the game places settlements: this puts Rafah town's hex (-15, 31), whose centre is in Sinai, in Gaza, Metulla's (15, 1) in Safad and Ein Gev's (12, 8) in Tiberias, which the centre rule had left outside, and moves the Kefar Sava / Ra'anana and Kafr Qasim hexes to their own sub-districts), land beyond the Mandate frontier the outside region (style none, so the bands run along the frontier and nothing is drawn beyond it), sea and lake hexes none (no coast band) -> content/map/subdistrict_hexes_1948.yaml. Every sub-district is Primary (blue). GameRoot.Begin and MapCapture.Promo draw it (Subdistricts1948.Show); the info panel names a hex's sub-district and district ("Jaffa Sub-District, Lydda District (Mandate sub-districts, 1947)"). Snapping moves a line by up to half a hex (about 4 km); the true line is the GeoJSON's. The "Why" screen lists sim causes only, so it does not name map lines (the sim must not read this file).
  • When it shows (owner ruling, 2026-10-11): from the 1948 start until the 1949 armistice lines replace it, each front switching to its armistice line on that agreement's date (February-July 1949) once that data is built; until then the layer stays, labelled on screen "Mandate sub-districts, 1947" (a UI label above the minimap while the layer is drawn; not the renderer's 3D banner, which the tilt-shift blurs; editor-only renders have no UI and so no label) (the baked file's label, shown_from: 1948-05-15, shown_until: null; Subdistricts1948.ShowsOn, applied on every turn; SubdistrictTests). Recorded in decision D-B.
  • Kept out of the sim: the sim reads files by name and its content hash covers only those, so neither file moves the hash or the goldens (SubdistrictTests.TheSimIgnoresTheFileTheContentHashDoesNotMove; test_the_sim_never_reads_the_border_files).
  • Licence: Government of Palestine works of 1945-47, public domain in Israel and the UK, with a possible URAA US copyright in the map's drawing that does not reach facts (Feist): decision D-B (docs/oss/DECISIONS.md), decided (owner-delegated, 2026-10-11): allowed; facts redrawn, no image copied; ASSETS.yaml row verified.
  • Rulings (owner-delegated to the controller, 2026-10-11): the near-frontier shore hexes (Taba, west of Rafah, the Lisan) stay as built; the Mandate frontier outline stays; Samaria District in Arabic stays "لواء السامرة" (no official Arabic Mandate text reachable online; "لواء نابلس" is the common usage; unverified, noted in the file); the Hebrew and Arabic names from Wikipedia are accepted for now; the sensitivity review is done (sensitivity-reviewed), while the 1947-48 amendment check stays open. Owner rulings, 2026-10-23 (#196): the Hebrew names follow the Gazette's 1945 Hebrew schedule headings, and the amendment check is closed (the research note above, sections 2 and 3).
  • 1949 armistice lines (armistice-lines, #196): content/map/armistice_lines_1949.geojson (true lines, draft) and armistice_hexes_1949.yaml (snapped), from the agreements and their annexed maps (research_notes/Map borders/armistice-lines-1949.md, "Built"): Syria (line and three-part DMZ), Lebanon (the frontier), Egypt (El Auja DMZ; the Gaza-Rafah line from the Wadi Hasi to Gaza; from 22 Feb 1950 the line on the signed modus vivendi map, S/1471), Jordan (low accuracy, from the low-resolution UNISPAL copy of map 1; line A by its Art. VI.3 stages; Jerusalem a gap). Each front switches on its agreement's date, which the sub-district layer records per front (Subdistricts1948.FrontReplaced, used by MapRegions): from then its zones are drawn over the sub-districts in the secondary style and its lines named in the info panel with their own agreement and date, while the sub-district bands stay as the base; the map label adds "1949 armistice lines: ", with "Jordan (low accuracy)" and "Egypt (and the 1950 modus vivendi)" (ArmisticeTests); its box takes up to the right column's 480 px (.map-layer, was 160 px), so the longest label is three lines, checked beside a real event card at seven window shapes (ArmisticeTests.TheLongestMapLabelFitsBesideTheEventCard). The sub-district layer stays as the base while the lines leave gaps. A line is visible as a band only where it bounds a zone or is the Mandate frontier; at 4.4-km hexes two of the three Syrian DMZ parts take no hex (true-polyline overlay: #237).

Shared edges and the far line (refs #196, .superpowers/borders-routes-r1/PLAN.md B1 and B3). Every sub-district is Primary, so where two meet both drew their full band: blue, white, white, blue, about 11-13 px across at gameplay03 against promo 03's one band of about 8 px, on every internal boundary. Promo 03 never shows two same-colour regions side by side. MapBorders.Use now also bakes, per hex, which border edges the neighbour draws in the same style (b's upper six bits; the style keeps the low two), and ScBorderApply draws 0.6 of the band and of the wash each side of such a shared edge (ScSharedBand), so the two make one band of the promo's width centred on the edge, with no jog where the drawing side would change. At exactly half, each half lost its partly covered outer pixel and the pair measured 6 px at gameplay03 against a full band's 8; at 0.6 the pair measures as one full band. The nearest edge is chosen in units of its own band, so where a shared edge meets the frontier at a corner each keeps its own width. Edges against a region of style none (the Mandate frontier against "outside") and primary-secondary edges (the double band promo 03 draws near Tel Aviv) keep the full band. Still one texel a pixel (TheBordersReadOneTexelAPixel). Far off, where the band is under about 4 px (diorama02), the 2 px floor band was mostly its light stripe and read as a white line; the light stripe now turns into the colour between 4.5 and 2.5 px and the wash fades with it, so a far band is a thin line of its colour. At gameplay03 (about 8 px) nothing changes. Tests: ASharedEdgeDrawsOneCentredBandAndTheFrontierKeepsTheFullBand (the proof's blocks, all primary and then primary against none), AtTheOverviewTheBandsAreThin (now also: few near-white pixels, a blueness gain), and UseBakesWhichEdgesBorderAnotherRegion (the shared bits). The proof pattern alternates styles, so the proof's measurements and ColourBlindPlayersTellTheBorderStylesApart are unchanged.

Two colours: who held which areas (owner rulings 2026-10-14, refs #196). The owner ruled that the borders use promo 03's two colours: a deeper blue for the Israeli side (about 89,119,156 on screen, against main's lighter 120,142,192) and orange-red for the other side, from sourced data only, with grey for contested ground, no-man's land and the demilitarized zones. Further rulings the same day: (a) the legend in English and Hebrew; (b) the localities stay as of 1948 ("start of game only"): the colour claims military control only; (c) the first-truce snapshot from 11 June may use the archive map's line dated 9 July (the closest sourced line, labelled as such), and grey marks areas changing hands only for the days an operation ran there; (d) no outward request to the archives: the best public copy of ISA map 6158/1, the accuracy (about +-1-2 km) stated in the data. Research note: research_notes/Map borders/control-areas-1948-49.md. Licence basis: decision D-C (owner-verify).

  • Data: content/map/control_1948_49.geojson (true polygons, status draft, owner-verify), written by tools/control_digitise.py from Israel State Archives map 6158/1 (1949; the NYU Faculty Digital Archive copy, 2368 x 4078 px: the ISA viewer refused a plain client and Commons has no copy, checked 2026-10-14). The scan is georeferenced on its own printed 10-km Palestine grid (a cubic fit, rms about 0.1 km); its red (9 July 1948), blue (18 July 1948) and green (end of October 1948) front lines are traced along their own pen colours between waypoints read off the scan; the partition line (teal) and the Mediator's demanded line (yellow) are not control lines and are not traced (where the green and the partition line run close, Resolution 181's own wording and the points the map labels in teal tell them apart). A later line is drawn only where the front moved, so each snapshot is the earlier lines less the stretches a later line replaces, plus enclaves leaving on the dates the sources give (Tira 16 July; Operation Shoter 24-26 July; Operation Hiram). The lines and the Mandate outline are polygonised and each face takes a side by parity from Tel Aviv; a face reached both ways fails the run. Closure checks: 44-50 sourced anchor places per snapshot land on their side unless within 2 km of a line (3 km once the low-accuracy Jordan line is in), and every holder place (HOLDERS) lands on the other side with no tolerance. The scan stops at the 60-km northing: south of it nothing is coloured until the Jordan agreement's line d closes the south on 3 April 1949. Each feature cites the source it comes from, with that source's accuracy: the 1948 snapshots and the operations' grey ISA 6158/1 (and the operations' pages); the 1949 sides, the zones and the Egypt closure the agreements' lines (armistice_lines_1949.geojson, D-A) and UNTS vol. 42, the Jordan-derived sides flagged low accuracy.
  • Periods (content/map/control_hexes_1948_49.yaml, tools/control_bake.py: each land hex of Mandatory Palestine takes the code at its centre, per period): 15 May-10 June 1948 the 9 July line, adjusted (owner decision of 16 October 2026, #674, option (b); research note control-15-may-1948.md section 9): the other side not named; grey where the CIA's "Military Situation - 18 May 1948" dashed (tenuously held) border is legible and closes on the 9 July line or the frontier (the Safad-Hula loop, traced on the scan georeferenced on its graticule ticks); cut by the 1947 sub-districts (no new line), the 9 July Israeli areas of the Jenin (Lajjun-Zir'in, 28 May) and Jerusalem (the CIA's corridor border, a single line along the road) sub-districts; in the Ramle and Acre sub-districts only the hex of the place that changed hands (contested_places, owner ruling 2 of 16 October 2026 on #686): Yibna's (5 June) and Acre's (17/18 May), so Rehovot, Rishon and Nahariya stay blue and the CIA's Rishon-Sarafand border is not drawn; the Gaza and Beersheba sub-districts without colour, the hexes of their Jewish settlements whose 1948 site is sourced pinned blue (Q3); every hex within 6 km of the Old City no-man's land; 11 June the 9 July line; 8-18 July (the Ten Days) grey where the 9 and 18 July lines differ; 18 July the blue line; 24-26 July grey at Ijzim, Jaba and Ein Ghazal; 15-31 October grey where the July and October lines differ in the south (Yoav, Ha-Har, Beit Jibrin, al-Dawayima), 29-31 October in the Galilee (Hiram); from 31 October the green line, until the Egypt agreement (Operation Horev is not drawn: its fighting lay south of the scan or beyond the frontier, and no source gives a line for January 1949); from each agreement's date (24 February, 23 March, 3 April with Art. VI.3's stages on 8 May, 22 May and 17 July, 20 July 1949) that front takes its armistice line, each 1948 piece of that front leaving on the date of the nearest line of its front, and the demilitarized zones are grey. Gaps closed: Jerusalem by the 1948 lines within 4.5 km of the city (map 2 is not available); the Egypt line south of Gaza straight to the frontier within Art. VI.3's bounds, the land within 3 km of that stretch on its Israeli side left without colour (the Strip itself stays Egypt's: Gaza, Deir al-Balah, Khan Yunis and Rafah read Egypt in every period, pinned by test_control.py).
  • Pinned towns (pinned_towns; the bake gives the town's hex the sourced side, as D-B's rule 5 does for Gazette places): Mishmar HaYarden, Syria's from June 1948 to 20 July 1949 (en.wiki), whose bridgehead is narrower than the map's accuracy (at 4.4 km its point falls in Rosh Pinna's hex, so the pin goes on the free neighbour nearest it, (15, 5): owner on #1796, 27 October 2026, "Move the pin"); Qalqilya, on the Jordanian side from 3 April 1949 (en.wiki quotes Hadawi, who argues that the 1949 lines severed its groves in favour of Israel, which implies the town stayed on the Jordanian side: an inference; on the UNISPAL copy of map 1 the pen line runs along the town's western edge, which 51 m a pixel cannot settle), which the low-accuracy Jordan trace puts 2 km to the Israeli side; Huleiqat, Egypt's from 8 July to 19 October 1948 (en.wiki: retaken by the Egyptian army on 8 July, captured by Giv'ati in Operation Yoav on 19 October), where ISA 6158/1's 9 July line, about 1.3 km away, puts it on the Israeli side (the source and the archive map disagree; recorded in the data); a pin covers whole periods only, so the Yoav period shows Huleiqat's hex Israeli from 15 October, four days before its capture.
  • Drawing: MapRegions draws Control1948's layer from 15 May 1948 (#674): Israel Primary, the other side Secondary, contested, no-man's land and zones the new Neutral style (BorderStyle.Neutral, the shader's third style, dashed along its edge every 2 km, 65% drawn: a cue that is not colour), the land beyond the frontier style none, no colour also style none (owner ruling of 27 October 2026, "Close the borders"; was left out of the layer, which left one-edge bands such as Shoval's in mid-air): it draws no band of its own and each side's band runs one-sided along it, as along the frontier, so every pinned settlement hex is ringed and no front ends in mid-air (ControlTests.NoBorderEndsInMidAirInAnyPeriod, every period). Before a first period (none with today's data) the sub-districts would be drawn in Neutral, since blue means the Israeli side. Colours: Primary pre-grade sRGB 0.29, 0.40, 0.56 (was 0.40, 0.50, 0.80), measured on screen at deltaE76 0.6 from 89,119,156; Neutral 0.46, 0.46, 0.44 with a pale stripe 0.98, 0.97, 0.93.
  • Label (map-layer, MapRegions.ControlLabel): the legend names each side before its colour, and grey as dashed: 1948 "Held by Israel: blue · Held by Arab forces: orange-red · Contested or no-man's land: grey, dashed", with the caption "Front lines, {date}, approximate (Israel State Archives)", from 15 May to 10 June 1948 "Front lines of 9 July 1948, adjusted (Israel State Archives; CIA map of 18 May 1948)"; once a front has its agreement "Israeli side of the armistice line: blue · Arab side: orange-red · No-man's land or demilitarized zone: grey, dashed", with "Armistice lines: {fronts}" (all switched) or "Armistice lines: {fronts}; elsewhere front lines of {date}" (some), Jordan named "Jordan (low accuracy)" (D-A ruling 2). Dates in the UI's language (Hebrew month names). The label may take up to six lines in its narrow box; ControlTests.TheLongestLegendLabelFitsBesideTheEventCard checks EN and HE at eleven window shapes. The info panel says who held the hex ("Held by Arab forces (Egypt)"; Jordan is named as the state from its agreement; in 1948 the West Bank's holders share one connected area no sourced line divides, so it is named only as "Arab forces"), nothing for a hex beyond the frontier, and no sub-district once its bands are gone.
  • Known staleness and limits: the October 1948 line stands until the Egypt agreement (24 February 1949), so Isdud (taken 28 October), Majdal (4-5 November) and Yad Mordechai stay on the other side until then (the caption names the source line's date). Huleiqat is pinned (above). The ground east of the Beersheba-Hebron front reads as held by Arab forces in 1948, since the source draws only front lines; no "not held" area is invented. Jerusalem's line falls between two hexes (the Old City's hex is the other side's, the one west of it Israeli; the grey rule for a hex holding both sides does not fire).
  • Not done, recorded in the data (coordinator rulings, 2026-10-14): the Syrian and Israeli truce lines of S/1353/Rev.1 (on its colour scan the Israeli line is legible only in parts and the Syrian line cannot be told from the zone's outline and the river; ISA 6158/1's red and blue cover that front, and Mishmar HaYarden is pinned); the CIA ORE 38-48 map as a cross-check (its reading-room copy was unreachable and the map exists online only as a JVL copy; the anchors come from Wikipedia); January 1949; the Lebanese villages held from Hiram to the armistice (outside the frontier: not coloured); sub-hex features (Kfar Darom, Mount Scopus, the Faluja pocket's exact edge) kept in the true polygons only.

Research item: region data (to source, then the owner's sensitivity review).

  • What: per-hex region labels for (a) the sub-districts of Mandatory Palestine at the end of the Mandate (six districts: Galilee, Haifa, Samaria, Jerusalem, Lydda, Gaza; sixteen sub-districts: Acre, Beisan, Nazareth, Safad, Tiberias, Haifa, Jenin, Nablus, Tulkarm, Hebron, Jerusalem, Ramallah, Jaffa, Ramle, Beersheba, Gaza), dated by the Gazette changes of 1942, 1945 and 1947; (b) the 1949 armistice lines, each dated by its agreement (Egypt 24 Feb, Lebanon 23 Mar, Jordan 3 Apr, Syria 20 Jul 1949; the events already in content/); (c) fronts, from war mode's data once #21 defines it.
  • Sources to check: Palestine Gazette, Supplement No. 2, No. 1163 (15 Jan 1942) pp. 119-132, No. 1415 (7 Jun 1945) pp. 621-629, No. 1552 (30 Jan 1947) pp. 125-133 (cited by https://en.wikipedia.org/wiki/Districts_of_Mandatory_Palestine, checked 2026-10-11); the Survey of Palestine 1:100,000 sheets (1944-48), which draw district and sub-district boundaries (Princeton Digital Maps, https://maps.princeton.edu/catalog/princeton-dc5999ng80p); UN map "Palestine: Land ownership by sub-districts" (1945 data), https://digitallibrary.un.org/record/498482; the armistice agreements themselves (UN S/1264/Corr.1, S/1296, S/1302/Rev.1, S/1353; per https://en.wikipedia.org/wiki/1949_Armistice_Agreements, checked 2026-10-11), whose lines are given as Palestine-grid map references; whether maps were annexed is to verify. No open, licensed vector file of the sub-districts was found (search 2026-10-11): AUB's "Palestine Population Sources" (2018, https://scholarworks.aub.edu.lb/handle/10938/22154) covers four sub-districts and states no licence; Palestine Open Maps states none either. So the boundaries must be digitised from a primary source (the Gazette schedules or the Survey sheets), with the method and its error recorded; their copyright status (Government of Palestine publications of the 1940s) is a licence question for the owner (docs/oss/DECISIONS.md) before any import.
  • Shape: sourced, dated geometry in content (the chapters design's lines.yaml, §448, or a polygon file beside it), and a tool that labels each hex by its centre (as tools/settlement_hexes.py does for settlements) into the per-hex layer the renderer takes (RegionLayer). Every entry is submitted fact-checked; the owner sets sensitivity-reviewed before a shareable build (chapters design §13).
  • Limits of the renderer's data seam (review M8), for the data work to plan around:
    1. No longer true as written: there are now three drawn styles plus None, Neutral = 3 (a grey band: contested, no-man's land, a demilitarized zone, or the 1947 sub-districts before any side is sourced; owner ruling 2026-10-14; MapBorders.BorderStyle). Originally two (Primary, Secondary): the 1949 lines set Israel against four neighbours, and the four still share the red-orange band. The texel's style takes b's low two bits, which the four values now fill, so a further style needs more bits as well as colours that stay apart for colour-blind players (a design task, not a constant).
    2. No longer true: layers do show together. MapRegions keeps the Mandate sub-districts as the base and draws each front's 1949 armistice zones over them from the date that front switches (unity/Assets/Scripts/Map/MapRegions.cs).
    3. Styles are per region, not per edge: an edge's look comes from the regions on its two sides.
    4. Hex snapping moves a line by up to half a hex (see the owner decision below). Follow-up: more styles or per-edge styles when the data and the owner's ruling on snapping need them.
  • Owner decision (ruled 2026-10-11: store the true lines, snap to hex edges only for display, as the promo does): the promo snaps borders to hex edges, but spec §3 (map state) says hex boundaries carry no political meaning and the chapters design draws lines as true polylines. Snapping a line to hex edges moves it by up to half a hex (about 4 km). Options: (1) hex-snapped regions (this renderer, the promo look); (2) true polylines (a separate overlay, not the promo look); (3) both, hex-snapped at the gameplay zoom and true lines in a map mode.

Measured.

  • Rendered tests (EditMode, BorderTests, 1920x1080 at gameplay03's distance and yaw, tilt-shift off, on the proof's primary-secondary edge (3, 11)-(4, 11) on the flat coastal plain, where the promo's lines are measured): along 9 points of the edge a blue stripe (graded sRGB 0.48, 0.58, 0.79; promo 0.49, 0.55, 0.71), a white one (0.93, 0.91, 0.89; promo 0.98), a red-orange one (0.74, 0.37, 0.18; promo 0.69, 0.36, 0.24), a cream one (0.92, 0.79, 0.53; promo 0.96, 0.74, 0.49). Band widths (median, stripes classified by colour so the wash is not counted): primary 8 px, secondary 6 px, against the promo's about 8 and 5 (bounds 6-11 and 3.5-8). Nothing changes 1.2 km or more from the edge; the middle of a coast edge is unchanged (0/134 pixels); no clearly-water pixel moves at gameplay03 (0 of 631,101); diorama02 3.2% of pixels changed (thin, label included); renders repeat exactly.
  • Colour-blind (GradeTests.ColourBlindPlayersTellTheBorderStylesApart, Machado 2009 severity 1, deuteranopia and protanopia, five points on a coastal-plain edge and five on a Negev desert edge): blue against red-orange 69-80; primary colour-land 54-77, light-land 38-47, colour-light 44-50; secondary colour-land 9-31, light-land 3-49, colour-light 30-35. Asserted per point: styles >= 15; each band's stronger stripe >= 15 from the land; its two stripes

    = 20 apart; its weaker stripe >= 5 unless the stronger one is >= 25. Not both stripes >= 15 (review M5), because a light stripe is meant to be light: on pale Negev sand the promo's cream stripe melts into the ground for normal vision too (3.2-6.1, logged), and the band reads by its red-orange stripe there (19-31). The red-orange is a little more saturated than the promo's (set 0.75, 0.31, 0.13) so it keeps that contrast on olive and on dry grass.

  • Frame cost (WebGL, 1920x1080, headless). Measured on the first, wider version (300 m / 230 m, before the water fade), when the release WebGL player still honoured ?borderproof; the machine was loaded (visual=none 10 ms against Task 8's 2.4 ms all), so read the differences:
    • Seven fetches on every land pixel: gameplay03 all 14.97 - -borders 11.64 = 3.3 ms, over the 2.5 ms budget; hence the early-out.
    • With the early-out, the proof drawn: gameplay03 13.33 (spread 0.31) - 12.42 (0.22) = 0.91 ms, within budget. diorama02 16.93 (1.06) - 15.82 (0.73) = 1.11 ms is within one spread: not resolved; re-measure on an idle machine with real data (promo-borderdata), in a development build or a release build with a data layer.
    • Without data: see I1 above (not measurable against main on this machine).
  • WebGL build: no Shader error, no "not supported" line (GLES3: LOAD_TEXTURE2D is texelFetch, static const arrays, no new derivative use beyond Task 8's ddx/ddy in uniform control flow); both goldens OK (Hash1952 62f8ec61 since #137), and golden&smoke=5 passes on the build of d1973db.
  • Texture units with promo-land merged (review M10): the terrain fragment's own textures are 6 (Bake, Coast, Ao and Land through the graph, _ScBiomeTex, _ScRegionTex; promo-land dropped Tracks), plus URP's (main-light shadow map, reflection probe and the like): about 10-11 by the source, under WebGL2's guaranteed 16. Built, not only counted: a local test merge of d1973db with origin/promo-land 7db1a3a (not pushed) built for WebGL with 0 errors and no Shader error (76.0 MB), and booted in Chrome for Testing with GOLDEN OK 1950 and GOLDEN OK 1952 and no "not supported" or Shader line (web_check's fail pattern).
  • Checkpoints (art-log rows): promo-borders-before (main 807e300); promo-borders-proof and promo-borders-after were run on 807e300's tree plus the then-uncommitted branch, so the art log records 807e300 for them (now marked 807e300+wip there); the rings in them are the old pattern. The review rounds' checkpoints are on d1973db (art-log rows promo-borders-noproof and promo-borders-r2-borderproof): promo-borders-noproof (no proof: the branch as played, goldens OK) and promo-borders-r2 (--borderproof: the editor renders show the labelled checker; its WebGL shots, a release build, ignore the flag and match -noproof's numbers exactly).

promo-sea: water, surf and beach (art track, open item 3, issue #38)

ScWaterShade (StatecraftMap.hlsl) is reworked on both sides of the edge, and a new ScShore paints the land's shore. Review round 1 (staff review I1-I2, M1-M8) is folded in; the values below are the shipped ones.

  • Depth. The map has no bathymetry (GLO-30 is a land model; its sea texels are 0 m), so depth is still the distance to the shore from coast.png. The ramp:

    • turquoise shallows ScShallowM 2.6 km wide, on a linear weight so the band reads at its width; the edge wanders +-35% on a 3 km noise;
    • then teal (mid), then a steel blue (deep) by 12 km (ScCoastRangeM), on pow(smoothstep(0, 12 km), 0.6).
  • Colours (sRGB, before the grade; Task 7's in brackets):

    Constant Value Task 7
    ScShallowSrgb (0.40, 0.72, 0.70) (0.38, 0.74, 0.72)
    ScMidSrgb (0.29, 0.50, 0.59) (0.13, 0.47, 0.58)
    ScDeepSrgb (0.27, 0.40, 0.54) (0.06, 0.24, 0.40)
    ScFlatSeaSrgb (the water toggle's flat sea, x 0.8) (0.20, 0.36, 0.47) (0.16, 0.42, 0.55)
    ScGalileeSrgb (fresh water, -300 to -100 m) (0.20, 0.45, 0.55) (0.15, 0.48, 0.60)
    ScSandSrgb / ScSaltSrgb (0.88, 0.76, 0.56) / (0.92, 0.90, 0.85) none
  • Waves. Two layers drive the wave normal:

    • the 700 m swell (unchanged);
    • a 170 m chop moving the other way, ScChopGain 1.4 times the swell, full under a 40 m pixel and gone by 170 m.

    The tilt is ScWaveTilt 0.6 under a 50 m pixel (gameplay03 is about 45 m: chop and sparkle show). It eases to ScWaveTiltFar 0.10 by 120 m (Task 7's 0.15 until sea-fix): the far sea reads calm, and the seam test's patch means stay within bound (at a flat 0.6, one 7x7 patch at 120 km carried 0.032 against 0.012). The glint (0.06) fades out between 60 and 120 m pixels (sea-fix, below).

    Bug B10 (sea-fix): both layers were value noise differenced along the map's axes, on axis-aligned lattices. Their slope jumped at every cell edge, so the open sea showed a blocky checker at 6-20 km. Both layers are now gradient noise with its analytic slope (ScGradientNoiseD, quintic fade: smooth across cell edges), the swell's lattice turned 0.6 rad and the chop's -0.9 rad (ScWaveSlope), scaled by ScWaveSlopeGain 0.22 to the old layers' strength (open sea luma std at 6 / 12 / 18 km: 5.5 / 6.6 / 6.8 before, 6.0 / 6.8 / 7.1 after, in 8-bit levels). Fewer hashes than before: 4 gradients per layer against 16 value-noise lookups. TheOpenSeaHasNoBlockGrid renders the open sea straight down at 4 m a pixel and folds the luma gradient by its position in each old lattice's cell; the block-edge energy ((largest bin - smallest) / mean, bound 0.25) was 0.660 before (170 m along z) and 0.122 after. MapCapture.SeaClose renders the sea at 6, 12 and 18 km. With the new swell the seam test's single +-2.5 km patches (one 7x7 patch is about one swell cell at 120 km) read up to 0.016 apart (bound 0.012) at the far tilt of 0.15, so the far tilt is now 0.10, and the glint, the near zoom's sparkle, fades out between 60 and 120 m pixels. Frame time (headless; another session loaded the machine, load average 4-18, so the spreads are wide):

    • PROFILE summary visual=all pose=gameplay03 runs=3 median_avg_ms=8.35 spread_ms=0.11 median_p95_ms=10.00
    • PROFILE summary visual=-water pose=gameplay03 runs=3 median_avg_ms=8.03 spread_ms=2.69 median_p95_ms=9.00
    • PROFILE summary visual=all pose=diorama02 runs=3 median_avg_ms=13.54 spread_ms=0.22 median_p95_ms=14.00
    • PROFILE summary visual=-water pose=diorama02 runs=3 median_avg_ms=13.24 spread_ms=0.21 median_p95_ms=14.00

    Water and shore cost 0.32 ms at gameplay03 (inside a 2.69 ms spread) and 0.30 ms at diorama02, against the 2.5 ms budget (0.89 / 0.13 ms in promo-sea round 1); p95 with everything on is 10 / 14 ms, under 16.7 ms.

  • Surf (review I2).

    • A swash line, full to ScSwashM 180 m and gone by 324 m.
    • Breakers through a ScSurfM zone (900 m here; 1.8 km since sea round 3; 2.4 km in #693; since #690's option C, 3.0 km on the dune coasts of DuneCoasts.json and 1.5 km elsewhere). They are a squared triangle wave in the distance from the shore, spacing ScBreakerSpacingM (300 m here; 450 m since sea round 3), rolling shoreward at ScBreakerRate 0.05 cycles/s (15 m/s here, a quarter pixel a frame at gameplay03 and 60 fps; 22.5 m/s, about half a pixel, since sea round 3).
    • The breakers are box-filtered in closed form (ScTriSqBox), so they never alias and tend to their mean (1/3) as the pixel grows. The previous pow(tri, 5) crest was about 21 m wide on 45 m pixels: hairline, dashed crests.
    • The box is fwidth(phase) x ScBreakerFilterPx 1.25, with the phase taken by ScBreakers outside the water branch in both callers (re-review N2). The phase's true gradient is not 1 / 300 m: its noise term adds up to 0.0025 cycles/m and the 8-bit coast's slope varies up to 2x, so a width from the nominal spacing was up to 2x too narrow in places.
    • The 300 m break-up noise fades to its mean between 75 and 150 m pixels. Since sea round 3, the foam toggle swaps it for a 200 m lace net at the gameplay zoom (below).
  • Beach (ScShore).

    • A sand strip along every sea-level coast (above -100 m): the Mediterranean and the Gulf of Eilat, both sand beaches, but not the rift lakes. ScBeachM is 650 m on average, +-40% on a 4 km noise. The width is symbolic, after the promo's beach: wider than the real 30-100 m strip, inside the coastal dune belt.
    • Wet sand (x 0.78) to 60-160 m; steep faces keep their rock.
    • White salt flats along the Dead Sea (below -300 m).
    • Both elevation thresholds are 20 m ramps (review M7: no hard contour where a shore slope crosses them).
    • The shore is off with the water toggle (review M1), so all - (-water) measures water and shore together.
  • Coast precision (review M6).

    • coast.png is L8 over +-12 km: 94 m a level (texels 79 m E-W by about 211 m N-S), linearly interpolated between texels. A crest or band edge can sit up to 47 m off its true place, about one pixel at gameplay03 (about 45 m), and under the breakers' 2-cycle phase noise.
    • The wet-sand band (60-160 m) is about one level wide: a soft darkening, not a measured feature.
    • A 16-bit coast, or a fine near-shore channel, is the fix if surf detail is pushed below about 100 m.
  • Test and capture hooks.

    • MapShaderGlobals.PinWaterTime(seconds) pins the water's clock (_ScWaterTime), so tests and captures can render chosen moments.
    • MapCapture.SurfFrames renders 60 gameplay03 frames on the web tier, consecutive at 60 fps or -surfDt seconds apart.
  • Not done here.

    • Boats: the owner wants no civilian boats; navy ships are units, on another track. (Superseded by #536: see "Sea and boats".)
    • Harbour water and piers at Jaffa need geometry, which belongs to the settlement track.
    • Issue #16 (the 1948 Dead Sea level, Lake Hula) is untouched.

Measured

WebGL ?shot, from tools/art_checkpoint.sh. The checkpoints:

  • before: promo-sea-before, at 0f13336;
  • round 0: promo-sea-after, at 45a357e;
  • round 1: promo-sea-r1, at af6e3cc.

Sea pixels are those with blue above red by 0.06, outside the top bar, the bottom-left panel, the minimap and promo 03's icon column. Spread is the p90 minus p10 of sea-pixel luma; "sea sat p10/p90" is the same for HSV saturation.

before round 0 round 1 promo
gameplay03 frame saturation 0.57 0.49 0.48 0.38
gameplay03 contrast 0.196 0.182 0.174 0.203
gameplay03 warmth 0.077 0.103 0.102 0.064
gameplay03 palette distance 0.69 0.61 0.60 -
gameplay03 sea mean sRGB (21, 77, 105) (49, 85, 107) (62, 101, 121) (57, 81, 103)
gameplay03 sea saturation 0.82 0.55 0.49 0.47
gameplay03 sea sat p10/p90 0.57/0.93 0.44/0.63 0.43/0.55 0.29/0.58
gameplay03 sea luma p10/p90 (spread) 0.16/0.45 (0.29) 0.26/0.39 (0.13) 0.31/0.44 (0.12) 0.09/0.48 (0.39)
diorama02 frame saturation 0.47 0.43 0.42 0.43
diorama02 contrast 0.203 0.192 0.187 0.200
diorama02 warmth 0.104 0.117 0.117 0.145
diorama02 palette distance 0.67 0.67 0.59 -
diorama02 sea mean sRGB (40, 82, 111) (63, 97, 118) (74, 109, 129) (80, 120, 151)
diorama02 sea saturation 0.66 0.48 0.43 0.47
diorama02 sea sat p10/p90 0.59/0.73 0.42/0.52 0.41/0.46 0.30/0.56
diorama02 sea luma p10/p90 (spread) 0.24/0.42 (0.19) 0.32/0.40 (0.08) 0.36/0.44 (0.08) 0.38/0.53 (0.15)

Against the bars:

  • diorama02 warmth stays in [0.087, 0.145], and saturation stays at or under 0.48.
  • The sea means now sit between the two promos.

What still differs:

  • The sea's spread is a third of promo 03's and half of promo 02's. Promo 03's dark tail is boat hulls and building shadows on the water, and its bright tail is glints and foam lace. (The boats are superseded by #536: see "Sea and boats".)
  • gameplay03 contrast falls as the sea brightens (0.196 -> 0.174).
  • gameplay03 warmth rises as the sea loses blue and gains a beach (0.077 -> 0.102 against the promo's 0.064).

SeaTests (round 1, against Task 7, each with its bound):

Check Task 7 Round 1 Bound
Foam over shallows (luma) 0.154 0.180 0.02
Shallows over the open sea 0.344 0.211 0.05
Dead Sea over the Mediterranean 0.46 0.27 0.1
Diorama open sea b - r 0.299 0.230 0.15
Haze luma shift 0.069 0.034 0.01
Open sea pre-grade HSV saturation 0.91 0.49 < 0.7

TheSeaBeyondTheEdgeMatchesTheTerrainSea: 32 lines pre-grade max 0.007 (bound 0.012; Task 7 0.004), graded max 0.014 (0.024).

New tests:

  • ASandStripLinesTheMediterraneanShore.
    • Strip off: RGBA(0.716, 0.656, 0.378), luma 0.65, b/g 0.58, 0.25 over the plain.
    • Strip on: RGBA(0.906, 0.771, 0.516), luma 0.78, b/g 0.67, 0.39 over the plain.
    • Bounds 0.72, 0.62 and +0.32: all three fail with the strip off.
  • TheBreakersAreFilteredToThePixel (re-review N1).
    • It renders at 30 km and 960x540, about 45 m a pixel: the same footprint as gameplay03's 60 km at 1920x1080. There an unfiltered crest is sub-pixel but not yet faded.
    • Each water pixel of a 49x49 patch around the surf zone (6067 pixels over four moments of a cycle) must match its 4x4 block in the same view rendered at 4x. Pixels touching land or markers are left out: the shoreline's step and the markers' edges are not filtered by design.
    • With the filter: p99 0.0082, worst 0.0176 (bounds 0.014 and 0.035).
    • Red with Round 0's crest swapped back in temporarily: p99 0.0200, worst 0.0620, and the test fails.
    • The same squared crest without its box measured p99 0.0087, worst 0.0247. At 300 m spacing it is about two pixels wide by itself, so the box is what keeps it exact as the zoom changes.
    • The surf pixel spans luma 0.642-0.788 over one breaker cycle (bound 0.03).
    • The overview is still: 0.0036 over 0.5 s (bound 0.01).

In motion (MapCapture.SurfFrames, a 200x400 px coast crop at gameplay03):

  • Over one second of consecutive 60 fps frames, no pixel changes more than 2/255 from one frame to the next, either before or after the change.
  • Over the whole second, the old crests moved single pixels by up to 46/255 (2.8% of the crop over 2/255). The filtered breakers move them by at most 9/255 (2.5%), the same with the fwidth(phase) box. The open sea's waves move them by at most 14/255 (1.7%).
  • A 15 s clip at -surfDt 0.25 shows soft crests rolling in, broken along the shore.

Frame time: headless, the machine loaded by other sessions (load average 4-12), so the absolutes are inflated; read the differences.

  • PROFILE summary visual=all pose=gameplay03 runs=3 median_avg_ms=8.19 spread_ms=0.05 median_p95_ms=9.00
  • PROFILE summary visual=-water pose=gameplay03 runs=3 median_avg_ms=7.30 spread_ms=0.05 median_p95_ms=9.00
  • PROFILE summary visual=none pose=gameplay03 runs=3 median_avg_ms=7.01 spread_ms=0.05 median_p95_ms=8.00
  • PROFILE summary visual=all pose=diorama02 runs=3 median_avg_ms=13.63 spread_ms=0.18 median_p95_ms=15.00
  • PROFILE summary visual=-water pose=diorama02 runs=3 median_avg_ms=13.50 spread_ms=0.14 median_p95_ms=15.00
  • PROFILE summary visual=none pose=diorama02 runs=3 median_avg_ms=10.31 spread_ms=0.03 median_p95_ms=13.00

Water and shore together (the toggle now covers both):

  • gameplay03: 0.89 ms, against a 2.5 ms budget. Round 0 measured 0.50 ms for the water alone; the shore was then unmeasured.
  • diorama02: 0.13 ms, below the noise (spread 0.18 ms).
  • p95 is at or under 16.7 ms at both poses.

Sea round 2 (branch sea-round2, refs #388)

At the gameplay and overview zooms the sea had faded to a smooth surface. The chop is gone by a 170 m pixel, and the wave tilt and glint fade out by 120 m (Task 7, sea-fix), so gameplay03's 46 m pixels and diorama02's 190-300 m pixels saw a flat gradient. Sea detail on main measured 0.013 against promo 03's 0.060, and 0.019 against promo 02's 0.063. Every change is in ScWaterShade and ScHazeWater (StatecraftMap.hlsl), which only the water's own pass runs since #195, plus ScShore's two constants. Sea.shader is unchanged.

  • The far sea's texture, all behind a new seatexture visual toggle (_ScSeaTexture), so the profile prices it as one feature:

    • wind streaks: a gradient-noise brightness variation (ScMottleCellM 2.5 km, 3x narrower across the wind, so 833 m there, with a finer octave of about 309 m, +-ScMottleGain 0.32) on the open Mediterranean, past the shallows, not on the rift lakes. Neither octave is faded by the footprint;
    • a ripple: the waves' slope towards or away from the camera, read as their sky reflection. Under a 50 m pixel it is +-0.54 (the tilted normal: swell and chop), faded out over 50-90 m; the 170 m chop's limit is 85 m. From 90 m it is +-0.28 (the 700 m swell's own slope), faded over 250-450 m. That leaves it at half strength at the swell's 350 m limit, and at 84-100% at diorama02's 190-300 m pixels (2.3-3.7 px a swell cell). It runs at 40% in the turquoise shallows, so they keep their pale band;
    • whitecaps: a fleck in some 120 m cells on the streaks' bright side, its brightness the share of the pixel it covers, gone by an 80 m pixel.
  • Glint: 0.12 under a 40 m pixel, easing to Task 7's 0.06 by 60 m, so nothing changes from 60 m out.

  • Surf line: the swash is never under 1.5 pixels wide (max(ScSwashM, 1.5 x footprint)). The breakers and their box filter are unchanged; at the breaker test's zoom the swash is still its own 180 m.

  • Beach: ScBeachM from 650 to 900 m; ScSandSrgb (0.88, 0.76, 0.56) to (0.90, 0.80, 0.62), drier and paler.

  • Water haze: the water's haze starts ScWaterHazeLeadM (80 km) nearer than the land's, at 120 km, so the overview's sea pales towards the horizon as promo 02's does. The gameplay zoom, under 100 km, has none.

  • The seam test (reviewer-agreed option A). A 2.5 km streak pattern varies by itself between the seam test's points 5 km apart, so TheSeaBeyondTheEdgeMatchesTheTerrainSea pins seatexture off and keeps its bounds and cases. The new SeaTests.TheFarSeaTextureHasNoStepAtTheMapEdge checks the edge with the texture on. It renders the west edge straight down at the seam test's 110 m a pixel, over a 40 km strip, with the water's clock pinned, and runs two checks:

    • per row, the change between the two pixel columns either side of x = 0 against neighbouring columns elsewhere. The RMS excess at the edge must stay under 0.012;
    • the mean signed change across the edge, summed over r, g and b, must stay under 0.012.
    run RMS excess mean change
    green (final values) -0.0032 0.0008
    red: the skirt's map coordinates offset by one streak cell (a texture that does not continue; streaks at 0.25) 0.101
    red: the skirt 4% brighter (an input step; streaks at 0.25) 0.0195

    Both red runs fail. Fitting a line to each side and extrapolating was tried first. Over the whole strip, a mismatched texture's steps cancel; per 880 m segment, the texture's own variation swamped the fit (0.049 with no seam, 0.059 with one).

  • MapEdgeTests.PanningIntoACornerShowsNoSkirtSeamOrFlatFill pins the water's clock. It compares a render with the skirt against one without, and the water moved between the two. A few pixels of the north-west corner's coast changed at random: 0-14 a run, with the texture on or off, and main has the same exposure. The far sea's ripple made a failure likelier (5 and 1 pixels in two branch runs). Pinned, it passed 3 of 3 runs with 0 pixels at every corner.

Measured: WebGL ?shot at main db99f3a (with towns round 2, land round 2 and #399's lighter far tilt-shift), then the branch on it (art-log rows main-db99f3a and sea-round2-r). The sea crops are those of #388, plus gameplay03's focus band (outside the tilt-shift's bands). Detail is the std of luma minus a 4 px blur, with the crop resized to 480 wide. Spread is luma p90 - p10.

main round 2 promo target
gameplay03 sea, #388 crop: detail / spread 0.014 / 0.089 0.045 / 0.160 0.060 / 0.20 >= 0.035 / >= 0.15
gameplay03 sea, focus band: detail / spread 0.017 / 0.088 0.055 / 0.185 0.075 / 0.32 >= 0.035 / >= 0.15
diorama02 sea: detail / spread 0.019 / 0.038 0.026 / 0.069 0.063 / 0.11 - / >= 0.07
gameplay03 palette distance 0.361 0.347 - no worse
diorama02 palette distance 0.563 0.534 - no worse
diorama02 warmth / saturation 0.109 / 0.406 0.111 / 0.400 0.145 / 0.43 [0.087, 0.145] / <= 0.48

The gameplay03 floors above (detail >= 0.035, spread >= 0.15) are superseded by "Sea and boats" below (#388 and #536; the detail floor was the maintainer's call, the spread floor the owner's). Promo 03's own open sea measures 0.012 in edge density, so the far sea is now calmed on purpose. The floors are now:

  • detail >= 0.025 in #388's crop and >= 0.033 in the focus band, just under the branch's 0.029 and 0.037;
  • spread >= 0.09, under the branch's 0.096 and 0.104. The author proposed it and the reviewer recommended it; the owner decided "Lower to 0.09" on 2026-10-14 (#536, "Changed acceptance criteria").

They are kept so the sea cannot go flat. #388's "Done when" was updated to match.

diorama02's spread misses its target by 0.001. Strengthening the streaks barely moves it (0.25 to 0.32 added 0.003). A stronger far ripple could close it, but the far ripple already runs at 84-100% at diorama02, at 2.3-3.7 px a swell cell. More contrast at that sampling would push it towards aliasing, so it was not raised. In the editor (web tier) the same spread is 0.079. The sea's faint hex lines (GridWaterOpacity 0.10, a measured grid ruling) are harder to see against the textured sea; that belongs to the grid, not here.

In motion (review of #402: the far ripple and the streaks' fine octave sit near or under their sampling limit at diorama02). A temporary editor probe rendered diorama02 (web tier, tilt-shift on) during a slow pan (60 m east and 30 m north a frame, about 0.25 px) and a slow zoom (0.25% a frame), 24 frames each, with the water's clock at 1/60 s a frame. It was run for main b10e95e, the branch, and the branch with the far ripple faded over 175-350 m. For each 64 px open-sea patch and each pair of consecutive frames, the patch's sub-pixel shift was found by phase correlation and taken out. What remained was compared with the patch's fine texture (luma minus a 3 px blur):

sea band main: fine texture / residual branch branch, far ripple at 175-350 m
in focus (y 30-60%), pan / zoom residual 38% / 38% of 1.0 level 19% / 19% of 4.7 levels 20% / 18%
far (y 13-30%, in the tilt-shift's band), pan / zoom 0.35 level, residual 0.24 / 0.35 level 0.60 level, residual 0.44 / 0.57 level -

In focus the texture moves coherently: its residual is half main's share, at far more texture. In the far band, which is already blurred by the tilt-shift, what is left is under one 8-bit level, and so is its frame-to-frame change. Moving the far ripple's fade to 175-350 m changes neither. No shimmer or crawl was found, so the tuning stays.

Frame time on the rebased head, against main db99f3a: one interleaved round, headless, under the exclusive lock (load about 5); three earlier rounds, before the rebase, agree:

pose all main / branch -water main / branch -seatexture (branch) none main / branch p95 all main / branch
gameplay03 8.04 / 8.23 7.86 / 7.99 8.17 5.27 / 5.48 13 / 13
diorama02 19.94 / 18.72 19.41 / 18.94 18.73 12.66 / 12.35 22 / 21

The texture costs 0.06 ms at gameplay03 and nothing measurable at diorama02. gameplay03's all rises 0.19 ms, as does its none (0.21 ms) with every feature off: build-to-build noise, not the water.

Boot (against 4abb64f, before the rebase; the change adds no boot work): headed.py boot --runs 3 --cpu 4, 3 rounds interleaved, 9 cold loads each, medians. Every phase but one equals main's (biome texture baked 433 / 436 ms, kits built 696 / 690, terrain water 20 / 20). terrain heights set is 1395 / 1471 ms: the wasm build's own tier-up draw (see "Issue #195": every build lands near 1390 or 1490 ms). That gives map_build_ms 3063 / 3131 and first_frame_ms 10958 / 11010. The change adds no boot work beyond one shader global and a toggle registration. The goldens and the smoke run are unchanged: GOLDEN OK 1950 71e9904d..., GOLDEN OK 1952 62f8ec61... and SMOKE OK 5 turns 1948-06-19 d3fdd459.... EditMode 747/747 on the merged head, with 0 warning CS.

Still different from the promo:

  • promo 03's sea has boats and building shadows in its dark tail (out of scope here: no civilian boats; superseded by #536, "Sea and boats");
  • promo 02's sea sparkles finely at the overview, where ours is held back from aliasing;
  • the open sea's turquoise near the coast reads a little paler against the brighter, textured open sea at diorama02. The shallow-to-open gradient is unchanged (19 levels on both), and the turquoise thins slightly (g - r 43 to 40, 12-24 px from the coast).

Sea and boats (branch sea-boats, refs #388)

The gap audit of 2026-10-13 (.superpowers/audit-2026-10-13/) found three gaps against promo 03:

  • the open sea at gameplay03 4x busier than the promo's (edge density 0.048 against 0.012, in its open-sea crop at 400 x 300);
  • the overview's sea flat;
  • 5 small wooden hulls with stub sails, where the promo has about 10 white-sailed boats and craft moored at Jaffa.

#402 had raised the far sea's texture on purpose, so the calm applies only far from the shore and only at the gameplay zoom. There, #402's surf, shallows and near-coast chop are unchanged. Everything below is behind a new seacalm toggle, and the calm and swells are also behind seatexture.

  • The zoom, not the pixel. The calm and the swells fade out, and the overview's shelf fades in, with the water's distance from the camera, not its pixel size, so a 1440p canvas or a device pixel ratio of 2 sees the same sea as 1080p. The calm and swells go between 110 and 190 km, and the shelf comes in between 150 and 220 km. The gameplay pose's water lies 50-90 km from the camera; the overview's (CameraController.MaxDistance, 250 km) lies 205-360 km away.
  • The calm (ScWaterShade, StatecraftMap.hlsl).
    • It runs from ScCalmNearM (3.6 km offshore, past the shallows, whose edge reaches at most 3.5 km, and past the surf) to ScCalmFarM (8 km), on the open Mediterranean only, not the rift lakes.
    • Across it the ripple eases to 30%, the mottle's fine octave to 30% and the glint to 40%, and the whitecaps go. The coarse streaks grow 1.3x.
  • Broad swells: a 14 km gradient-noise brightness variation, +-0.25, on the same far sea and at the same zooms as the calm, so the calmed water keeps promo 03's broad lighter and darker reaches. It is off at the overview, where its cells would read as blotches.
  • The overview's coastal shelf. The sea within 12 km of the shore lightens towards a pale teal (0.36, 0.62, 0.66), by 60% times how far past the shallows' gradient the water lies, fading to nothing at 12 km. That is promo 02's lighter shelf. It tints the shallows' outer gradient a little at the overview and is not applied to the rift lakes.
  • The boats (Kits/SeaCraft.cs, one more mesh under the kits' root, drawn with Kit.mat, so the "kits" toggle covers it):
    • 10 under way, 2.4 to 7.5 km out: 3 off Jaffa, 2 off Tel Aviv, 1 off Bat Yam, 2 off Haifa, 1 off Acre and 1 off Gaza. The Jaffa, Tel Aviv and Haifa points are the ones the kits build those ports at; the others are approximate shore points chosen for placement.
    • Two sailing rigs of the period: lateen-rigged feluccas and gaff cutters.
    • 7 moored: 2 in Jaffa's basin inside its breakwater, 2 at Tel Aviv's mole (in its lee or at anchor south of it, clear of the Jaffa label the gameplay pose shows north-west of it), and at Haifa 2 fishing boats by the quays and a small coastal steamer at anchor off the breakwater.
    • No motor launches or modern craft.
    • Each hull, plus a margin, lies in open sea and meets no breakwater or mole piece. The kits now record each piece as a "breakwater" in the census. SeaCraftTests.NoHullEverMeetsABreakwaterOrMole sweeps five landmark scales and 16 draws each.
    • Each boat is its real length x SettlementKits.ModelScale (the train's 24x, #1094): a felucca 12 m (about 290 m), a cutter 10, a moored fishing boat 8, the coastal steamer 50 (about 1,200 m). Design values, not sourced. Before #1094 a felucca was 34 landmark metres (2 km at the default scale).
    • Every boat heads east or west, give or take 35 degrees, so the north-looking gameplay camera sees its sails side-on.
    • The sails are lit as cloth turned mostly to the sky (each face's normal tipped up by 0.7). Lit as flat plates, the face the camera saw was often the one turned from the sun, and the sails read grey.
    • KitMesh.Triangle gains an overload with an explicit normal for this. The old overload produces the same vertices as before.
  • Jaffa's old boats are gone from SettlementKits.Port. Their random draws stay, so every later Jaffa house is drawn from the same stream. In pinned-clock renders at 25 km against main:
    • every pixel of the old town, the lighthouse and the clock tower is unchanged (0 of 371,200 differ);
    • so is Tel Aviv with its port;
    • the breakwater differs only on its 29 edge pixels where it meets the calmer sea; none inside it differs.

Measured: WebGL ?shot of main 58d7b5d8 (with #525's frame) against the branch rebased on it, with the audit's measure.py crops and #388's crop:

main branch promo
gameplay03 open sea: edge density (audit crop) 0.048 0.020 0.012
the same, sea only (x 0.02-0.22, clear of the boats) 0.044 0.015
gameplay03 open sea: luma contrast 0.054 0.055 0.052
gameplay03 sea, #388 crop: detail / spread 0.046 / 0.163 0.029 / 0.096 0.060 / 0.20
gameplay03 sea, focus band: detail / spread 0.055 / 0.185 0.037 / 0.104 0.075 / 0.32
diorama02 sea pixels (inside the frame): luma std / p10-p90 0.100 / 0.279 0.099 / 0.279 0.086 / 0.11 (audit crop)
gameplay03 / diorama02 palette distance 0.35 / 0.58 0.36 / 0.58
diorama02 warmth / saturation 0.129 / 0.36 0.129 / 0.36 0.145 / 0.43
  • #525's frame now crops diorama02's sea to the coastal strip. The audit's overview crop holds the frame, and the strip's own spread is already above promo 02's, so the "flat overview sea" gap is closed by the frame.

  • The coastal shelf moves diorama02's sea pixels by 1.7 levels on average (18% of them by more than 2). In the editor test below, on the open water off the strip, it is 0.086 against main's 0.048.

  • The same at any resolution. These are pinned-clock editor renders at 1920 x 1080 and at 3840 x 2160 (a device pixel ratio of 2, scaled back to 1080p to measure), comparing seacalm on and off:

    pose 1080p 2160p
    gameplay03, open-sea edge density, off / on 0.051 / 0.016 0.061 / 0.018
    gameplay03, share of sea pixels changed 61% 64%
    diorama02, share of sea pixels changed / mean change 18.4% / 1.68 levels 18.4% / 1.68 levels
  • Before the rebase (main dc4b272, no frame), the audit's overview crop measured contrast 0.030 on main and 0.031 on the branch.

  • #388's crop detail falls under #402's 0.035 floor. That crop's promo figure counts the promo's boats and Jaffa, and promo 03's open water alone measures 0.012, so the floor contradicted the target. The new floors are in "Sea round 2" and #388. The near-shore texture is unchanged at the gameplay zoom.

  • The shelf itself: SeaTests.TheOverviewSeaIsLighterAlongTheCoast. At diorama02, water 5 km out is 0.086 lighter in luma than water 25 km out; main measures 0.048, and the bound is 0.07.

  • SeaTests.TheOpenSeaCalmsAwayFromTheShore, at gameplay03's distance with the water's clock pinned: the fine texture 20 km out against 2.2 km out is 0.23 on the branch and 0.89 on main (bound 0.6).

  • Three tests compare seacalm on and off, pixel for pixel. Each was run red against a mutated shader:

    test what it checks mutation red
    TheGameplayZoomKeepsTheNearShoreSea gameplay03, 0.3-3.4 km out on nine rows: no change ScCalmNearM 0 12/255
    the shelf at the gameplay zoom 19/255
    TheOverviewGetsNoCalmOrSwells diorama02, 15-40 km out: no change no zoom fade 13/255
    TheShelfShowsAtTheOverview diorama02, 4-8 km out: changes by more than 3/255 (15/255 now)
  • SeaCraftTests checks the boats:

    • the count and the kinds;
    • every hull and its margin in open sea, not a rift lake, with no two boats overlapping;
    • the size against the hex;
    • no hull meeting a breakwater or mole piece, over 1,337 boats from 5 scales x 16 draws;
    • the seeding, and that the boats are drawn under the kits and toggled with them.
  • The census's cost. The kits now run with the census on, to read the ports and their works. The game towns' kits, built 5 times each way, take a median of 11 ms without it and 12 ms with it.

Frame time: headed.py profile, under the exclusive lock. The final round is main 58d7b5d8 first, then the branch, at a load of 4-6:

pose all main / branch -seacalm (branch) -water main / branch none main / branch p95 all
gameplay03 3.35 / 3.62 3.47 3.51 / 3.16 2.64 / 2.50 5 / 5
diorama02 12.69 / 12.56 12.36 12.73 / 12.41 8.37 / 8.11 14 / 14
  • seacalm prices the calm, the swells and the shelf at about 0.15 ms at gameplay03 and 0.2 ms at diorama02.
  • all rises 0.27 ms at gameplay03, against spreads of 0.14 and 0.33 ms, and falls at diorama02.
  • p95 is unchanged.
  • Two earlier rounds, before the zoom-keyed fades and on a busier machine (load 5-27), gave gameplay03 all +0.35 ms main first (spread 1.1-1.3 ms) and +0.97 ms branch first. In the branch-first round -kits and -seatexture rose by as much as all.

Boot (headed.py boot --cpu 4 --runs 3, one round each, before the rebase): map_build median 3389 ms on main and 3106 on the branch, and kits built 828 against 785 ms in the last run. The branch is not slower: the boats' placement and one more mesh are not measurable against the run-to-run spread.

Sea round 3: the surf as lace (branch visual/c1-foam-lace, refs #388)

Plan C1 part 2 (.superpowers/plan-C1-beach.md). The 2026-10-15 audit still saw one swash line and a smooth turquoise band where promo 03 has 2-4 broken white lines. Everything is in the water's own pass (ScWaterShade); the terrain function is unchanged.

  • Constants: ScSurfM 900 m to 1.8 km, ScBreakerSpacingM 300 to 450 m (ScBreakerRate kept, so 22.5 m/s), and ScFoamWhite 0.95 (0.85 before) on the foam's lerp to white.
  • The foam toggle (_ScSeaFoam, MapShaderGlobals): a 200 m value-noise net, kept under 0.62 with a fixed 0.25 edge, replaces broken so each crest breaks on its own (one 300 m blob of broken covered two crests 450 m apart).
    • A footprint-scaled edge aliased: TheBreakersAreFilteredToThePixel p99 0.067 at 140 m. The fixed edge keeps it at 0.0113 against the unchanged 0.014 bound.
    • The net's mean is 0.676 against broken's 0.5, so the breakers carry about 35% more foam at the gameplay zoom. Both fade to 0.5 at 75-150 m pixels, so the overview is unchanged and zooming out through that range dims the breakers by about a quarter.
  • Test: SeaTests.TheSurfShowsABrokenBreakerBeyondTheSwashAtTheGameplayZoom guarantees the swash and at least one breaker past it, at 3 of 4 moments, west of Tel Aviv and south of Jaffa, with a foam share in [0.3, 0.7]. Promo 03's 2-4 lines are not yet met; they are read again after #519's grade.
  • Profile (?vprofile headless, 1920x1080, exclusive lock, against main e6f98962): gameplay03 all 8.52 ms, -foam 8.48 (foam 0.04 ms, budget 0.5), none 7.41 (main 7.47), p95 11 (main 11); diorama02 all 16.47, -foam 16.52, p95 18 (main 18).
  • Bands (gameplay03 ?shot, south of Jaffa, before #519): foam band 9.0 to 20.6 px (promo 44, band 25-55), foam share 0.59 to 0.48 (promo 0.54), the strict lace-line median still 1 (promo 2); north foam 5.8 to 13.0 px.

Promo track: land use (issue #33, branch promo-land)

The flat mustard/olive ground becomes the promo's patchwork. tools/land_bake.py (no network, about 20 s, deterministic: test_the_committed_textures_are_the_bake checks the committed PNGs against a fresh bake) decides where from real data and paints what procedurally, baked:

  • where (158 x 211 m grid): fields on gentle slopes (< ~7%) with enough rain for dry farming (CHELSA bio12, station-corrected as hexgrid.py: smoothstep 180-380 mm; the irrigated Jordan valley floor at least 0.7); the citrus belt on the flat coastal plain and valley floor below 150 m near the 1948 towns (the hex-wise biome inputs softened over 1.5 km, so nothing steps at hex edges); garrigue, terraces and scrub where the rain is 250-600 mm, the ground unfarmed and not the shader's rock;
  • what (landuse.png, 2048 x 4096 RGBA, about 79 x 105 m): strip-field parcels in rotated 5 km blocks (green, ochre stubble, ploughed, fallow, grass, groves), soft grassy rims, hedges, some block edges as farm tracks; on the wet hills patches of stony soil and grass, terraces along the real contours every 40 m of rise, scrub; 2 x 2 samples per texel; sRGB colour (the sampler decodes it) spread into the empty texels so mips and BC3 blocks never blend with black; weight in alpha;
  • no roads between the towns: the first rounds joined each town to its three nearest with seeded meandering tracks (tracks.png); a controller ruling removed them, as invented routes would draw a second road beside the real 1948 roads and railways the roads track traces from the 1946 Survey of Palestine map (tools/routes_1948.py). A finer track network, if wanted, is traced there, not generated. The shader lays the land between ScGroundBase and ScGroundShade (relief, AO and breakup shade the fields), up to the waterline: promo-sea's beach band (ScShore, drawn after) covers it where there is a beach, and fields reach the water where there is none (the Sea of Galilee, cliffs). The land visual toggle lerps it away: a visual switch, not a price (the sample still runs).

Why baked, not procedural in the shader. The first cut drew the same pattern in the shader (about 150 instructions). WebGL headless, same night: base 2.72 / 9.76 ms (gameplay03 / diorama02 all), procedural 13.94 / 15.44; with a per-pixel [branch] and one shared noise 10.37 / 13.60; branch-free 9.58 / 13.58; flattened 10.91 / 15.01. The cost stayed when no pixel took the branch (-land 8.11 at gameplay03), and sampling the textures alone cost nothing (2.69 ms): most likely register pressure / occupancy from the shader's size (the GPU reserves registers for the worst path, and ANGLE may flatten the branch), not its arithmetic. Single builds differ by a few tenths of a millisecond, so the bisecting steps (fields alone +4.3 ms) are indicative only. The decision to bake holds either way.

Cost, measured base vs branch (separate builds; the toggle cannot price it).

  • Night of 2026-10-11, idle machine: main base 2.72 / 9.76 ms -> promo-land (4b3c307, before the review fixes) 2.68 / 9.93 ms (PROFILE summary visual=all pose=gameplay03 runs=3 median_avg_ms=2.68 spread_ms=0.09 median_p95_ms=5.00, ... pose=diorama02 ... median_avg_ms=9.93 spread_ms=0.02 median_p95_ms=11.00): within the noise between builds.
  • Review round 1 (24b540f on promo-sea), A/B against the same branch with the land call cut from the shader (Builds/Web-noland), profiled back to back while the machine was busy (another browser on the GPU: every number about 5 ms above the night's): gameplay03 all 8.20 -> 9.09 ms (+0.89, spreads 0.09 / 0.07), none 7.10 -> 7.73; diorama02 all 14.07 -> 13.70 (-0.37, inside the spreads 0.49 / 0.25). p95 10 / 16 ms.
  • Idle re-measure (re-review R1): not measured. Waiting for a 1-minute load average under 3, checked every 5 minutes, 2026-10-11 09:02-10:47 and 14:32-15:32: it went under 3 once (2.64 at 09:22), but the A/B taken then was still inflated (none 7.36 ms at gameplay03 against the night's 2.33) and the branch half was disturbed (spreads up to 10 ms), so it is not used; otherwise the load stayed between 3.9 and 51. The busy-machine A/B above is the only cost figure; re-measure on an idle machine before the tier numbers are re-ratified.

Size and memory. landuse.png is 10.7 MB in the build report (2048 x 4096 BC3 with mips: 10.7 MiB of GPU memory). The WebGL .data is LZ4-compressed, so the download grows less: Web-disabled.data 21.53 MB (promo-sea) -> 24.40 MB (+2.86 MB, then with the since-removed tracks texture's 0.02 MB). Halving the texture (the platforms' 2048 cap gives 1024 x 2048) blurred the hedges, rims and rows (gameplay03 local gradient on the plain 0.0220 -> 0.0174, -21%), so it stays 4096 rows; BC1 cannot hold the weight, so BC3 stays.

Platforms. The default web build targets desktop browsers (S3TC/BC, as the map's other textures already need). The land texture's WebGL override pins only the size (4096) and leaves the format automatic, so the build's texture subtarget picks it: DXT5 by default; Build.Web -textureSubtarget astc (or etc2) makes a build for browsers without S3TC (iOS Safari, most Android Chrome), where the land is ASTC 6x6, 4.8 MB (unity-build-astc.log), and golden + smoke=5 pass. A desktop build opened on such a browser would decompress the land to RGBA32 on the CPU: 42.7 MiB. WebGL2 guarantees only 2048-texel textures; virtually every GPU offers 4096.

Tests, review round 1: EditMode 465 passed (LandUseTests: import settings for every platform, the subtarget argument, -extraPose validation, the land follows slope/rain/towns, gameplay detail; TheTerrainShaderCompilesForWebGL2); tools pytest 194 passed; dotnet test 628; golden and smoke unchanged on Web-disabled and Web-astc.

Colour (WebGL ?shot, checkpoint promo-land-r5 against promo-sea-r2): gameplay03 saturation 0.4792 -> 0.4681 (promo 0.38), warmth 0.1018 -> 0.0921 (promo 0.064), contrast 0.1738 -> 0.1730, palette distance 0.5971 -> 0.5148; diorama02 saturation 0.4183 -> 0.4164 (bar <= 0.48), warmth 0.1171 -> 0.1154 (bar 0.087-0.145), palette distance 0.5869 -> 0.5868. Open item 3: the land lowers gameplay03 saturation by 0.011. Open item 5: the olive plain / gold hills seam is now fields against garrigue, a softer step (owner to judge). Licence: CC BY-SA 4.0, decided by the owner on 2026-10-14 (D13 in docs/oss/DECISIONS.md). Still different from the promo: soft at close zoom (rows, hedges and bushes read at the gameplay zoom, blur at 15-20 km and closer); terraces read faintly; the orchard ground and promo-art's groves follow separate rules; the promo's ground is warmer and browner under its grade.

Land round 2 (branch land-round2, refs #33)

Round 1's land read as yellow-olive. It also lost most of its field lines, because the bake's 2 x 2 supersampling averages an 18 m hedge or a 30 m track away inside a 79 x 105 m texel. Round 2 changes only the colours and the bake; the shader, the texture's size and format, the trees and the sea are untouched.

  • Ground colour. The land's colours in tools/land_bake.py (SRGB) and the plain, valley, lowland, hill and mountain colours in MapPalette.Ground move toward promo 03's ground, (177, 144, 114) in the ground crop below: less green, more blue, at its hue (28.6 degrees). The hill colours (soil, hillgrass, stone, and MapPalette's hills and mountains) are an orange brown at 22-25 degrees, saturation 0.24-0.34. With round 2's first values (soil at 8 degrees and saturation 0.26; stone, dust and fallow at about 18) the Judean hills read rose (review of #392): a red hue at low saturation. At 28.6 degrees they rendered at 34, yellower than promo 03's hills (26) and the video's (28), and diorama02 lost promo 02's browns, about (208, 144, 112) and (176, 112, 80). The brightness (luma) of each colour is kept. The desert is unchanged.
  • Fields. Hedges and tree lines are HEDGE_HALF_M 75 m either side of a parcel edge. Farm tracks are TRACK_HALF_M 70 m either side of a block edge. Each line is chosen once per edge, so the parcels on both sides draw it. The crops are further apart in luma: a darker green, darker ploughland, paler stubble and fallow. test_field_lines_are_wide_enough_for_the_texture holds both lines to at least 1.5 texels across. It fails at round 1's widths, measuring 15 m and 28 m.
  • Hills. Round bushes about 200 m across, at most one per 300 m cell. A 1.5 km noise sets the share of cells with one (20-80%), so they clump. Each sample takes its dot from the 2 x 2 cells round it, so no dot is cut. Each bush has a softer shadow offset east-south-east, away from the sun (MapLighting.SunAzimuth 285 degrees). The garrigue's weight rises from 0.7 to 0.9. Terraces cover more of the slopes, with stronger stone lines.
  • Bake. About 30 s. Two runs give byte-identical PNGs. test_the_committed_texture_is_the_bake passes. The land weight mean is 0.192 (round 1: 0.171), because the garrigue's weight rose.

Measured: WebGL ?shot at main 4abb64f, then the branch on it, art-log rows lc-r2-main and land-round2-fix5. The ground crop is gameplay03's land (x 0.36-0.88, y 0.30-0.82), pixels at luma >= 0.4. The towns' tan walls and their ground tint are in it and unchanged, so it moves less than the ground does. The Judean hills are -extraPose judea30:35.10:31.85:30000:10 (editor, web tier): the median hue of their pixels at luma >= 0.25.

main round 2 promo
gameplay03 saturation / contrast / warmth 0.447 / 0.181 / 0.070 0.429 / 0.178 / 0.062 0.38 / 0.203 / 0.064
gameplay03 palette distance 0.435 0.378 -
diorama02 saturation / contrast / warmth 0.414 / 0.184 / 0.113 0.406 / 0.183 / 0.109 0.43 / 0.200 / 0.145
diorama02 palette distance 0.5633 0.5643 -
gameplay03 ground crop (sRGB) (166, 148, 99) (169, 143, 106) (177, 144, 114)
judea30 hill hue (degrees) 39.8 30.5 26 (promo 03), 28 (video 0:21)

The bars hold: diorama02 warmth 0.109 is inside [0.087, 0.145], and its saturation is under 0.48. diorama02's palette distance is 0.001 above main's; promo 02's browns (about (208, 144, 112)) are what the hills' hue trades against. Every ground channel is within 8 levels of the promo's (it was within 15).

Frame time: headless, main and the branch back to back under one exclusive lock, 3 rounds, rotated. The load average was about 6 from other sessions, so the absolute times are inflated. The table gives medians over the rounds, in ms:

pose all main / branch -land none
gameplay03 8.13 / 8.05 8.05 / 7.81 5.09 / 4.97
diorama02 19.49 / 19.71 19.08 / 19.45 12.36 / 12.72

Every difference is within about 0.4 ms, which is also how far none moves between the two builds. The review fix (the hill hues) changes only colours: the same texture size and format, no shader, so these numbers and the boot below stand. The shader and the texture's size are the same. diorama02's p95 is 21-22 ms on main and on the branch alike, under that load.

Boot: headed.py boot --runs 3 --cpu 4, 3 rounds interleaved, 9 cold loads each, medians:

  • map_build_ms: main 3032, branch 3039;
  • first_frame_ms: main 10842, branch 10819;
  • biome texture baked (MapPalette's colours): main 437, branch 431;
  • downloaded 53 files: main 1048, branch 1040.

Both builds' map build is over #321's budget at this load; that is main's, not this change's. The goldens and the smoke run are unchanged: GOLDEN OK 1950 71e9904d..., GOLDEN OK 1952 62f8ec61... and SMOKE OK 5 turns 1948-06-19 d3fdd459.... The tree count is unchanged (7,814).

Still different from the promo: the trees and bushes, so far only painted in the bake (promo 03's land is 27% dark trees and shadow; ours is 10%). The grade's shadows are cool, where the promo's are warm brown (#391).

Open items for the art track

Each item below was ruled to the art track (controller, rank 10). None blocks landing the branch.

  1. Border skirt look. The skirt (the slab around the map) reads as a different material from the terrain next to it. The fix is the table or frame and the skirt in the terrain's own look (promo 02), not a colour patch. A patch, BorderLandColour (0.76, 0.66, 0.47), was measured and rejected: the south-west slab rendered (0.69, 0.58, 0.37) against the terrain's (0.82, 0.69, 0.43), so the step shrank, but the skirt is 17% of diorama02 and its saturation rose 0.34 -> 0.47. That took the frame to saturation 0.488 against the 0.48 bar (warmth 0.129). Do not move the bar to exclude the skirt. The earlier line saying a lower-chroma colour is the next candidate is withdrawn. Done in #36 (frame round the map). The skirt is gone: the board ends in a bevel (cut earth where the edge stands above the frame, the frame's lip where the sea and the low coast meet it), a walnut frame and a dark wooden table (MapBuilder.BuildFrame, MapFrame.shader). The frame and table read the haze at sea level or higher (at world y 0 the rift boost made it 1.6x), and the table takes 0.4 of that; the full haze lifted it to a flat khaki. WebGL diorama02 against main (c2774ecc, with #502's olives): saturation 0.405 -> 0.359, contrast 0.173 -> 0.222 (promo 0.200), warmth 0.118 -> 0.129 (bar 0.087-0.145). gameplay03 and both hero shots: no land pixel moves by more than 3/255; 1 sea pixel of gameplay03 moves by up to 13/255 (the water's own clock, unpinned in the player). The sea no longer runs past the map's west edge, so the sea-seam tests went with the skirt.
  2. Settlement markers float. At the web shadow texel (about 100 m at 60 km with a 2048 map) the markers (symbolic scale, 300-700 m) show long detached shadows, so they look as if they hover. Set the sun's shadow bias (depth and normal bias) at kit scale: offsets of about 100-150 m at that texel. Do not just turn marker shadows off.
  3. Saturation of gameplay03 and warmth of diorama02. gameplay03 saturation is +0.18 over promo 03, mostly the sea (0.83 against 0.51) and the gold hills biome. Sea part done in promo-sea (frame 0.57 -> 0.48, sea 0.82 -> 0.49); the hills biome remains. Measure the SeaTests margins before any sea change. diorama02 warmth (0.1038) sits near its 0.087 floor, so any sea desaturation must re-check it.
  4. Tilt-shift bottom band. The bottom blurred band measures 17-22% of the frame against about 15% in the promo. It awaits the owner's verdict on the offer BottomHalfWidth 0.16 / BottomFeather 0.19.
  5. Olive plain against gold hills. The seam between the olive coastal_plain/valley palette and the gold hills biome is for the owner to judge.
  6. Overview shadow overrun. The shadow pass overruns the budget by 0.1-0.4 ms at the overview pose; the levers are in "Tier decisions".
  7. Sea brightness spread (promo-sea re-review N3). The sea's luma p90 - p10 is 0.12 at gameplay03 against promo 03's 0.39, and 0.08 at diorama02 against promo 02's 0.15. Its saturation p10/p90 is narrower too (0.43/0.55 against 0.29/0.58). The means match ("promo-sea", Measured). The causes:
    • promo 03's dark tail is boat hulls and building shadows on the water. Boats are out of scope: navy ships are units, and there are no civilian boats (owner). Superseded by #536 ("Sea and boats").
    • Its bright tail is foam lace and glints. The glint is held at 0.06 by the seam test (Task 7), and the surf is a smooth band, not lace.
    • diorama02's sea has a near-to-far gradient that ours keeps flat.

Rulings digest

One line each: ruling - reason.

  • White balance Temperature 5 (plan: 9) - the merged diorama warmth must land below the promo's 0.145 (measured 0.127-0.133 at 9).
  • AmbientProbe removed - dead code: Trilight ambient replaces any assigned probe, so it never rendered.
  • Sea glint 0.06 (was 0.25) - at 0.25 the lobe lit a warm patch over 6 km of open sea at yaw 180 and broke the seam test.
  • Skirt hold smoothing (running mean, 8 km beyond the brief) - the skirt follows the coast without steps.
  • Grid opacities land 0.35 / water 0.10, plus a 30 -> 10 px footprint fade and fade to 0.35 - the sea lattice was about 5x the promo's weight; land lines were invisible at 0.25 once graded.
  • Grade Saturation 0 (plan: 16) - the palette now carries the colour; graded/ungraded saturation x1.042 (bar 0.98-1.12).
  • Selection ring ivory (promo ring g/r 0.915, b/r 0.785) - matches the promo and passes the colour-blind separation tests.
  • Olive coastal_plain/valley palette - gameplay03 was one flat mustard; warmth 0.108 -> 0.077, saturation 0.582 -> 0.565.
  • Web shadow map stays 2048 - 1024 measured equal cost, so shrinking would lose quality for nothing.
  • Web render scale 1.0 - measured WebGL tier decision (see "Tier decisions").
  • Bloom, SSAO and the depth/opaque copies off - bloom changed at most 2/255 of any pixel; see "Tier decisions" for the costs.
  • NOT taken: dropping the bake's R channel - zero memory gain (BC1 either way) and it churns the assets; R stays as the AO reference.
  • NOT taken: deduplicating the OpenMap helper across SeaTests and TerrainShaderTests - test-only duplication; refactor risk with no behaviour gain.

Art spike: procedural settlement kits and trees (branch promo-art)

Question: can code-generated settlement and vegetation art at Civ-style symbolic scale bring gameplay03 close to promo 03? Code: unity/Assets/Scripts/Map/Kits/ (built at load, seeded per settlement id by KitRandom, no sim state touched), Assets/Shaders/Kit.shader (Kit.mat buildings, Tree.mat instanced trees), KitTests.

  • Archetypes (SettlementKits): white city (Tel Aviv, Haifa: 2-9 storey flat-roof blocks, ribbon windows and balcony lines in the shader, L-shaped wings, roof rooms), old town (Jaffa, Acre, Gaza, Jerusalem: packed stone houses, domes; Jaffa's clock tower, belfry, mosque, breakwater, lighthouse and boats), Arab town (tan cubes on slopes, a mosque), moshava/kibbutz/moshav (red gable roofs, yards, a water tower; Nahalal as a ring). One merged mesh per settlement; all towns share one lot list, so a smaller neighbour never overlaps a bigger town's buildings. The kits toggle off shows the marker boxes again.
  • Footprint = (1945 population / 60,000) hexes, floored at 0.08 hex. The populations live in content/map/settlement_population_1945.yaml (Village Statistics, 1945, read from the Wikipedia table on 2026-10-11: owner-verify until checked against a scan; Beersheba's row there is garbled, 5,570 is its likely reading), read through MapInputs.PopulationYaml. The sim never reads the file: SaveGameTests.TheKitPopulationFileIsOutsideTheContentHash, and the player's BOOT OK content 96035735... is unchanged. Towns share seams (distance over radius), stay on land, and tint their ground in the baked biome texture.
  • Trees (Vegetation, VegetationLayer): citrus groves in 2x2 patches (every tree of a patch checked for water and towns, else one tree) round the coastal-plain towns, olives over the hills (and lone ones on the plain), palms by the coastal towns, cypress windbreaks; Graphics.RenderMeshInstanced per 30 km tile from beginCameraRendering, thinned as (110 km / distance)^2 past 110 km. The WebGL cost is per instance, not per triangle (at 25x a 30% triangle cut changed nothing; the 2x2 patches took the instances from 51,000 to 21,165 and the gameplay03 cost from 2.9 to 1.3 ms).
  • Scale (owner ruling 2026-10-11, superseding 10-25x): buildings and trees match promo 03's measured scale per class: city blocks about 60x real size, houses and trees about 100x (KitScale; landmarks take the city-block scale). Counts were tuned at 25x and fall with the square of the scale. ?kitscale=, ?housescale=, ?treescale= (clamped to 5-200) override them in the player; MapCapture.Kits takes -kitScale, -houseScale, -treeScale.
  • #37: the kits use the sun's own URP shadow bias, unchanged (review I3: a kit-only bias and receiver offset were removed, as the probe could not show a need for them). KitTests.AKitShadowStartsAtItsWalls walks out from a lone box's shadow-side wall from 1 m in 1 m steps, on both tiers: the first shadowed sample is 3 m out at 12 km (pixel 7 m), 10 m at 60 km (pixel 35 m) and 1 m at 250 km, the same as with the old marker material, and the sunlit roof darkens at most 1.9% with shadows on (no acne). No bias gap reproduces, so the floating look of the old markers was not a bias gap.
  • Grid readability: KitTests.TheGridStaysReadableThroughTownsAndGroves: at gameplay03, 76 of the 90 land edge midpoints whose line shows on bare ground still show it with kits and trees on (84%; the floor is 75%).
  • Measured at the 60x/100x defaults (d4e7275, headless WebGL, machine_lock.py --exclusive --wait-unity, load 4-9, other sessions running; "none" means every visual feature off, not only the kits and trees): gameplay03 all 2.34 ms (p95 3), -trees 2.57, -kits 2.34, none 2.34; diorama02 all 11.22 (p95 12), -trees 10.41, -kits 11.35, none 7.29. Trees -0.23 / 0.81 ms and kits 0.00 / -0.13 ms (gameplay03 sits in the ~1 ms clock's bimodal 2.0 / 2.4 ms noise, so its negative costs are noise): every feature is within budget, all - none 0.00 / 3.93 ms, p95 3 / 12 ms. The main build in the same session: gameplay03 all 2.78, diorama02 all 10.12 (so +1.10 ms at diorama02). The diorama02 shadow overrun is unchanged (2.95 ms against main's 2.91). Headed DPR 1: all 2.46 / 10.95, trees -0.24 / 0.75. Headed DPR 2: all p95 6 / 20 ms (floor 33), trees 0.35 / 0.89. Boot (headed.py boot --runs 3, main then this build): cpu 1 map_build 491 -> 508, first_frame 3631 -> 3610; cpu 4 1874 -> 2035, 6795 -> 6991 (within max(300 ms, 50%)). The player logs KITS buildings=306 vertices=12374 triangles=8556 trees=8147 build_ms=20 plant_ms=19.
  • After merging main with promo-sea and promo-units (689880d, headless, load 18.7 at the start and 4.4 at the end, so absolute times are noisy): gameplay03 all 9.57 (p95 11), -trees 8.75, -kits 9.45, none 7.47; diorama02 all 14.99 (p95 16), -trees 14.28, -kits 15.40, none 10.92. Trees 0.82 / 0.71 ms and kits 0.12 / -0.41 ms stay within budget, as do all - none (2.10 / 4.07 ms) and p95 (11 / 16 ms). The rise in "none" (2.34 -> 7.47 ms at gameplay03) is not the kits (both "none" runs draw no kits or trees); it should be re-measured on a quiet machine against main.
  • Build size (WebGL, main -> this branch): uncompressed +318,885 bytes (+0.44%: wasm +286,873, data +28,978, framework.js +3,034); Brotli (quality 11, as shipped) +33,407 bytes (+0.17%: wasm +13,092, data +20,091). No texture or model files: everything is generated at load.
  • Authored meshes later: trees swap as is (Vegetation.Build(kind) returns a mesh in real metres). Buildings will need Emit to return placements (archetype, variant, position, yaw, scale) drawn through the same tile instancing and a textured, instanced variant of Kit.shader; that refactor waits for the first authored mesh.

The 1948 localities and Bedouin camps (branch promo-villages)

Owner rulings (2026-10-11): show the places of 1,000 or more people first, by their 1948 names, as town and village kits beside the 25 game settlements, and a light Bedouin encampment scatter sized by the Village Statistics tribal totals, inventing no names or points. Owner rulings (2026-10-12, after the staff review): the "Today:" tooltip only where a living place carries on; Lydda and Ramle for the game settlements lod and ramla; Nesher drawn as a mixed locality.

  • Data: content/map/localities_1948.yaml (status draft; its header lists every field and rule). It has 271 places and 8 camp areas.
    • The places are the 293 of 1,000+ in VS 1945, less the 20 game settlements among them, the Jahalin tribe row (counted with the camps) and Huj (810 plus 230 Jews of kibbutz Dorot: the village alone is under 1,000).
    • The camp areas total 64,670 people: Beersheba 47,980, Jerusalem 7,070 (its "Tribal Units" row only), Ramle 3,780, Jaffa 2,270, Hebron 2,000, Safad 820, Gaza 530 and Nablus 220. Tulkarm's "Tribal Units" row has no figure.
    • The file is presentation only, outside the content hash (SaveGameTests.TheLocalitiesFileIsOutsideTheContentHash), tools/tests/test_localities.py checks it. The player reads a baked runtime form, localities_1948.tsv (see "Boot" below), through MapInputs.LocalitiesTsv.
    • Left for later rounds: places under 1,000, the 70 VS rows with no point, and the 29 foundings dated only "1948".
  • Today's name: a today entry ({name, name_he, source}) exists only for a place that carries on. There are 161 of them.
    • None is given where Wikidata classes the place's item as a depopulated village or a former settlement. The test's QID list is tools/tests/data/localities_no_today_qids.txt; the content records no such class.
    • There are two exceptions:
      • Majdal, whose item is today's Ashkelon (the owner's example).
      • Sha'b → Sha'ab (שעב). Its item, Q2655669, is also the living Arab local council. This applies the owner's rule (2026-10-12 re-review).
    • Walaja, Samakh and Lubiya are classed both ways too, but no living place carries them on, so they have no today name.
    • Jewish places merged into a later municipality take the municipality: Magdiel and Ramatayim become Hod HaSharon, and Kfar Ganim becomes Petah Tikva.
  • Kits:
    • Villages and Arab towns take the Arab-town archetype (village type: a smaller mosque).
    • Jewish places take the red-roofed archetype: kibbutz, moshav ring or moshava, and the 14 Jewish towns too, since the repo maps them to town and several began as moshavot.
    • Nesher (VS: 1,430 Muslims, 1,400 Jews) is a mixed locality: Arab houses on its west side and Jewish houses on its east, with a mosque and a water tower. Its tooltip says "Mixed locality".
    • No Templer colony reaches 1,000. When they come, the moshava archetype is the closest.
    • The landmark follows VS's religion columns: 11 churches (Shafa 'Amr by its Urban row) and none in the 5 Druze villages.
    • Every place shares the one lot list. Every game settlement's landmark claims its spot first, then every locality's, then all houses, largest place first. The 25 game settlements keep their cells and their landmarks (tested); a village fills in around them.
    • Localities and camps are merged into one mesh per 30 km tile.
  • Camps: one per 1,000 of an area's total, rounded half up with no floor, so Nablus draws none: 65 camps, 175 tents (after #250).
    • The areas are the sub-districts as content/map/subdistrict_hexes_1948.yaml draws them, read at load (Localities.FillHexes).
    • Each camp is placed at a seeded spot, dry, gentle, outside every town and off every route.
    • It has two to four brown goat-hair tents in a loose line.
  • Labels: smaller pills (.map-label-locality), only nearer than 40 km (MapLabelsPresenter.LocalityDistanceM), placed after every game settlement with the same declutter rules.
    • Hovering one shows its tooltip, in the UI's language (ui.lang).
    • A touch tap on a pill shows it, and a second tap on the same pill, or a tap anywhere else, hides it (#245; MapLabelsPresenter.Tap, called from GameRoot). Keyboard players still have no route to it: only a hovering pointer or a touch reaches the pills.
  • Units in play (review I5, merged with #250): the unit tokens keep off the villages and camps as they keep off the game towns.
    • The units are placed before the kits are built, since their clearings feed the kit build. So MapScene.SettlementFootprints, which the units place against, holds the game settlements' and the localities' discs (SettlementKits.FootprintRadiusM), the same discs the kit build claims.
    • A token standing in a town or village gets a clearing. The kits take the clearings as corridors, so village houses, tents and trees keep off them, as they do on main.
    • The camps are placed after the units. They keep off every token outside the towns (KitGround.Tokens): a camp's centre stays SettlementKits.CampReach house-scale metres plus the token's ring away from it.
    • UnitsInPlayTests.NoTokenStandsInAVillageOrACamp builds the map with the kits on. It checks that every token outside a town has its ring clear of every MapScene.KitFootprints entry (towns, localities, camps), and that every token in one has its clearing.
  • Grid readability with the villages: 79% (70 of 89 edge midpoints) at gameplay03, against 84% before. The floor stays at 75%.
    • Merging #250 made the units keep off the villages and the camps keep off the units. The camps and village houses moved, and the share fell to 72%, under the floor.
    • The fix is the plan from review M10. A 1948 locality's houses and a camp's tents keep off the edge midpoints of the hex they stand in (KitGround.OnGridMark). This costs a few locality buildings (1,332 after #251) and a few tents (175).
    • The game settlements are unchanged.
    • The next data round (places under 1,000) gets the same rule. Don't lower the floor.
  • Bands stay legible through the villages: BorderTests.TheBandsStillReadThroughTheVillages reads the bands with the kits and trees on (18 of 18 walks at gameplay03 and Tulkarm). The other border tests measure the bands with the kits and trees off.
  • Start of game only. The file records no later fate, depopulation date or post-May-1948 status, and the test test_start_of_game_only_no_later_fate keeps it so. A dated map (#61) needs the owner's framing decision first: whether the settlement layer is ever dated after the start, and if so how places that ceased to exist are shown. The research note .superpowers/handover-2026-10-11/promo/localities-1948.md ("Sensitivity note") lists the options.
  • Measured.
    • Frame time (headless WebGL vprofile, exclusive lock, aa1a40a against main c2089b4, 2026-10-12):

      gameplay03 diorama02
      all 13.83 ms (p95 15); main 14.33 (16) 18.01 ms (p95 20); main 17.53 (20)
      all − none (budget 8 ms) 3.33 ms; main 3.91 5.29 ms; main 4.96
      trees (budget 2.5 ms) 0.48 ms; main 0.52 1.26 ms; main 0.90
      labels 0.02 ms 0.06 ms

      The kits' cost is within the noise. The 20 ms p95 at diorama02 is main's too.

    • Boot (headed.py boot --runs 3, main c2089b4 → this branch, 2026-10-12; budgets at cpu 4: 2720 / 10445 ms):

      map_build first_frame
      cpu 1 601 → 640 ms 3860 → 3884 ms
      cpu 4 2506 → 2572 ms (148 ms under budget) 8141 → 8006 ms

      Before this review round, cpu 4 map_build was 2712 against 2720.

      Re-measured after merging #248, #250 and #251 (2026-10-12, cpu 4, three runs, exclusive lock):

      build map_build first_frame
      main f974c9f 2581 ms 8114 ms
      this branch 2768 ms 8508 ms
      this branch, light read of the sub-district file 2745 ms (25 over 2720) 8489 ms
      the same after merging #253, #255, #261 and #267 (4b9283a) 2850 ms (130 over) 8904 ms
      • The localities parsed phase included a second, full parse of subdistrict_hexes_1948.yaml (the camps' areas, review I4), done with the reflection deserializer. GameRoot parses that file in full again later in the boot.

      • The map build now reads only the regions' keys and the hexes' regions, with the event reader (Localities.FillHexes(doc, string, ...)); KitTests checks that it gives the same areas. The phase went from 221 to 172 ms.

      • Main itself grew 75 ms since c2089b4 (#250's unit placement, #251's routes). The branch adds about 165 ms over main at cpu 4, nearly all of it the localities file's YAML scan.

      • Resolved by a baked runtime file (coordinator ruling 2026-10-12: fix the root cause, not defer the kit build or re-baseline the budget):

        • localities_1948.yaml stays the sourced, hand-edited source of truth.
        • tools/localities_bake.py bakes it into localities_1948.tsv: 26k characters against 176k, one tab-separated row per place or camp.
        • The player reads the file with Localities.ParseRuntime (a split, no YAML), with the same checks as the YAML reader. Errors name file:line.
        • The manifest ships the .tsv and no longer the YAML.
        • pytest fails a stale bake; KitTests.TheRuntimeFileYieldsTheSameRowsAsTheYaml compares every row.
        • The localities parsed phase went from 172 ms to 53 ms. Main and the branch now boot within noise of each other.

        Measured at cpu 4, three runs each, interleaved, exclusive lock, 2026-10-12:

        build map_build first_frame
        main 316a042, round 1 2573 ms 8113 ms
        this branch f715717, round 1 2588 ms 8229 ms
        main 316a042, round 2 2610 ms 8234 ms
        this branch f715717, round 2 2553 ms 8131 ms
        this branch after merging main (05390da), final checks 2658 ms 8367 ms

        The branch is 130 to 170 ms under the 2720 ms budget: real headroom for the towns' round 2. That round should bake the same way.

    • The growth is the localities file's parse: the localities parsed boot phase is 50 ms at cpu 1 and 220 ms at cpu 4. It is not the ground tint, which the player logs as KITS ground_tint_ms=0.8. My earlier attribution to the tint was checked with that timer and was wrong; the earlier "terrain heights set" growth was the parse that ran before it.

    • Two changes cut the parse:

      • An event-based reader (Localities.Read) replaces YamlDotNet's reflection deserializer.
      • Each row's sources are written as a plain URL list, so the file went from 266k to 176k characters. Raw scanning is about 90% of the parse, so file size is the lever.
    • The tint was made cheaper anyway: rows are clipped to the circle, the inner half skips the smoothstep, and the blend is in integers.

    • The next data round (places under 1,000) must not ship as a bigger file of this kind without re-measuring.

    • History: a first version scanned every lot and footprint linearly, taking 350 ms (kits) and 330 ms (trees) in the editor. DiscIndex (a 4 km grid) brought that to about 30 ms each.

    • Build: wasm +111,913 bytes and data +14,441, plus the localities file fetched at boot.

Towns round 2: dense, era-correct towns (branch promo-towns-2, issue #31)

Question: why do Tel Aviv, Jaffa, Lod, Ramla, Rishon and Petah Tikva read as a sparse handful of boxes at gameplay03 when promo 03 shows dense miniature towns, and what closes the gap within the frame-time and boot budgets?

Diagnosis (main bdb5cd2, measured before any change)

Measured with TownCensusTests.TheCensusIsLogged. It builds the kits as MapScene.Build does, with the routes' and the units' clearings as corridors, and logs one CENSUS line per town. It reports:

  • Buildings: counted per town. Wings are a block's lower part.
  • Cover: the share of the town's own land under a building or landmark footprint. "Own land" is its footprint disc, on land, in its cell. It is sampled every 100 m, over the whole disc and over its inner half radius ("core").
  • Wall heights: symbolic metres; one storey is 198 m at 60x and 330 m at 100x.
  • Wall colours: distinct sRGB colours at 4 bits per channel.

The player logged KITS buildings=261 vertices=10244 triangles=6892 trees=7283.

town footprint R buildings (+wings) landmarks cover: all / core wall height min / median / max wall colours
Tel Aviv 7.6 km 43 (+11) none 24.7% / 24.9% 396 / 594 / 1188 m 7
Jaffa 5.7 km 16 mosque, clock tower, church, lighthouse, breakwater 42.6% / 14.6% 330 / 660 / 990 m 4
Haifa 6.9 km 34 (+10) none 17.8% / 23.7% 396 / 594 / 990 m 8
Jerusalem 7.4 km 70 a generic mosque 32.2% / 16.2% 330 / 660 / 990 m 8
Lod 2.4 km 5 mosque 13.6% / 11.8% 330 / 330 / 660 m 4
Ramla 2.3 km 5 mosque 19.5% / 27.6% 330 m only 3
Petah Tikva 2.4 km 3 water tower 10.5% / 7.0% 330 / 330 / 660 m 3
Rishon LeZion 1.7 km 2 water tower 9.5% / 2.7% 330 m only 2
all 25 235 (+21) 26

Jaffa's whole-disc cover is above its core cover because its cell is small (Tel Aviv's footprint takes its north) and four landmarks of up to 1 km sit in it.

Causes, in order of weight:

  1. Counts. TargetCount was tuned at 25x and falls with the square of the scale. At the 60x / 100x ruling, every town under about 30,000 people gets its floor of 6 buildings, and then places fewer: 1 to 5 in the moshavot, and 5 in Lod and Ramla.
  2. Placement. Buildings went to random spots over the whole footprint (radius drawn as u^0.62), tried against every lot. Each town had about 30 tries per building, and the footprint could stretch to 1.6 R.
    • That spreads a town evenly at low cover, with no dense core and no street rows. Promo 03's towns are a solid core with streets as the only gaps.
    • The routes' corridors (300-680 m ribbons plus a 100 m margin) and the units' clearings then cut each disc further.
  3. Old-town houses too big. Old towns drew 7-14 m houses at 100x, so each was 0.7-1.4 km wide. Jaffa came out as 16 large blocks, not a packed old city.
  4. Heights. Tel Aviv's median block was 3 storeys and its tallest 6. Promo 03 shows 5-8. The owner's "promo over era" ruling covers this.
  5. Landmarks. Only Jaffa had its own. Jerusalem had a generic mosque, and Haifa and Tel Aviv had none. Jaffa's lighthouse stood at the breakwater's head; the real one (1865) stands on the hilltop above the old port.
  6. Palette. It is not the gap: each archetype's colours already match promo 03 (white and cream city, honey stone, tan villages, terracotta roofs). Variety is limited by the counts.

Plan

  • Plot lattices replace the random tries (SettlementKits.PlaceHouses).
    • Each town is laid out in rows of plots at its street grid's angle. Old towns and villages stagger alternate rows and jitter each plot.
    • Each plot is built with a probability that is near 1 in the core and falls towards the edge, to none at 1.1 R. That is UnitGround.FootprintMargin: the units assume no building past it, and PaintGround tints the same disc.
    • A plot by a road, or by another town's lot, tries a smaller building, then a small nudge within the plot, before it stays empty.
    • Plots are visited nearest the centre first. Each plot draws its random numbers whether or not it is built, so one plot's fate never shifts another's.
  • Building styles per quarter, at the ruled scales (city blocks 60x, houses 100x).
    • Tel Aviv: white International Style blocks with flat roofs (the White City, 1930s, Geddes plan 1925-29), some stepped. Taller in the core under the promo ruling.
    • Jaffa and the old towns: small stone houses (4.6-6.6 m real), packed, with domes and roof rooms. Some red tiled roofs outside the core.
    • Haifa: the lower town by the port in stone houses (below about 35 m elevation). Hadar HaCarmel up the slope in white blocks.
    • Jerusalem: stone-faced blocks in the new city to the west (Storrs' 1918 stone ordinance), a few with red roofs. Arab-quarter houses round the other sides. Old-town houses within the walls.
    • Large moshavot (Petah Tikva, Rehovot, Rishon LeZion, Hadera; 7,500+ people in 1945): a plastered two- and three-storey centre inside red-roofed houses.
    • Arab towns and villages: smaller stone houses (4.4-6.6 m) on a staggered lattice.
  • Landmarks, placed before any town's houses. Positions keep the real bearing from the settlement's point to the landmark's sourced coordinate; distances and sizes are symbolic.
    • Jaffa: the Mahmoudiya Mosque, the clock tower on the plaza east of it, St Peter's Church, the lighthouse on the hill above the port (red and white bands), and the breakwater and boats.
    • Jerusalem: Suleiman's walls round a symbolic square Old City with five gates, the citadel, the Haram platform, the Dome of the Rock and Al-Aqsa.
      • The Dome of the Rock is blue and white tiles with a dark lead dome: the gilded dome dates from the 1959-64 restoration, so it would be an anachronism in 1948.
      • Al-Aqsa's dome is lead.
    • Haifa: the 1933 port (main and lee breakwaters, quay sheds) and the German Colony: the Templer houses of 1868 in stone with red roofs, in two rows along its north-south main street.
    • Tel Aviv: its port (1936/1938) at its sourced coordinate.
    • Ramla: the White Tower east of the town's heart.
  • No anachronisms:
    • The Shrine of the Báb's dome (1953) is not drawn.
    • Neither is the Dagon silo (1955), the gold Dome of the Rock, or any post-1948 tower.
  • Performance: one merged mesh per town as before, so the draw calls do not change. The lot index is promo-villages' DiscIndex, ported byte for byte with its KitGround hunks so the two branches merge cleanly there. It also takes the trees' lot test off its linear scan.
  • Tests (TownCensusTests):
    • The towns are deterministic.
    • Each city's landmarks are present.
    • No building stands on a route or in the sea.
    • Each of the four cities has a floor on its built cover.

Result (after merging the villages, #291)

The census as the map now builds it: with the 1948 localities, whose landmarks claim their hearts before any town's houses (review M8). "Main" is a3202ad, with the villages and before towns round 2; same census, same corridors.

town buildings, main → round 2 cover all / core, main → round 2 median wall height landmarks, round 2
Tel Aviv 36 (+12 wings) → 43 (+11) 18.4 / 19.7% → 17.4 / 18.5% 396 → 792 m (2 → 4 storeys) port, 3 hangar sheds
Jaffa 15 → 33 29.0 / 21.8% → 32.6 / 31.9% 660 m Mahmoudiya Mosque, clock tower, St Peter's, lighthouse on the hill, breakwater
Haifa 30 (+10) → 45 (+5) 16.4 / 19.4% → 21.0 / 20.6% 594 → 660 m port (main and lee breakwaters, 4 quay sheds), German Colony (Templer houses in two rows)
Jerusalem 70 → 110 25.7 / 19.8% → 22.7 / 24.9% 660 → 330 m walls (a gap where a road crosses), 13 gates, citadel, Haram, Dome of the Rock, Al-Aqsa
Lod / Ramla 5 / 5 → 11 / 11 mosque; Ramla also the White Tower
Petah Tikva / Rehovot / Rishon 2 / 3 / 2 → 8 / 5 / 3 water tower
all 25 216 → 365 (413 after the review: the small-town floor, #317) 27 → 55

Distinct wall colours: Haifa 7 → 16 (white blocks, the lower town's stone, the Templer houses), Jerusalem 8 → 14 (stone blocks, village houses, red-roofed villas), Jaffa 6 → 7, Tel Aviv 8 → 7.

  • Tel Aviv's cover did not rise. Before the merge, round 2 had 62 blocks and 30.0% / 25.6%.

    • The villages' landmarks inside its 7.6 km disc (Sheikh Muwannis, Summeil, Jammasin, Ramat Gan, Bnei Brak and others) now claim their spots first. The lattice loses every plot their mosque or water-tower lots reach.
    • What changed there is the form: narrower (12-16 m real), taller blocks in rows on the Geddes grid, with a 700 m beach strip in front. The close-up crops show the difference better than the cover does.
    • Taking the villages' spots back would break the villages' rule that a village keeps its heart, so it is left for the owner.
  • Jerusalem's Old City is symbolic and squeezed by the roads.

    • It is 4 km across, about 4 times the real one. The drawn 1948 roads (300 m ribbons) meet by Jaffa Gate and loop round the north and east, so at any readable size they cross inside the square, through gaps in the walls. Since #319 the map cuts each road and the railway at the walls (docs/routes.md, "Jerusalem's Old City"), so none runs inside them.
    • Round 2 fitted the Haram's platform where it kept clear of every route's traced line: 1.3 x 1.1 km, wider than long, with Al-Aqsa in its south-east corner, and it drew a gate wherever a road met the walls (13). Since #319 the platform runs north-south (1.1 x 1.7 km, the real one's shape) against the east and south walls, Al-Aqsa stands across its south end, and the walls have the seven gates open before 1948 on their real sides, the Golden Gate left sealed (sources in SettlementKits.Jerusalem). The roads' traced lines inside the walls still keep houses off them.
    • The citadel is moved east off the drawn Hebron road.
    • The houses inside the walls are capped at two storeys, and the Dome's walls and drum are 30x high, so the walls and the Dome stand above the roofs.
  • Readability and colour:

    • The grid stays readable: 78% of edge midpoints (69 of 89), against the 75% floor and main's 79%.
    • The beach beside Tel Aviv reads as sand at 26 of 33 samples (79%, floor 75%).
    • At gameplay03 (WebGL), the palette distance to promo 03 is 0.43 (main 0.44), warmth 0.068 (0.070) and saturation 0.45 (0.45).
  • Trees: Vegetation.Plant now draws each tree's random numbers before its lot and route test. A town's new lots remove only the trees they touch, instead of reshuffling every later grove.

    • Two grade tests sample the selection ring against the land and could land on a crown. They now measure with the trees off: the ring against the ground's graded colour, which is what they test.
  • Frame time (headless WebGL ?vprofile, machine_lock.py --exclusive --wait-unity, main a3202ad against 49bc88e, two interleaved rounds each, 2026-10-12, load 3.6 at the start and 7.3 at the end):

    gameplay03 all (p95) gameplay03 all − none diorama02 all (p95) diorama02 all − none
    main 13.70 / 14.22 ms (15 / 16) 3.00 / 3.45 ms 18.07 / 18.71 ms (20 / 20) 4.84 / 5.60 ms
    round 2 14.15 / 13.97 ms (16 / 15) 3.59 / 3.15 ms 18.06 / 18.70 ms (21 / 21) 5.54 / 5.53 ms
    • The kits stay one merged mesh per town: 63 draws, as on main. The triangles grow from 61,054 to 65,684.
    • "-kits" draws the marker boxes instead, so the kits' own cost is within the noise (all − (-kits) is −0.6 / −1.1 ms at gameplay03).
    • Every budget holds: all − none is under 8 ms, and the p95 at gameplay03 is under 16.7 ms. The diorama02 p95 of 20-21 ms is main's too (the villages section).
  • Boot (headed.py boot --runs 3, three interleaved rounds per side, so 9 runs each; main 2b7212e against this branch's final head; machine_lock.py --exclusive --wait-unity, 2026-10-12).

    • Other sessions loaded the machine (load average 6.5 rising to 25), so single runs reach 3.3 s at cpu 4. The medians and the "kits built" phase are the fair comparison.
    map_build median kits built phase median first_frame median
    cpu 4, main 2739 ms 570 ms 9494 ms
    cpu 4, round 2 2679 ms 563 ms 8750 ms
    cpu 1, main 659 ms 132 ms 4029 ms
    cpu 1, round 2 662 ms 138 ms 4020 ms
    • The towns add about nothing to the boot: the kit phase is 7 ms lighter at cpu 4 and 6 ms heavier at cpu 1.
    • The lattice checks are cheap first, and only towns with water near them run the beach test.
    • The earlier session on a quieter machine (main a3202ad, before those cuts) measured map_build 2668 → 2683 ms at cpu 4 (6 runs each), with the cpu 1 kit phase 129 → 137 ms.
    • The budget at cpu 4 is 2720 ms; #195 needs about 100 ms of it. Towns round 2 takes none.

Review (2026-10-12, on the merge with main 471c300)

  • Fixed on the branch:

    • The beach setback tested any water, so it also applied on the Sea of Galilee. It now tests the sea only (Sea: water that is not a lake hex). Degania goes from 0 houses back to main's 1.
    • Nahalal's ring ended at 0.88 of the radius, but its lattice's first ring of plots stands at 0.76 and 0.89, so it drew no houses. The ring now runs out to the radius: 0 → 1 (main 1).
    • Tel Aviv's blocks are capped at eight storeys, promo 03's tallest: the cap was ten. The tallest block is now 7 storeys (1,386 m) and the median is still 4.
    • TownCensusTests.SmallTownsKeepTheirHousesAndNoBlockOutgrowsThePromo covers all three.
  • Each test fails under a mutation:

    • Reverting the lake test, the ring and the cap fails the new test on all three counts.
    • A per-build counter in the plot yaw fails EveryBuildPlacesTheSameTowns.
    • Dropping the plots' route test fails NoBuildingStandsOnARouteOrInTheSea (268 placements).
    • Thinning the core to 0.4 fails TheCitiesCoverTheirGround.
  • Small-town floor (#317, fixed on the branch; reworked after the review of #339):

    • At 100x a plot is about 1 km, so a small town (1.3-2 km footprint radius) has about 7 plots, and the routes, the sea or a neighbour's lots take most of them. On the first lattice Tiberias drew 0 houses (main 3), Netanya 1 (main 2) and Beersheba 3 (main 5).
    • The lake was half the cause. The kits' KitGround.Water counted every lake hex as water, and a lake hex also holds its shore. Tiberias stands in one, 60-160 m above the Sea of Galilee (about -214 m), so it had no ground left.
      • A point is now lake water only at or next to the lake's surface, in any hex that is not sea (a shore hex of another biome holds some of the lake too). It uses the rule the drawn water uses (MapBuilder.LakeBelowM and LakeFlatM, tools/terrain_bake.py): below -150 m and flat within 0.25 m over the heightmap's 3 x 3 window, at the point or within one heightmap sample of it.
      • The drawn shore (coast.png's distance, interpolated between texel centres) runs up to about a sample past the last flat sample. A first version without that sample's margin put a cypress in the drawn Kinneret, and one that looked only in lake hexes put a palm and a citrus tree in it by Ghuweir Abu Shusha.
      • The test runs only in a lake's hexes and their neighbours, and each sample's flatness is cached: the rift's floor is asked about thousands of times while the kits are built.
      • Sea hexes keep their test. The Hula is not water in the heightmap's rule (its surface is about 70 m above sea level), on main or here.
      • KitTests.TheLakeIsItsWaterNotItsWholeHex checks that Tiberias's shore at 35.530 E, 32.7944 N (lake hex, -111 m) is land and that the lake at 35.58 E, 32.80 N (-214.5 m) is water. Each half fails when the lake branch is made always water, or never water.
      • KitTests.NothingStandsInALakesDrawnWater decodes coast.png itself, the file the water shader samples. It checks that no lot centre and no tree trunk in the rift (a lake hex, or ground below -150 m; 129 of them) lies under the drawn water. It fails on the cypress without the sample's margin, and on the palm and the citrus tree with lake hexes only.
      • Tiberias's lattice now draws 4 houses on its shore, and the fallback a fifth, full-size, beside them.
    • The fallback. A town whose lattice places fewer than 6 houses (round 1's floor, SmallTownFloor) is filled up the way round 1 drew it.
      • Each house is the archetype's full-size building, the size main drew, and keeps its whole footprint as its lot.
      • It tries 240 spots (ScatterTries) from its own stream (KitRandom(id, 3)), nearest the town's sourced point first. Every spot is tried at full size before any at 0.72; the half-size fit is never used. Every house keeps the beach setback.
      • Build runs the fallback after every place's own houses (towns, the 1948 localities) and the camps, so it takes only ground no other place uses. KitTests.TheSmallTownFloorMovesNoOtherPlace checks this: every lot of every place is the same, in the same order, with the fallback off and on.
      • It fires in 15 towns (the moshavot and kibbutzim, Tiberias, Netanya, Beersheba and Acre) and adds 27 houses in 13 of them. Acre and Yad Mordechai have no free ground left.
    • Every town now draws at least main's count: Tiberias 5, Netanya 4, Beersheba 5, Rosh Pinna 2, Nahalal 3. The 25 towns draw 397.
    • Tests:
      • SmallTownsKeepTheirHousesAndNoBlockOutgrowsThePromo pins every town at main's count (MainBuildings).
      • TownCensusTests.NoTwoHousesOfATownOverlap checks the footprints by separating axes. It fails (20 pairs) when the fallback skips its own lattice's lots.
    • The grid stays readable at 80% (71 of 89).
    • Perf, after the review's fixes (WebGL, main 4abb64f against the branch, machine_lock.py --exclusive --wait-unity, two interleaved rounds, load 5-9, 2026-10-13):
      • Frame time is equal. At gameplay03 all is 8.08 / 7.73 ms on main against 8.14 / 8.02 ms on the branch, and at diorama02 19.43 / 19.50 against 18.90 / 19.42 ms. The p95s are equal, and there are 63 draws on both.
      • Boot at cpu 4: the "kits built" phase is about 45 ms heavier, 691 / 709 ms on main against 743 / 750 ms.
      • An attribution round (main, the branch, the branch with the fallback off, and the branch with the old whole-hex lake rule) put about 10 ms of that on the fallback: 689 / 738 / 729 ms in the quieter round. The lake rule costs nothing measurable. The rest is round 2's own placement against this main.
      • The author's boot against 2b7212e measured the kit phase within noise. The ~35 ms not explained here is not attributed further.
    • Perf, one interleaved round with the floor (95710e7 against main 471c300, exclusive lock, the machine loaded by other sessions to load 15, so every number is high):
      • Frame time at gameplay03 is 25.62 ms (main) against 25.24 ms (branch), and at diorama02 27.78 against 27.67 ms.
      • Boot map_build median is 2982 against 3075 ms. The "kits built" phase is 746 against 686 ms.
      • The editor's kit build_ms is 85-89 ms, with or without the floor.
      • The floor adds no measurable cost; the boot difference is load. Still 63 draws.
    • The label refresh is unchanged: MapLabelTests.RefreshWhileTheCameraMovesIsCheapAndAllocationFree measured 162-163 µs with the floor and 171-180 µs without it, interleaved. The single 238 µs (fail) came in a full run at load 24.
  • Still open, filed as issues:

    • #318: Tel Aviv's cover is held down by the localities' landmark lots.
    • #319: the symbolic Old City has roads crossing inside its walls.
  • Unchanged after the fixes: the grid stays readable at 78% (69 of 89), and the beach beside Tel Aviv reads as sand at 79% (26 of 33).

  • Perf, re-measured by the review: frame time is headless WebGL ?vprofile; boot is headed.py boot --runs 3 --cpu 4. Main 471c300 against the branch at 4a68d05, each round under machine_lock.py --exclusive --wait-unity, in interleaved order. A third round was cut short by a stop.

    gameplay03 all (p95) gameplay03 none diorama02 all (p95) diorama02 none boot cpu 4: map_build / first_frame
    main, round 1 (load 3.1) 14.21 ms (16) 10.72 ms 19.47 ms (23) 13.20 ms 2584 / 8431 ms
    branch, round 1 (load 7.5) 15.40 ms (17) 11.32 ms 18.18 ms (21) 13.06 ms 2598 / 8369 ms
    branch, round 2 (load 13.6) 14.36 ms (16) 11.01 ms 17.80 ms (20) 13.00 ms 2619 / 8367 ms
    main, round 2 (load 4.2) 14.23 ms (16) 11.03 ms 17.81 ms (20) 13.02 ms 2595 / 8390 ms
    • Round 2 is equal within noise. In round 1, the branch's gameplay03 cost rises with its none row (+0.6 ms both), that is, with the machine's load.
    • Boot's map_build is 14-24 ms over main, within the runs' spread, under the 2720 ms budget at cpu 4.

Sources (checked 2026-10-12)

Settlement quick wins (branch art/settlements-quickwins, refs #31)

The 2026-10-13 gap audit (.superpowers/audit-2026-10-13/renders: measure.py, metrics.json) ranked the settlements the biggest gap to promo 03. Measured with its crops and poses: editor renders (MapCapture.Promo, Mobile tier, with -extraPose close30:34.83:32.00:30000:10) and the WebGL ?shot of gameplay03. Main is 7aeec3f.

Changes.

  • White City (Style.WhiteBlock):
    • Walls are lime-white plaster (0.94, 0.945, 0.94; was 0.94, 0.92, 0.87), with 8% of blocks cream (was 12%).
    • Roofs are paler (0.90, 0.89, 0.86).
    • Each upper storey has a balcony with a solid plastered parapet on the street side, and on most blocks the back too (WhiteBlockDetail). Each balcony stays inside the plot's lot.
    • Roof rooms are on 80% of blocks (was 50%), and a second room is on 45% of the blocks that do not step.
    • Kit.shader, White City walls only: discrete windows with plaster piers between them, the glass near black (_CityWindowDark 0.95; the other styles keep _WindowDark 0.55), and a darker band under the balcony slab. The storey pattern holds until a storey is about two pixels (was three), so the 30 km look keeps its window rows.
  • Old towns, Arab towns and villages (Style.OldHouse, Style.Village, Archetype.OldTown, Archetype.ArabTown; StoneHouse):
    • Three in four houses of two storeys or more step back: the top storey is a smaller block on one corner of the lower roof. Single-storey houses often have a roof room.
    • Walls are grey-buff limestone (0.77, 0.70, 0.60, jittered from grey to buff) and roofs lime plaster.
    • A village below 120 m (the coastal plain and the valleys) is mud brick (0.78, 0.69, 0.56) with flat roofs.
    • Domes are fewer and lower: a cap rising 0.6 of its radius, not a hemisphere, in the roof's lime colour, not white.
      • Old towns: 8% of flat-roofed houses (was 22%).
      • Hill villages: 7% (was 10% of every village).
      • Plain villages: none.
    • The mosques' and landmarks' domes are unchanged.
  • Historical basis. The owner's ruling (m3-rulings, "Owner rulings"): where the promo departs from the era, the promo wins, but the architecture stays plausible for 1948. Domes are reduced, not removed.
    • The hill country built in limestone, and its houses were roofed with stone vaults and cupolas.
    • The coastal plain and the valleys built in mud brick, with flat roofs of beams and earth.
    • So domes stay on hill houses only, as the low cap a vault shows from outside.
    • Sources are under "Sources" below.
  • Layout unchanged. A 1948 locality's houses share the locality's random stream with its spots. Emit(Archetype.ArabTown) still draws exactly five numbers from it (wall, roof, the dome roll), and each house's detail comes from a stream seeded off it. The lattice towns' houses already had a stream of their own.
    • Every building stands where it did: same lots, same trees, buildings 1954 (WebGL) / 1953 (editor), locality buildings 1339, trees 8358, draws 63.
    • So #318 (the localities' landmark lots in Tel Aviv) and #249 (tree holes under the figures) are untouched.

Result (main → branch; promo in brackets):

main branch promo 03 target
gameplay03 white share (V > 0.72, S < 0.28; WebGL shot, map area) 2.49% 6.55% 7.46% (audit) >= 4.5%: met
gameplay03 white share (editor, Mobile) 2.36% 6.84%
close30 Tel Aviv edge density (crop 0.32-0.58 x 0-0.30, 600x400) 0.038 0.054 0.114 >= 0.07: not met
close30 Tel Aviv p5 / p95 0.241 / 0.753 0.221 / 0.787 0.136 / 0.896 p5 <= 0.18: not met
close30 Tel Aviv warmth 0.131 0.114 0.090 <= 0.09: not met
close30 Jaffa edge / warmth 0.041 / 0.141 0.047 / 0.112 0.100 / 0.107
close30 village crop edge / warmth 0.035 / 0.222 0.036 / 0.198 0.071 / 0.132
  • Why Tel Aviv's edge, p5 and warmth fall short.

    • At close30 a storey is about three pixels. The crop's top third lies in the far tilt-shift band: the edge per band rose from 0.031-0.046 to 0.039-0.058, against the promo's 0.05-0.16.
    • About half the crop is the town's sand ground. Its tint comes from PaintGround, and SeaTests.TheBeachBesideTelAvivStaysSandWithTheKitsOn pins it.
    • The walls are already near-neutral white; the remaining yellow is the sun and the grade.
    • Closing the gap needs denser, smaller blocks (a layout change that meets #318), palms between the buildings (the vegetation) and a ground or light change. None is a quick win, so they are left as follow-ups.
  • Kit numbers (WebGL KITS line, main → branch): buildings 1954 → 1954, vertices 85,401 → 99,131, triangles 66,436 → 68,366, draws 63 → 63.

  • Boot (tools/headed.py boot --runs 3 --cpu 4, three rounds, order rotated, each round under one exclusive machine_lock, 2026-10-14; builds of main 7aeec3f and this branch; 9 cold loads each):

    map_build median (range) kits built phase median first_frame median
    main 3067 ms (3042-3085) 747 ms 11075 ms
    branch 3055 ms (3036-3077) 743 ms 11081 ms
    • Within noise: the per-round medians alternate (+19, -26, -16 ms), and the kit phase is unchanged.
    • The machine was loaded (other sessions' jobs), so both sides sit over the 2720 ms budget. Main does too; that is #321's margin, not this change.
  • Goldens: the WebGL golden&smoke=5 run shows GOLDEN OK 1950 71e9904d, GOLDEN OK 1952 62f8ec61 and SMOKE OK 5 turns 1948-06-19 d3fdd459, unchanged.

  • Tests: EditMode passes 768/768 with 0 warning CS, on the branch and on main 7aeec3f alike. An earlier run on the old base lost 52 UI, label and border coroutine tests to destroyed scene objects. It did not reproduce: all 768 pass on both.

Sources (checked 2026-10-13)

Arab villages: dense, terraced cores (visual gap S1, branch visual/arab-villages-s1, refs #32)

The 2026-10-15 gap audit ranked this 9 of 10: promo 03's Qalqilya, Lydda and Ramla are dense packs of layered stone houses, while main drew each 1948 Arab locality as a handful of large cubes with a domed mosque at its heart, so the hills read as evenly spaced dice, nearly every one with a dome.

Cause on main c02d4eb0.

  • The 250 Arab localities (type village or arab_town, mixed Nesher aside) took the promo-villages scatter (BuildOne). They are 234 Muslim (216 villages and 18 Arab towns), 11 Christian (a church) and 5 Druze (no landmark). Each wanted six buildings (TargetCount's floor: at 100x the built-up share asks for fewer), each 6.5-12 m real, so 650-1,200 m wide, on random spots. Counted from their lots: 1,063 houses, median 5, and 17 with none.
  • Every Muslim locality built the domed town mosque at 0.8 size (a 672 m hall, a 264 m dome and a 1.25 km round minaret).

Changes (SettlementKits.cs, SettlementKits.Landmarks.cs):

  • Layout. An Arab locality is laid out on the village plot lattice (PlaceHouses, towns round 2), as Lod and Ramla are, nearest its heart first. As before, it builds only in its own cell, off the routes, the grid's edge midpoints (review M10) and every other place's lots. It keeps no beach setback: the coast's villages met the shore, as their scatter drew them. Where the routes, a bigger neighbour or the cell's seam leave fewer than ten houses (VillageFloor), small village houses fill free spots over its plot reach, nearest the heart first.

  • The village lattice (Style.Village, so also Lod, Ramla, Tiberias, Safed and Jerusalem's Arab quarters):

    • The pitch is 6.6 x 6.2 m (was 7.6 x 7.2) with less jitter. Houses are 3.8-5.4 x 3.6-5.0 m (was 4.4-6.6 x 4.2-6.0), so the lanes run about two metres real.
    • It is built solid to 0.75 of the radius (was 0.5), and 0.85 at the radius (was 0.6).
    • More two-storey houses in the core: 60% falling to 15% (was 45% to 10%). Single-storey houses keep their roof rooms, and the roofs stay flat lime plaster.
    • House domes on hill houses fall from 7% to 3%. A village now draws about twice the houses, so it keeps about the domes it had.
  • Terraces. A hill house (above 120 m) whose ground falls more than 0.8 m across its footprint, beyond its 2 m plinth, stands on a terrace. The terrace is a dry-stone retaining wall up to the house's uphill ground, with a yard 0.3 m wide round the walls. The house's lot covers the terrace.

  • The mosque rule. Every Muslim locality keeps its mosque: all 271 localities are places of 1,000 or more people, none a hamlet. A village's mosque is now a village mosque (VillageMosque):

    • a one-storey stone hall, 12 x 9 m at the landmark scale (0.8 x 60x: 576 x 432 m);
    • a flat lime roof and no dome;
    • a short, thick square minaret at one corner, about 14 m tall (672 m), inside the hall's footprint.

    The domed mosque with its tall round minaret stays with the 18 Muslim Arab towns (Hebron, Nablus, Majdal, Qalqilya and the others), and with the game towns, which keep their own landmarks. So the domed mosques at the localities go from 234 to 18. House domes stay: 3% of hill houses, about main's count per village (see above).

  • The game towns keep their ground. A village keeps its houses out of every smaller game settlement's plot reach, plus the lot of that settlement's largest fallback house. That settlement places its houses, and its small-town fallback (#317), after the village. Without this rule, Hiribya's packed lattice left Yad Mordechai 1 house against its measured floor of 3.

Historical basis (sources below):

  • Village houses were single rooms, ranging "from three meters in small houses" to more than ten in big ones, clustered round yards "in a compact and convergent manner" (Abuarkub 2018). The kit takes the small end (3.8-5.4 m), so a village at 100x holds more houses. A caution: IJCIET's publisher, IAEME, is widely listed as a predatory publisher. The quoted phrases are in the paper and only a visual parameter rests on them, but a stronger reference (such as Amiry and Tamari, The Palestinian Village Home, 1989) has not been checked yet.
  • Hill villages such as Battir stood among stone field terraces round the village. The terrace under a sloped house is a modelling choice that carries that stepped look into the village itself, not a sourced house form.
  • A village mosque was a one-storey hall with a chunky minaret at one end (Qaysaria), and a village's square minaret stood 10-15 m (Kawfakha, from Petersen's gazetteer).
  • No source here supports leaving out the mosques of the villages of 1,000 or more people. So the rule makes them smaller and domeless; it does not drop them.

Result (EditMode census: VillageKitTests, TownCensusTests; main from the localities' lots on the same build):

main c02d4eb0 branch
Arab localities' houses (250, churches and Druze villages included) 1,063, median 5 2,617, median 11 (VILLAGES: max 50)
localities with no house 17 9, all inside Jerusalem's or Tel Aviv's footprint (Silwan, At-Tur, Lifta, Abu Dis, Sheikh Muwannis...)
hill villages (150): median houses, core cover by houses not recorded (main records no locality house) 12, 17.5%
domed mosques at the localities / village mosques 234 / 0 18 / 216
Lod / Ramla houses (census) 11 / 11 16 / 16
WebGL KITS line: buildings, vertices, triangles, draws 1,972, 103,537, 70,572, 64 3,632, 167,329, 89,178, 64
grid readable at gameplay03 (floor 75%) 79% (70 of 89) 83% (74 of 89)
  • The villages under six houses outside the game towns are Yibna, Taibe, Beit Dajan, al-Bireh (beside Ramallah), al-Jura, az-Zeeb, Wadi Hunayn, Tantura, Nazla and al-Jiyya (10 of about 200). The routes, the coast and bigger neighbours take their ground. The test holds the count at 5% or fewer.
  • The core cover counts houses only (review of #647). Counted with the mosque, main's domed mosque alone covers 35.8% of the median hill core (49 of 137 samples), and the branch's houses and smaller mosque together also come to 35.8%. That match is a coincidence of the two medians, not a shared cause, and it is why the first version of the test passed on main. The test's floor is 15%, against the measured 17.5%. Main records no locality house, so on main the test reads 0%. That proves the test sees the houses; it does not show that main's large cubes covered less of the core.
  • Each house stays at the house scale (100x): no village house is over 5.4 m real. The census and town tests pass, every town is at or over its floor, and Yad Mordechai is back to 3.
  • The goldens are unchanged (GOLDEN OK 1950 71e9904d, 1952 62f8ec61).
  • Against promo 03 (p2-villages-s1-compare/):
    • The villages now read as packs of small houses with lanes, not as dice, and the domed mosques are gone from the villages. House domes remain on 3% of hill houses.
    • The promo's Qalqilya still has more and taller houses: 30-40 houses of two or three storeys on terraced rings. At 100x a village's 1.3-1.4 km footprint holds about 19 plots, so 30+ houses would need a smaller house scale or a larger village footprint (an owner call).
    • The promo's warmer honey stone and its shading are the palette and Kit.shader, out of scope here.

Frame time (headless WebGL ?vprofile, 1920x1080 at DPR 1, each slot under one machine_lock.py --exclusive --wait-unity, two slots in opposite orders, load 2-4, 2026-10-15; builds of main c02d4eb0 and this branch before its rebase onto #557, which adds the ma'abara kit and no village):

gameplay03 all (p95) gameplay03 all - none diorama02 all (p95) diorama02 all - none
main, slot 1 / 2 8.21 / 8.21 ms (10 / 10) 0.15 / 1.26 ms 15.92 / 16.29 ms (17 / 17) 4.98 / 5.28 ms
branch, slot 1 / 2 8.23 / 8.23 ms (10 / 10) 1.25 / 1.24 ms 16.10 / 15.94 ms (17 / 17) 4.98 / 5.18 ms
  • Frame time is equal within the runs' spread. Main's 0.15 ms in slot 1 is its noisy none row (spread 1.1 ms).
  • The localities stay merged per 30 km tile, so the draws stay at 64. The triangles grow from 70,572 to 89,178.
  • The kits' own cost, all - (-kits), is within noise on both: -0.32 / -0.09 ms on main and -0.05 / -0.11 ms on the branch at gameplay03.
  • Boot (headed.py boot --runs 3 --cpu 4, in slot 2): map_build median 2147 ms on main against 2293 ms on the branch (+146 ms), and first_frame 7780 against 7906 ms. The map build stays 427 ms under the 2720 ms budget.
    • The cost is the kit build: the player's KITS build_ms is 45 ms on main and 77 ms on the branch at cpu 1.
    • Most of it is the village lattice's plots and their fits, and the floor's spots for the villages under ten houses.
    • One round only. A tighter kit phase (fewer floor spots) is the lever if #321 needs the headroom.

Sources (checked 2026-10-15)

Tel Aviv's White City (visual gap S2, branch visual/telaviv-bauhaus-s2, refs #31)

The 2026-10-15 gap audit (.superpowers/visual-gaps-2026-10-15.md, S2) found promo 03's Tel Aviv to be white Bauhaus blocks of four to eight storeys, with balcony bands, ribbon windows, roof terraces and palms in the streets, about 45-60 px wide in a 1920 frame. Main drew white boxes of two to seven storeys, about 25-30 px wide, with no street trees.

Owner rulings on #652 (2026-10-15):

  • "Lower floor to ~30, wider blocks": blocks about 1.6-2 times main's width, more height variety, and open plots for palms. The first round kept the 50-building floor; its widest block that held it was 15.8 m, 1.2 times main's.
  • "Keep strips (promo look)": the ribbon windows stay.

Changes (SettlementKits.TelAviv.cs; Tel Aviv only, on its own lattice style Style.TelAviv; Haifa's Hadar keeps Style.WhiteBlock):

  • Blocks:
    • Size: 21-24 x 13-15 m real (main 12-16 x 9-11.5), on a 27 x 19 m pitch (main 18.5 x 13.5).
    • Plots: every plot is built in the core, and nine in ten at the edge (main seven in ten).
  • Storeys, in three tiers as promo 03's skyline (promo over era, owner ruling 2026-10-10):
    • tall blocks of seven or eight storeys, 60% of the core's and 35% at the edge;
    • low blocks of four, 36%;
    • the rest five or six.
  • Facade: a dark ribbon window runs along each long side of every storey, above the balcony bands that WhiteBlockDetail already drew.
    • It is geometry: two triangles per storey and side, set a hair proud of the wall.
    • It is not done in Kit.shader, which is unchanged. At gameplay03 a 2 m window bay is under a pixel, so the shader's discrete windows fade to their mean, while a strip stays a band.
  • Pilotis under a quarter of the blocks: a recessed, shaded ground storey under the block (the White City's first was Engel House, 1933).
  • Roof terraces: a plastered parapet round every flat roof, warmer paving inside it, the stair head and, on four in ten, the laundry room.
  • Street palms in Tel Aviv's own mesh:
    • Where: street crossings, street middles, and a 3 x 3 grove over each plot left open.
    • Size: each is as large as its spot allows, 0.7-1.6 times the city blocks' scale, with 85% of its crown clear of every block.
    • Order: they are placed after every place's houses and the camps, and before the small towns' fallback, so they move nothing.
    • Each claims a lot, so no tree grows through one.
  • #507: everything is in tel_aviv_kit, so a model that turns the town's kit off hides all of it.
  • Plaster (#658, follow-up): Tel Aviv's walls and parapets are a warm cream, TelAvivPlaster (0.97, 0.90, 0.87), and the terrace paving (0.87, 0.80, 0.75), where they were the shared lime white (0.94, 0.945, 0.94) and (0.84, 0.81, 0.76). Haifa's Hadar keeps WhitePlaster. Census mean wall: Tel Aviv (0.93, 0.93, 0.92) to (0.96, 0.90, 0.86); Jaffa (0.76, 0.69, 0.60) and Haifa unchanged. Editor render, lit plaster (85th-percentile luma in Tel Aviv's box), gameplay03: (0.886, 0.824, 0.714) to (0.902, 0.820, 0.706) at blue 0.84, the same median at 0.87; promo 03's Tel Aviv reads (0.969, 0.855, 0.796). The review measured a first cream, blue 0.84, as too yellow against the promo's balance (red minus blue past it), so blue went to 0.87. Editor, the brightest 15% of the pixels that changed in Tel Aviv's box: red minus green 0.053 on main, 0.080 at both blues; red minus blue 0.156 on main, 0.198 at 0.84, 0.191 at 0.87 (promo 0.114 and 0.173).

Result (editor census, TownCensusTests; main 98515c52 → branch):

Tel Aviv main branch
buildings (+wings) 55 (+18) 27 (+4)
cover, whole / core 25.5% / 25.6% 27.2% / 23.7%
storeys min / median / max 2 / 5 / 7 4 / 5 / 8
height spread (cv) 0.30 0.31
median block width (real m) 13.1 21.9 (1.67x)
street palms 0 26
  • Floors: Tel Aviv's floors follow the ruling: 24 buildings (in Floors and MainBuildings), and cover 24.5% / 21%.

    • All are about 10% under the measure.
    • They were 50 buildings and 22.5% / 25% cover, with a MainBuildings row of 36.
    • Every other town's floor is unchanged.
  • The ruling said about 30 buildings. At 1.6 times main's width the lattice holds 27; the sweep's 29-32-building variants kept median widths of only 15.8-17 m. The owner then chose "27 wide blocks (Recommended)" (#652, 2026-10-15).

  • Neighbours:

    • Jaffa draws 39 buildings (main 38), with cover 39.0% (main 39.0%).
    • Petah Tikva draws 9 (main 9), with cover 20.7% (main 20.3%).
    • The 1948 localities draw 2,898 houses (main 2,895).
    • The other 22 towns are unchanged.
  • Kit numbers (editor KITS line): vertices 167,301 → 173,789, triangles 89,166 → 92,616, draws 64 → 64.

  • Frame time: not re-measured for this round. The vertices rise by 4%, less than the first round's 13%, which measured as below (WebGL ?vprofile, headless, 1920x1080, DPR 1; main c02d4eb0 and the first round back to back under one exclusive machine_lock, 2026-10-15; PROFILE summary median avg / p95 in ms):

    pose, toggles main c02d4eb0 first round
    gameplay03, all 8.33 / 10 8.33 / 10
    gameplay03, -kits 8.61 / 10 8.39 / 10
    diorama02, all 15.65 / 17 15.89 / 17 (spread 0.84)
    diorama02, -kits 15.70 / 17 16.05 / 17
    • The kits' cost (all minus -kits) is within the spread on both.
    • The machine was shared with other sessions' jobs between runs. Diorama02's p95 of 17 ms is over the 16.7 ms budget on main as well, so it is not this change.

Accuracy. The White City article (https://en.wikipedia.org/wiki/White_City,_Tel_Aviv, checked 2026-10-15) has:

  • pilotis from 1933 (Engel House);
  • long, narrow balconies;
  • flat roofs used as shared space;
  • white and light colours;
  • small parks between the buildings;
  • the Geddes plan, 1925-29.

It describes the windows as small and recessed, not ribbon windows. The strips follow promo 03 (owner ruling on #652, "Keep strips (promo look)"), as the storeys do.

Issue #195: the water and the border lookups out of the terrain's fragment shader

  • Why. promo-sea (#138) put the water shading (ScWaterShade, the breakers' fwidth, the footprint) in the terrain graph's function, and promo-borders (#206) six neighbour fetches in ScBorderApply's loop. Either one makes the shader big enough to slow every terrain pixel in WebGL, the water and borders toggled off included (.superpowers/handover-2026-10-11/promo/perf-investigation.md): none@gameplay03 measured 2.5 ms without both, 7.3 ms with the borders' fetches alone, about 10 ms with both.

  • Water. StatecraftTerrain.hlsl outputs a black, fully metallic, fully occluded surface with no emission on the water (Lit adds 0 there). MapBuilder.BuildWaterMesh copies the terrain's surface wherever the water's coverage can be above 0: wet heightmap samples (the sea within 0.5 m of 0 m, flat rift lakes) and two samples around them, at full resolution, with flat blocks as T-junction-free fans (129,882 vertices, 239,958 triangles). It is drawn with a copy of Sea.mat in its _ScTerrainWater mode: additive, colour times coverage (the land adds 0: no per-pixel discard to split the shore's 2x2 quads, the lesson of #306), no depth write, its depth pulled two pixels' worth towards the camera (the terrain's LOD splits the surface its own way). The terrain draws one queue earlier (Terrain.mat 1998) and the water at 1999, before every other opaque, so boats and harbours cover it.

  • Borders. MapBorders.Use bakes, per hex, which of its six edges border another region (alpha: 192 plus a bit per edge, MapBorders.NeighbourSteps order). ScBorderApply reads its own hex's texel only, where it read up to seven, and skips a hex with no border edge (it would draw nothing). Same math, same pixels: BorderTests unchanged, and the border pixels of the player's screenshots are identical. An overlay mesh along the bands, like the water's, was the plan; the bake met the target with no mesh, no boot or memory cost and no rebuild hitch, so no mesh was built.

  • Exactness. On the water and its shore the sum is what the terrain drew before, except where a border band meets the 140 m coverage ramp at the shore, and the extra rounding of one blend.

  • Measured (WebGL, headless, 1920x1080, DPR 1; PROFILE summary medians over 3 interleaved rounds, main 1cdb356 and the fix back to back under one exclusive machine lock per round, order rotated; [min-max] across rounds):

    pose, toggles main 1cdb356 fix
    gameplay03, none 10.48 [10.29-10.79] 2.52 [2.44-2.55]
    gameplay03, all 13.94 [13.85-14.14] 3.19 [3.16-3.19]
    diorama02, none 13.56 [12.40-13.78] 8.18 [8.02-8.33]
    diorama02, all 18.72 [17.92-19.03] 12.05 [11.93-12.35]
  • Baked. The mesh's samples and triangles are baked into Assets/Textures/Map/terrain_water.bytes (3.4 MB; Statecraft.Game.EditorTools.TerrainWaterBake.Bake, re-run whenever heightmap.r16z changes), referenced by Map.unity's MapScene, so a player only places the vertices on the heights it loaded. The file carries the heightmap's resolution and a check of its heights; a stale or missing bake is refused (a warning) and the mesh derived at load as before (22 ms unthrottled, about 87 ms at --cpu 4). SeaTests.TheBakedWaterMeshMatchesTheHeightmap fails on a stale bake.

  • Boot (tools/headed.py boot --runs 3 --cpu 4, interleaved under one exclusive lock per round, order rotated; six rounds, 18 runs each, 2026-10-13, main 244723f against the fix merged on it, cc3c954): map_build_ms median main 3023 [3006-3042] ms, the fix 3039 [3017-3064] ms. terrain water takes 19 [17-22] ms, the fix's whole share; terrain heights set is 1389 [1374-1411] ms on main and 1397 [1372-1409] ms on the fix.

  • SetHeights moves with the build, not with the fix. The review measured the fix merged on main 6d439a7 (44e1d63) at about 105 ms over that main, 85 ms of it in terrain heights set, and it reproduces here (1473 [1462-1496] ms against 1385 [1375-1418], nine runs each). It is not the 3.4 MB bake, and not the WebGL heap:

    • The same commit rebuilt without Map.unity's reference to the bake (the same wasm, byte for byte; the water mesh then derived at load) keeps the slow terrain heights set: 1487 [1466-1503] ms against 1491 [1463-1499] with it.
    • With heap probes (a WebAssembly.Memory.grow hook and the heap size after every boot phase) in builds of main 244723f and of the fix merged on it, both grow the wasm heap at the same points to the same sizes (286 MB to 343 MB at new TerrainData, none inside SetHeights); the bake adds 3 MB of reserved native memory and no grow. The SetHeights call alone measured 1185-1198 ms on both.
    • A small C# change that runs only after SetHeights (reading the bake through TextAsset.GetData, which cut terrain water from 20 to 17 ms) made terrain heights set 1511 [1489-1530] ms in its build, against the fix's
      1. It was dropped.
    • With V8's tiering off (--js-flags=--no-wasm-dynamic-tiering: every function compiled by TurboFan up front), five builds (both mains, the fix, 44e1d63 and the GetData variant) all take 1426-1455 ms in terrain heights set. So each wasm binary gets its own SetHeights time, about 1390 or about 1490 ms at --cpu 4, from when V8 tiers up the code it runs. Any change can move it either way, and the next merge of main draws again. So judge a boot change by its own phases, and read terrain heights set apart from them.
  • A layer change (MapBorders.Use) takes 0.8 ms for the 870-hex 1948 layer, texture upload included: no rebuild, no hitch.

  • Memory. The water mesh keeps about 4.5 MB on the GPU (its CPU copy is dropped in players, UploadMeshData); placing it from the bake allocates its 3.4 MB of bytes and about 1.5 MB of vertices, short-lived. The borders add nothing (the edge bits ride in the existing texel).

Issue #390: a lighter far tilt-shift band at the gameplay zoom

  • What happened. #320 did not change the tilt-shift. Before it, normal play on the development machine was slow enough that the runtime quality fallback stepped down to level 3 and turned the tilt-shift off (headed Chrome, 1920x1080, New Game then End Turn: QUALITY level=1 p95 51 ms, level=2 45, level=3 ... tiltshift=off 42). #320 made the frame about 3.7x faster, so the fallback stays at level 0 and the far band drew in full: Tel Aviv, Hadera and the brigade token were a smear. Main before #320 with ?nofallback draws the same frame as 4abb64f (0.109% of pixels more than 24 levels apart, identical gradient energy per tenth of the rows). ?shot and ?vprofile turn the fallback off, so the gameplay03 pixel A/B never showed it.

  • Owner's ruling (2026-10-13). Keep a soft tilt-shift, so the game still reads like promo 03, but keep the far towns readable at the play zoom; the wide views (diorama02) keep the full effect. This also settles the open bottom-band note above: the bottom band is not changed.

  • What changed. TiltShift.FarStrength(distance) scales the far (top) band's radius and its blend: 0.35 (FarGameplayStrength) up to 150 km, rising to 1 at the 250 km overview. The near band is unchanged. It is one global, _ScTiltShiftFar, set with _ScTiltShift in TiltShift.Apply: no new keyword or variant.

  • Readability bound. Promo 03's far band (rows 10-18% from the top, columns 30-80%) has 0.34 of the gradient energy of its focus band (rows 45-60%); a sharp frame of ours (?notiltshift, normal play) has 0.93 there, so the promo keeps about 37% of the far detail. TiltShiftTests.TheFarBandStaysReadableAtTheGameplayZoom asks for at least 60% (shifted / sharp, rows 0.78-0.90 from the bottom, under the web bar) at 60 km and at the 120 km starting zoom. The fix keeps 0.72 and 0.69; 4abb64f kept 0.17 and 0.14. TheDioramaKeepsBothBands asks for at most 0.30 far and 0.45 near at diorama02 (0.20 and 0.33, unchanged). Mutants killed: the far strength at 1 at the gameplay zoom (4abb64f's look), a shader that ignores _ScTiltShiftFar, a far strength stuck at 0.35 at the overview, and the far strength applied to both bands.

  • WebGL captures, 4abb64f against the fix. Gradient energy per tenth of the rows from the top, columns 30-80%:

    frame build 0-10% 10-20% 20-30% 30-40% 40-70% 70-100%
    normal play, 22 May 1948 (fallback on) 4abb64f 13.4 5.2 6.6 18.3 29.4 26.3 23.2 13.9 3.3 2.5
    fix 14.6 17.7 17.4 23.4 29.4 26.3 23.2 13.9 3.3 2.5
    ?notiltshift 15.3 24.7 23.1 25.8 29.4 26.3 23.4 21.6 20.5 16.7
    gameplay03 (?shot) 4abb64f 12.4 3.4 4.4 12.1 17.3 17.1 17.0 8.9 2.8 2.2
    fix 13.2 12.7 13.6 15.0 17.3 17.1 17.0 8.9 2.8 2.2

    diorama02 is pixel-identical (0 pixels more than 24 levels apart). The near band and the focus band do not move.

  • Frame time (WebGL headless ?vprofile, 1920x1080, 2 interleaved rounds under the exclusive lock, medians): gameplay03 none 5.02 / 5.06 ms against 4abb64f's 5.07 / 5.20, all 7.90 / 7.99 against 7.89 / 7.87; diorama02 none 12.63 / 12.62 against 12.20 / 12.21, all 19.51 / 19.49 against 19.08 / 18.94. Main before #320 measured 20.98 ms at gameplay03 none in the same session, so #320's gain stands. The 0.4 ms at diorama02 repeated in both rounds; there the far band's strength is 1 and the frame is pixel-identical, so it is the extra uniform read and compare per tilt-shift pixel or build-to-build noise, within the tilt-shift's 1.5 ms budget.

Issue #458: a lighter near tilt-shift band, by distance

  • Owner's ruling (2026-10-14). The near (bottom) band should match promo 03: less blur at the gameplay zoom and almost none at the overview. The far band stays as ruled for #390 (FarGameplayStrength 0.35). This replaces the 2026-10-13 note above that the bottom band is not changed.

  • The gap (the 2026-10-13 audit, .superpowers/audit-2026-10-13/renders/measure.py): mean |Laplacian| of the luma in rows 80-95% from the top, columns 30-75%, on the WebGL ?shot. Main had 0.0056 at gameplay03 against promo 03's 0.021, and 0.0064 at diorama02 against promo 02's 0.077: the near band drew in full at every distance.

  • What changed. TiltShift.NearStrength(distance) scales the near band's radius and its blend, as FarStrength does the far band's: NearGameplayStrength (0.75) up to 150 km (NearLightBelowM, the far band's breakpoint), falling to 0 at the 250 km overview. One global, _ScTiltShiftNear, set with the others in TiltShift.Apply; no new keyword or variant. The band's height (HalfWidth 0.13, Feather 0.17) is unchanged.

  • Why 0.75, not 0.35-0.5. The ruling's brief guessed 0.35-0.5. The half-resolution copy the blur composites is softer than the frame on its own, so the blend, not the radius, sets most of the measure. A model of the pass in numpy, checked against main (it gives 0.0050 for main's 0.0056), puts 0.4 at about 0.046 (twice promo 03's sharpness), 0.5 at 0.039, 0.7 at 0.024, 0.75 at 0.021 and 0.8 at 0.018. Scaling the radius alone cannot reach the target: a zero radius at full blend still measures only 0.023. 0.75 is the value that matches promo 03.

  • Measured (WebGL ?shot, main 88787ab against the branch, same machine):

    pose main branch promo target
    gameplay03 0.0056 0.0195 0.021 (promo 03) 0.015-0.025
    diorama02 0.0064 0.0575 0.077 (promo 02) at least 0.05

    diorama02's ?notiltshift shot measures 0.0575 too: the overview's foreground is now as sharp as the frame. Above the focus, main and the branch differ less than two ?shot runs of main do (gameplay03: 7346 pixels more than 2 levels apart against 14749 run to run, all on the animated sea), so the far band is unchanged.

  • Tests. TheNearBandIsGentleAtTheGameplayZoomAndGoneAtTheOverview covers the curve. At the gameplay pose the near band now keeps 30-80% of its gradient energy (TheBottomBlurs..., TheFarBandStaysReadable...; it kept about 20-25% before), and TheDioramaKeepsTheFarBandAndASharpForeground asks for at least 95% of the near band at diorama02 (main: 33%) with the far band still at most 30%. Measured: near 0.37 / 0.39 (Mobile / PC, no bar), 0.41 at 60 km and 0.35 at 120 km under the web bar, 1.00 at diorama02; far unchanged (0.71, 0.68, 0.20). Mutants killed: a shader that ignores _ScTiltShiftNear (near 0.19-0.33: the three render tests) and the near strength applied to both bands (far 0.34 at 60 km, 1.00 at diorama02: three render tests); the C# curve held at 1, or not reaching 0 at the overview, fails the unit test.

  • Quality fallback. Unchanged: TiltShift.Draws still needs !FallbackOff, so level 3 and every lower-quality step still turn both bands off (QualityFallbackTests pass). The pass draws as before whenever the tilt-shift is on, even where the near band is 0, so frame time does not depend on the new strength.

  • Fill rate. Headed ?vprofile at DPR 2 (headed.py profile --dpr 2), main, branch, branch, main under one machine_lock.py --exclusive --wait-unity, 2026-10-14. The machine was loaded by other sessions (load 6-13), so the spreads are 1.5-7 ms. diorama02 all: main 41.81 / 33.81 ms, branch 41.09 / 39.46; none: main 28.02 / 23.97, branch 26.36 / 24.83. gameplay03 all: main 11.95 / 13.99, branch 13.27 / 12.40. The branch sits inside main's range in every row: within noise, as expected for one more uniform in an unchanged pass.

Issue #38: grade retune (land against promo 03, sea kept blue)

  • What changed. MapGrade: contrast 14 -> 18, post exposure 0.15 -> 0.45, saturation 0 -> -7, temperature 5 -> 3, tint 2 -> -2, split-tone shadows teal (0.36, 0.52, 0.60) -> near-neutral (0.43, 0.46, 0.49), highlights amber (0.96, 0.78, 0.55) -> near-neutral (0.84, 0.87, 0.835), balance -15 -> -31. The tonemapper and the vignette (0.26 / 0.45) are unchanged. The look is crisper: more contrast, the low end pulled down by the balance and the mid-tones lifted by the exposure, while saturation -7 takes back what contrast 18 adds, so the land keeps main's warmth and saturation.

  • How it got here: four owner decisions on #519.

    1. The first fit (temperature 7, saturation -4, highlights (0.96, 0.78, 0.62), vignette 0.34 / 0.84) darkened gameplay03's sea. The owner's rule: wait for the sea PR (#536), re-measure, and "if the sea is still dark, use main's vignette". After #536 the open sea read luma 0.254 / p5 0.172 against promo 03's 0.351 / 0.295 and main's 0.340 / 0.290, more than 0.02 below both, so the vignette went back to main's.
    2. With main's vignette diorama02's whole-frame warmth was 0.169, over #38's 0.145: the owner chose "Re-fit the warmth" (the highlights' colour and the balance, temperature and tint in small steps, no exposure change that darkens the sea).
    3. No such re-fit met both 0.145 and #555's border-stripe test at exposure 0.55; the owner ruled "Exposure 0.45, sea slightly darker" (2026-10-15). That grade (temperature 3, saturation -2, pale rose highlights (0.89, 0.80, 0.88), slate shadows (0.45, 0.45, 0.53), tint 2) met the numbers but read pink, turned the table around the diorama maroon, and moved the Israeli border stripe 8.1 deltaE76 from the owner's ruled blue (0.6 on main).
    4. After side-by-sides the owner chose "Less pink, then merge" (2026-10-15): keep the exposure and contrast, take the rose out of the land and the table, and bring the stripe back to about 3 deltaE76 of the ruled blue or less.
  • The less-pink re-fit. Each candidate was checked by EditMode (GradeTests, and ControlTests' ThePrimaryStripeIsTheDeeperBlueOfTheRuling, which logs the stripe's deltaE76 to the ruled blue (89, 119, 156)), then built and shot in WebGL. The stripe was too blue (94, 119, 169 at the pink grade), and the slate shadows' blue was most of it: near-neutral shadows took it to 3.8. Green-highest highlights and tint -2 took the rose out of the land; they warm it, and saturation -7 (with temperature 3) brings diorama02's warmth back under 0.145. A rose index, (R - G) - (G - B) of the land's mean colour, measures the cast: above 0 reads pink or rose, below 0 olive or yellow.

    candidate (temperature, saturation, tint, shadows, highlights) d02 warmth (bar 0.145) stripe deltaE76 land rose tests
    the pink grade (3, -2, 2, 0.45/0.45/0.53, 0.89/0.80/0.88) 0.144 8.1 +8 pass
    3, -3, 0, 0.46/0.47/0.50, 0.88/0.85/0.82 0.157 3.8 -1 pass
    4, -3, -2, 0.45/0.46/0.49, 0.87/0.86/0.83 0.158 2.6 -4 pass
    3, -6, 0, 0.46/0.47/0.50, 0.85/0.87/0.84 0.149 2.9 -3 pass
    4, -7, -2, 0.45/0.46/0.49, 0.85/0.87/0.84 0.150 1.8 -3 pass
    3, -8, -1, 0.46/0.47/0.50, 0.86/0.86/0.85 2.6 saturation x0.980 (floor)
    3, -7, -2, 0.43/0.46/0.49, 0.84/0.87/0.835 (this grade) 0.141 2.1 -5 pass
  • What is measured. The targets are region-based (maintainer's decision on #38): the land against promo 03's land, the sea held, #388's sea floors, and diorama02's whole-frame warmth in [0.087, 0.145]. Shots: WebGL ?shot, 1920x1080 (tools/art_checkpoint.sh's web_check.py --query "golden&shot" --shots). Crops, measured with .superpowers/audit-2026-10-13/renders/m.py: land (0.36, 0.20, 0.95, 0.80) (promo (0.36, 0.30, 0.83, 0.80)), open sea (0.02, 0.15, 0.28, 0.45) (promo (0.05, 0.12, 0.25, 0.30), the 2026-10-13 audit's; promo 03's wider sea crop takes in its dark UI buttons), frame (0, 0.09, 0.83, 0.83) (promo (0, 0.07, 0.83, 0.80)), diorama02 board (0.31, 0.12, 0.71, 0.97) (promo 02 (0.02, 0.13, 1, 1)), diorama02 open sea (0.29, 0.12, 0.37, 0.26) (promo 02 (0.05, 0.15, 0.30, 0.55)), diorama02 Negev (0.35, 0.75, 0.55, 0.95), diorama02 table (0.01, 0.3, 0.12, 0.7). The whole-frame warmth and palette distance are tools/art_compare.py's; #388's detail and spread are #536's measures. Main and this grade were both measured on main 732ceab8.

    promo main this grade
    g03 land luma / p5 / contrast 0.425 / 0.145 / 0.187 0.481 / 0.181 / 0.170 0.501 / 0.179 / 0.187
    g03 land saturation / warmth 0.363 / 0.181 0.370 / 0.201 0.373 / 0.202
    g03 land mean colour, rose index (130, 104, 84), +6 (141, 121, 90), -11 (148, 125, 97), -5
    d02 Negev mean colour, rose index (190, 156, 128), +6 (204, 169, 106), -28 (212, 180, 126), -22
    d02 table mean colour, rose index (50, 47, 41), -3 (58, 44, 31), +1
    g03 open sea luma / p5 / warmth 0.351 / 0.295 / -0.248 0.335 / 0.285 / -0.219 0.331 / 0.276 / -0.199
    g03 frame p5 / contrast / warmth 0.148 / 0.184 / 0.064 0.235 / 0.157 / 0.040 0.230 / 0.176 / 0.046
    #388 detail / spread (crop; focus band) 0.060 / 0.201; 0.075 / 0.316 0.036 / 0.102; 0.037 / 0.104 0.038 / 0.113; 0.041 / 0.119
    whole-frame warmth g03 / d02 0.064 / 0.145 0.060 / 0.131 0.066 / 0.141
    palette distance g03 / d02 (#388 bars 0.44 / 0.56) 0.36 / 0.58 0.35 / 0.57
    d02 board luma / warmth 0.556 / 0.152 0.590 / 0.230 0.623 / 0.204
    d02 open sea luma / p5 0.437 / 0.371 0.410 / 0.399 0.418 / 0.403
    border stripe deltaE76 to the ruled blue 0.6 2.1
  • What is left. The land is brighter than the promo's (luma 0.501 against 0.425; main 0.481); its warmth and saturation stay at main's (0.202 / 0.373 against main's 0.201 / 0.370 and the promo's 0.181 / 0.363). Its hue sits between main's yellow and the promo's slight rose (rose index -5; main -11, promo +6). The table around the diorama is a little warmer brown than main's. The open sea is 0.020 under promo 03's luma and 0.019 under its p5 (main's own sea is 0.016 / 0.010 under). diorama02's palette distance (0.57) is over #388's 0.56, as main's (0.58) already is. Met: diorama02's whole-frame warmth (0.141), the land's contrast (0.187 against the promo's 0.187; main 0.170), #388's detail and spread floors, and the stripe within 3 deltaE76 of the ruled blue.

  • Why one global grade cannot do it all. Contrast works per channel in log space, so it raises every colour's saturation; taking the land well under main's saturation needs a large negative Saturation, which turns the sea grey-navy (a -25 trial, on dc4b272: sea warmth -0.12, TheOpenSeaStaysBlueUnderTheHaze b - r 0.149 against its 0.15). Warm split-tone shadows (#391) push the sea warm in the same way.

  • Tests. No test or threshold changed (a stale saturation note in GradeTests was updated). EditMode 1157 of 1159 on main 732ceab8 with this grade (2 explicit tests not run), 0 warning CS; the filtered Grade, Sea, SeaCraft, Border, BorderEdge, Map, MapEdge, MapLabel, BiomeColourBake, TerrainShader, Control, Settlement and Kit suites 222 of 222 on main 988279c3. The grade's own tests: saturation 0.411 -> 0.411 (x0.999), clipped 0.00% (Negev and gameplay03), vignette corner/centre 0.754 -> 0.576, selection deltaE 32.6 graded at gameplay03 (28.9 ungraded) and 38.6 at diorama02 (33.4), colour-blind sea-land 47.1 / 40.1, selection-land 31.2 / 32.3, selection-sea 59.1 / 54.8 (deuteranopia / protanopia). The sea tests: TheOpenSeaStaysBlueUnderTheHaze b - r 0.216, BandsStayOffTheWater 633,139 sea pixels (100k bar), TheBreakersAreFilteredToThePixel p99 0.0129 (0.014 bar), TheBakesCarveTheRelief 0.1021 -> 0.1134 (x1.111, 1.08 bar). ThePrimaryStripeIsTheDeeperBlueOfTheRuling 2.1 deltaE76 from the ruled blue, against 10.6 for main's lighter blue. Thin margins: the highlights' red is over their blue by 0.005 (the warm-highlights check), diorama02's warmth is 0.004 under its bar, and the open sea sits at the coordinator's hold (luma 0.33, p5 0.27).

  • Frame time (headed WebGL ?vprofile through tools/headed.py profile, 1920x1080, DPR 1, frame rate uncapped, under machine_lock.py --exclusive --wait-unity; main 732ceab8 and this grade interleaved main, grade, grade, main; medians of the average): gameplay03 all 8.51 / 8.72 ms against main's 9.26 / 8.44, diorama02 16.27 / 17.16 against 19.73 / 16.55; p95 10-11 and 17-18 ms (main 10-11 and 17-22). The grade's own cost (all minus -grade) is -0.12 / -0.24 ms at gameplay03 and -0.27 / 0.00 at diorama02, against main's +0.22 / +0.05 and +2.71 / +0.10 (main's first round ran on a busy machine): noise. Only the LUT's inputs change; the passes and shader variants are the same.

  • Labels (the label pills are UI, drawn after the grade, over the same 7 gameplay03 and 11 diorama02 label boxes as main's shot). Text against the pill's fill (WCAG): gameplay03 at least 5.5 (main 5.7), median 8.6 (9.0); diorama02 at least 6.3 (6.7), median 7.8 (8.2). Fill against the ground around it: gameplay03 at least 2.07 (2.01), diorama02 at least 2.34 (2.33).

HUD round 1: the info panel and the gilt trim (#398, #397)

Art-track gap 4, first step. The gap list and the later steps (the minimap; the top bar and card head, which wait for owner rulings) are in the issue. The settlement badge, the population tier in a city's or town's pill, has shipped: owner-approved as shipped (#710, 18 October 2026, "Accept as shipped"; MapLabelsPresenter.TierClass).

  • The info panel: before a hex is chosen it is a one-line hint (promo 03 shows no panel until a tile is chosen). The hex's facts are icon rows with gold signs (UiIcon region, house, village, flag): the region, the settlements, the 1948 localities and the placement. There are no yields (owner ruling) and no land-use name, because the land-use bake is modelled, not a sourced record. The region line wraps (#397). One refusal shared by both map actions is said once, without buttons, which takes the 1948 panel from about 385 to 300 px tall.
  • UiFrame.Filigree: a second inset rule and a scroll at each corner (on a panel at least 120 px tall the scrolls also run down the sides). It is drawn with Painter2D on the date plate, the event card, the info panel and the minimap, and takes no clicks.

Allocations: GC.Alloc samples over 200 renders (Recorder("GC.Alloc"), a temporary probe run on both trees) were the same on main e5bbf0f and on the branch: the info panel 15,600 and the top bar 12,800. No per-frame work is added.

Frame time, ?vprofile, WebGL headless 1920x1080, DPR 1, machine_lock.py --exclusive. Two interleaved rounds per build, main e5bbf0f against the branch on it (median avg ms, spread in brackets):

Pose main all branch all main none branch none
gameplay03 8.09 (0.03), 8.03 (0.07) 7.96 (0.16), 8.06 (0.06) 5.25 (0.10), 5.32 (0.23) 5.32 (0.04), 5.27 (0.10)
diorama02 18.87 (0.72), 19.10 (0.61) 18.88 (0.61), 18.99 (0.59) 12.23 (0.02), 12.33 (0.19) 11.66 (0.32), 11.57 (0.16)

Headed at DPR 2 (headed.py profile --dpr 2). The first run compared main e5bbf0f with the branch; then two interleaved rounds compared main db99f3a with the branch merged with it. Median avg ms, then p95, main first:

Pose, config run 1 (e5bbf0f) round 2 (db99f3a) round 3 (db99f3a)
gameplay03 all 14.61 / 14.62, p95 18 / 18 14.53 / 14.59, p95 18 / 18 14.59 / 14.56, p95 18 / 18
gameplay03 none 8.66 / 8.64, p95 14 / 14 8.78 / 8.40, p95 14 / 14 8.57 / 8.43, p95 14 / 14
diorama02 all 30.77 / 31.03, p95 36 / 40 30.95 / 31.26, p95 34 / 39 31.90 / 30.99, p95 36 / 39
diorama02 none 20.46 / 20.46, p95 23 / 22 21.32 / 20.69, p95 24 / 23 21.17 / 20.63, p95 23 / 23

The averages are within the spread everywhere; diorama02 has no hex chosen, so the branch draws a smaller panel. The p95 at diorama02 all at DPR 2 was 3 to 5 ms higher on the branch in all three runs, so it was not noise.

Bisect (DPR 2, diorama02 all p95, two interleaved rounds, main db99f3a first). At the shot poses the event card is closed, so variant (c) draws no visible filigree:

Build round 1 round 2
main 36 35
(a) the branch without the filigree 36 36
(b) only the info panel and refusal changes 35 36
(c) the filigree on the event card only 35 36
the branch (filigree on the date plate, info panel, minimap, card) 39 39

Cause. The filigree drawn as Painter2D curves. An EditMode probe found no regeneration: over 120 idle frames, and with the minimap's view box set every frame (still or moving), no UiFrame regenerated its mesh and no element got a GeometryChangedEvent. So the cost is in drawing the curves' tessellated, antialiased geometry each frame, not in rebuilding it.

Fix. UiFrame.Ornament rasterises the scrolls and the leaf once, in code, into a 64x64 alpha texture (two variants, with and without the side scroll; no asset). Each corner draws one tinted, mirrored, textured quad, and the second rule is four plain quads. The look is unchanged at 1920x1080. Re-run, DPR 2, interleaved, main db99f3a, the fix, then the old branch (avg ms, then p95):

Pose, config round 1 round 2
diorama02 all 30.78 / 30.61 / 31.29, p95 35 / 33 / 39 31.57 / 30.53 / 31.93, p95 36 / 34 / 38
diorama02 none 20.49 / 20.77 / 20.82, p95 23 / 24 / 23 21.35 / 20.35 / 21.09, p95 24 / 23 / 23
gameplay03 all 14.64 / 14.99 / 16.10, p95 18 / 23 / 23 14.51 / 14.53 / 14.45, p95 18 / 18 / 18
gameplay03 none 8.74 / 8.77 / 9.66, p95 14 / 14 / 17 8.64 / 8.67 / 8.37, p95 14 / 14 / 14

The fix's diorama02 all p95 (33, 34) is back at or under main's (35, 36). In round 1, gameplay03 all was disturbed for both branch builds in the same stretch: spreads 1.36 and 1.68 against main's 0.01, and the old branch's none was up too. Round 2 is clean, with all three at p95 18. golden&smoke=5 is unchanged on the fix build.

HUD round 2, part 1: the minimap (#398)

The owner chose all three round-2 pieces (2026-10-14), one PR each: the minimap, the left icon column, the top bar. This is the minimap, against promo 03's (T54: "the minimap is a pixelated hex strip").

  • The picture is a bake, tools/minimap_bake.py -> unity/Assets/Resources/Map/MinimapRelief.png (256 x 1024, powers of two so it is block-compressed on the GPU, mipmapped, trilinear): a hill shade of the heightmap (heights averaged over 3 x 3 samples, x12, lit from the north-west) over the biome colour bake and the land use (softened, a third parchment), the sea from coast.png (dark navy, a lighter band within 6 km of the shore, a thin pale line at the edge), and land outside Mandatory Palestine greyed with a gold rim. It replaces the 128-pixel point-filtered strip that MinimapPresenter.Paint drew at boot, so boot loses that loop and nothing is drawn per frame. tools/tests/test_minimap_bake.py re-bakes it in CI and compares the pixels.
  • Promo 03's three round buttons, beside the picture: zoom all the way out (the camera's MaxDistance, over the middle of the map), show or hide the settlement names (the labels switch), show or hide the borders and their caption (the borders switch). A switched-off layer dims its button. The picture stays mouse-only (accessibility audit B1d); the buttons take Tab after End Turn.
  • The panel is 222 x 380 (was 160 x 340): the picture 133 x 356, the buttons 50 px.

Measured on the WebGL ?shot frames at 1920x1080 against promo 03's minimap (T54's definitions; the picture resized to 300 px tall first; land is every pixel that is not blue-dominant):

promo 03 main gameplay03 branch gameplay03 branch diorama02
panel, share of the frame 4.97% 2.62% 4.07% 4.07%
picture, share of the frame 3.11% 1.80% 2.30% 2.30%
land luma, contrast (std) 0.467, 0.109 0.609, 0.078 0.492, 0.098 0.497, 0.106
land p10-p90 0.260 0.163 0.259 0.271
land relief (mean abs Laplacian) 0.063 0.024 0.069 0.074
land saturation 0.327 0.353 0.277 0.275
dE to promo: land, sea 0, 0 17.2, 27.4 3.3, 2.4 3.7, 2.6

The picture is still taller and narrower than promo 03's, which crops the country; the minimap shows all of it. The promo's region tints (orange and blue by side) are not in the bake: the two-colour borders (#555) change by date.

Frame time, ?vprofile headless 1920x1080 DPR 1, and boot (headed.py boot --runs 3), one machine_lock.py --exclusive slot, main df17ac6f first, then the branch (median avg ms, spread):

main branch
gameplay03 all / none 8.39 (0.09) / 7.18 (0.10) 8.38 (0.06) / 7.11 (0.03)
diorama02 all / none 15.96 (0.48) / 11.41 (0.09) 16.52 (0.73) / 11.99 (0.10)
boot map_build / first_frame, cpu 1 565 / 3924 ms 563 / 3917 ms

That first pair was not interleaved. The re-run (one --exclusive slot, 10 minutes; both builds with a temporary, uncommitted FrameProfile patch that measures only all and none, in the order base, branch, branch, base; two diagnostic branch builds; a headed DPR 2 pair) shows no cost:

run (order) diorama02 all diorama02 none gameplay03 all gameplay03 none
1 main 16.64 (0.48) 11.83 (0.29) 8.60 (0.61) 7.12 (0.28)
2 branch 16.46 (2.90) 11.81 (1.21) 9.92 (0.08) 8.19 (0.09)
3 branch 15.59 (0.44) 11.24 (0.07) 8.37 (0.09) 7.08 (0.05)
4 main 15.45 (0.33) 11.23 (0.06) 8.25 (0.07) 7.07 (0.09)
5 branch, picture without mipmaps 15.39 (0.42) 11.20 (0.06) 8.30 (0.03) 7.01 (0.03)
6 branch, main's 160 x 340 panel, no buttons 15.42 (0.37) 11.19 (0.15) 8.30 (0.07) 7.07 (0.02)
DPR 2 headed, main 33.73 (0.15) 27.90 (0.15) 19.71 (0.94) 18.33 (0.56)
DPR 2 headed, branch 33.49 (0.05) 27.86 (0.10) 19.34 (0.29) 17.81 (0.02)

Runs in the same position agree within 0.1 ms (1 and 2; 3 and 4), and the 0.6 ms step follows the position in the slot, not the build: the machine was slower early in the slot (run 2's gameplay03 is the same drift). The variants that remove the mipmaps or the larger panel and its buttons cost the same as the branch. The first pair's +0.56 ms was this drift.

HUD round 2, part 2: the left icon column (#398)

Promo 03 and T29 have a column of round icons at the left edge under the bar. This step moves the four screens' buttons (Cabinet, Budget, Defence, Why?) there from the top bar. It adds no action and no second copy of a button.

  • Icon-only, as in the promo. Each button keeps its name as its text, because the event log and the strings read it, but draws it clear; the name is also its tooltip, and HoverTip (from #584) shows it beside the disc on hover and on focus.
  • The column comes before the top bar in Game.uxml, so Tab still runs Cabinet, Budget, Defence, Why?, New Game, End Turn. It is covered where the top bar is (the start screen, an outcome clip) and stays usable under the decision card, as the top bar does: the card's backdrop starts at its right edge. The card stays centred in the window.
  • The screens, the wide screens, the narrow HUD's full-width screens and the toast start right of the column (84 px), so a screen never hides the button that closes it. The toast is 460 px wide (was 520), to stay clear of the screens.
  • The Hebrew UI mirrors rows, not the screen's corners (#550): the column stays at the left, as the minimap stays at the right and the info panel at the left.
  • The top bar now holds New Game and End Turn. Its height and its tiers are part 3.

Measured on the WebGL ?shot frames at 1920x1080. The discs are found as the navy fill in the left 5% of the frame, between 7% and 47% of its height:

frame discs disc width, share of height pitch left edge, share of width first disc's top
promo 03 5 5.21% 6.32% 0.87% 12.1%
demo2 T29 5 5.56% 6.30% 0.31% 12.1%
branch gameplay03 (EN, HE, diorama02 the same) 4 4.72% 6.30% 0.94% 10.8%

Frame time (?vprofile with the same temporary all/none-only FrameProfile patch on both builds, headless 1920x1080, then a headed DPR 2 pair) and boot (headed.py boot --runs 3), one --exclusive slot of 8 minutes, interleaved base, branch, branch, base. Base is #584's head 916cbad. Median avg ms (spread):

base branch branch base
gameplay03 all 8.16 (0.05) 8.21 (0.08) 8.18 (0.01) 8.21 (0.04)
gameplay03 none 7.02 (0.04) 6.97 (0.04) 7.02 (0.03) 6.98 (0.03)
diorama02 all 15.49 (0.40) 15.55 (0.43) 15.61 (0.36) 15.46 (0.43)
diorama02 none 11.08 (0.08) 10.73 (0.09) 10.74 (0.17) 11.16 (0.05)
boot map_build / first_frame, cpu 1 533 / 3877 ms 530 / 3866 ms 532 / 3862 ms 528 / 3852 ms

DPR 2 headed, base then branch (spread): gameplay03 all 19.14 (0.29) / 19.09 (0.19), none 17.81 (0.03) / 17.84 (0.05); diorama02 all 33.79 (0.21) / 33.81 (0.25), none 27.91 (0.04) / 27.87 (0.01). The column costs nothing measurable. Diorama02 none is 0.35-0.4 ms lower on the branch in both positions, outside the spreads; at DPR 2 the same cell is equal (27.91 / 27.87). It is not explained (the branch draws four fewer captioned buttons in the bar, but that is not measured on its own), so it is not claimed as a gain.

HUD round 2, part 3: the top bar (#398)

T54 measured the bar at 8.8% of the frame's height against promo 03's and T29's 6.5%. The one-row bar is now 70 px:

  • The date plate hangs below it, as promo 03's "1936 / Turn 1 / January 1936" plate does. It is at least 98 px tall and grows if the date wraps to a third line. #598's "Score {n}" stays at the far end of the year line.
  • The meters keep two lines, caption then value and change. Security's "Level 5 of 5" sits beside its caption; it was a third line. Captions are 15 px, values 21 px (24 on the roomy bar), with the icons as before.
  • New Game is a round icon with no caption, as promo 03's bar buttons are. Its name is its text, drawn clear (the event log reads it), and its tooltip, which HoverTip shows beside it on hover and on focus.
  • End Turn is 20 px with its attention line at 14 px.
  • With the four screen buttons in the left column (#602) the bar needs about 248 px less, so each width tier starts that much narrower (HudPresenter). Logical width is the window's width over the panel's scale, which is the plain mean of the two ratios, (w/1920 + h/1080) / 2 (logged by TopBarTests). The one-row edge is 1640, between a fullscreen 5:4 and 4:3 (review of #624 asked whether a fullscreen 5:4 could keep the one-row bar; at 1563-1612 it wrapped "Public Support" and "International Standing" (the captions' names then; #622 and #683 renamed them) with the widest values, while at 1635 it still fit, so 1640 is a margin, not a limit). So 16:10 (1819) and 4:3 (1646) keep the one-row bar with its icons, a fullscreen 5:4 (1585) takes the stacked bar with them, and a square window (1382) the stacked bar without. The taller tiers' heights are unchanged.
  • The panels under the bar keep their 112 px top, which clears the hanging plate. Under a decision card a full-width band of the backdrop runs from the bar's bottom to the column's top (event-scrim-top), so no strip of map stays live and the hanging plate is dimmed evenly; the column's strip starts at the column's top (TopBarTests.WithACardOpenTheBackdropMeetsTheBar).
  • The plate is a label occluder (date-plate in MapLabelsPresenter.OccluderNames), since the bar's box no longer covers it (MapLabelTests.APillUnderTheHangingPlateIsHidden).
  • Security's level and the attention line are 15 px, the accessibility audit's floor (H4); they were 13 and 14.
  • The owner's text floor (15 October 2026): the bar's small text stays at least 12 px on screen at small windows. The panel scales with the window (0.67 at 1280x720, 0.62 at 1024x768), so HudPresenter sets it to the whole px that keeps 12 on screen (18 px at 1280x720, 20 at 1024x768; in a browser, CSS px: the panel's scale over devicePixelRatio). Raised captions stay on one line, the turn box widens with them, and the tiers read the bar's width less the room this takes (FloorGrowth), so 1024x768 takes the stacked bar without icons; above 18 px the one-row bar's turn box no longer fits its 70 px, so 1366x657 takes the stacked bar. The phone tier (three rows) keeps 15 px: its fixed rows have no room for 25 px captions (a follow-up, by the owner's answer). The plate's turn and date line takes the same floor (the owner's second answer): the plate keeps its 230 px width, the turn and the date take a line each, and the plate's padding (18 px at the sides, 9 at the top and bottom) keeps every line inside its gilt frame's inner rule and at least 0 px from the 18 px corner zones (the scrolls' reach of about 14 px, plus 4; the score sits on a zone's edge at 1280x720, 4 px clear of the scroll); the plate grows to hold them, taller and, where the date needs it (23 px at 960x600), wider than 230 px, and the tiers count the extra width.
  • In Hebrew the bar's row mirrors (#550): the plate is at the right and End Turn at the left. The left column and the minimap keep their corners.

Measured on the WebGL ?shot frames (measure_bar.py: the rows from the top until the frame's middle third stops being dark; promo 03 and T29 include a 3 px gold rule):

frame bar share of height plate bottom share plate below the bar
promo 03 103 px 6.7% 130 px 8.5% 27 px
demo2 T29 73 px 6.8% 91 px 8.4% 18 px
before (#602), gameplay03 and diorama02, EN and HE 95 px 8.8% 95 px 8.8% 0
after, gameplay03 and diorama02, EN and HE 69 px 6.4% 97 px 9.0% 28 px

At 21:9 (2560x1080) the panel scales by the plain mean of the two ratios, so the 70 px bar is 7.6% of the frame.

Frame time (?vprofile headless 1920x1080 with a temporary, uncommitted all/none-only FrameProfile patch on both builds, then a headed DPR 2 pair) and boot (headed.py boot --runs 3), one --exclusive slot of 8 minutes, interleaved base, branch, branch, base; base is #602's head 192fdf31. Median avg ms (spread):

base branch branch base
gameplay03 all 8.19 (0.07) 8.14 (0.07) 8.12 (0.04) 8.19 (0.04)
gameplay03 none 7.02 (0.06) 6.96 (0.07) 6.94 (0.05) 7.02 (0.07)
diorama02 all 15.75 (0.39) 15.72 (0.41) 15.74 (0.54) 15.55 (0.38)
diorama02 none 10.74 (0.05) 10.69 (0.10) 10.74 (0.11) 10.72 (0.11)
boot map_build / first_frame, cpu 1 540 / 3761 ms 567 / 3786 ms 562 / 3792 ms 541 / 3779 ms

DPR 2 headed, branch then base (spread): gameplay03 all 19.09 (0.24) / 19.10 (0.20), none 17.75 (0.02) / 17.85 (0.02); diorama02 all 33.96 (0.27) / 33.79 (0.19), none 28.08 (0.14) / 27.97 (0.10). Frame time is equal within the spreads. Boot's map_build median is about 24 ms higher on the branch in both positions, but its single runs overlap base's (base 524-552 ms, branch 546-570 ms), the change touches no map code, and both are far inside the 743 ms budget at cpu 1.

Issue #391: the terrain shader's frame-time cliff (warm land shade, not shipped)

#391 asked for the shaded hill ground to read warm brown, as promo 03's does (about (67, 53, 46)), rather than olive-grey. The look was solved; the cost was not, so nothing shipped. What was measured is a constraint on any future addition to StatecraftTerrain.hlsl.

Diagnosis (main, editor, Mobile tier). On bare terrain (kits, trees, routes, borders and tilt-shift off) the shaded ground at hills12a (35.05, 31.76, 12 km) measured (71, 62, 55): hue 27 degrees, G/R 0.87. Most of the olive cast in the full frame comes from tree crowns and kit shadows, which belong to vegetation round 2 (#34). The grade's split toning cannot fix the ground's part: every shadow-colour candidate also moved the sea at gameplay03 by 4-18/255.

The look that worked. Scaling green by 0.87 and blue by 0.76 against red, weighted by 1 - smoothstep(0, 1, N.L) for the fixed sun, gave on bare terrain: hue 24.0 / 25.7 / 27.0 degrees and G/R 0.845 / 0.847 / 0.838 at hills12a, judea30 and gameplay03, sea drift 0.00-0.01/255, palette distance 0.463 -> 0.454 (gameplay03) and 0.581 -> 0.558 (diorama02), and WebGL diorama02 warmth 0.121 -> 0.127. Baked into terrain_bake.png's R (unread at run time) it cost 120 KB of data.unity3d under LZ4.

The cost. Headless ?vprofile, 1920x1080, DPR 1, interleaved under --exclusive, median avg ms, main 7040351:

Shader change all gameplay03 none gameplay03 all diorama02 none diorama02
main 8.23-8.42 4.98-5.07 19.3-21.6 12.6-13.2
no-op function call (control) 8.05 4.86 22.55 14.67
GetMainLight(), dot, smoothstep, lerp 14.96 13.23 21.4 16.8
_MainLightPosition uniform, dot, smoothstep, lerp 12.59 11.71 25.26 17.46
baked mask: bake.r kept, one constant lerp 14.87 12.55 22.09 17.35
baked mask used at once, AO sampled after it 16.72 13.72 21.22 18.06

Every variant that adds work to the function costs 4-8 ms at gameplay03, and as much with every visual toggle off as on: the toggles do not remove shader code. The no-op control matches main, so the build and the call itself are free.

Constraint. Treat the terrain fragment shader as at a cliff: one extra lerp of a value already in a register cost as much as per-pixel lighting. The likely cause is register pressure or occupancy in the WebGL (ANGLE/Metal) compile of the whole function, not the arithmetic, as #195 found for the water (3 ms from its size alone). Shortening a value's live range did not help. Before adding anything to StatecraftTerrain.hlsl, profile it as above against main; prefer changing baked data the shader already reads in its existing operations (land colours in tools/land_bake.py, MapPalette), or moving work out of the function as #195 did.

Issue #321: boot headroom, the biome colours from a bake

The budget is map_build_ms at --cpu 4 within 2720 ms (Task 1's 1813 ms plus 50%). After #502's hill scatter, main measured about 30 ms under it on a quiet machine and over it under load.

Where the map build goes. Main 72e2aee, tools/headed.py boot --runs 3 --cpu 4, two rounds under one machine_lock --exclusive (machine load 4.5-7), median dt of each phase in ms. The branch adds three ?bootlog markers (terrain object created, units placed, kit buildings built) that split the old phases, so its rows are the same work in finer pieces.

phase main branch what it is
heightmap decoded 190-193 190-192 r16z decode (counted from its start)
hexes parsed 12 10-12
localities parsed 81-83 78-79 the baked TSV (#291)
terrain heights set 1386-1395 1390-1400 new TerrainData (about 195 ms, #320's diag) and SetHeights
terrain object created (in the next row) 82-88 Terrain.CreateTerrainGameObject (the collider) and the layer
biome texture baked 438-442 38-41 main: the blend of 1024 x 1024 texels; branch: the bake read, the tint, the upload
border skirt 35-36 34-35
terrain water 17-19 19-21 the bake (#320)
settlements placed 7 7-8
units placed (in kits built) 354-357 route YAML read and corridors, footprints, unit data and spots
kit buildings built (in kits built) 253-260 SettlementKits.Build and the kit meshes
kits built 812-828 216-221 main: all three rows above; branch: the trees (Vegetation.Plant, VegetationLayer.Init)
routes built 47 44-46 the ribbon bake (#436)
map built 14-15 14
map_build_ms (median of 3) 3050 / 3073 2745 / 2761

The three largest: terrain heights set (about 1390 ms, 46%), the kits phase (about 820 ms: units placed 355, kit buildings 255, trees 220) and the biome texture (about 440 ms). first_frame_ms: main 11225 / 11162, branch 10821 / 10929.

The change. MapBuilder.BiomePixels (the BiomeBlender weights of every texel over MapPalette.Ground) is baked by Editor BiomeColourBake into unity/Assets/Resources/Map/BiomeColours.bytes (3 MB raw, 71 KB under Brotli -q 11, and 155 KB more .data in a -compression disabled build, whose .data is LZ4: "SCB1", version, resolution, a check of the frame, hex size, blend radius, palette and every hex's centre and biome, then RGB). MapScene reads it through BiomeColourFile.Load; a bake for other inputs is refused with a warning and the colours are blended at load as before. The ground tint, the texture and its mipmaps are made as before, from the same texels: BiomeColourBakeTests.TheMapGetsTheSameGroundFromTheBake compares every texel handed to the tint with and without the bake. tools/tests/test_biome_colour_bake.py recomputes the header's check in CI, re-blends the texels with a numpy port (within 1 per channel: 7 of 3,145,728 channels differ by 1, at x.5 roundings) and pins a fingerprint of the C# that makes them, so an edit to the blend fails CI until it is re-baked (review of #529). Re-bake with -executeMethod Statecraft.Game.EditorTools.BiomeColourBake.Bake after changing hexes.json, the heightmap's bounds, the palette or the blend.

The biome phase went from about 440 ms to 120 ms (82-88 terrain object, 38-41 the rest); the map build from 3050-3073 to 2745-2761 ms, about 310 ms less. At this load both are over 2720 ms; on the quieter machine where main measured about 2690 ms, the same cut leaves about 340 ms of headroom.

What is left, by expected saving at --cpu 4:

  • terrain heights set (about 1390 ms): ship the TerrainData as an asset instead of SetHeights at boot. Up to about 1.2 s, but the riskiest: the asset must be copied before the map owns and destroys it, its LOD data must load without a recompute, and the build grows by about 8 MB before compression. Prototype and measure before planning further.
  • units placed (about 355 ms): not yet split. Likely the two YAML reads (route paths, units) and the unit spots; a baked runtime form, as #291 did for the localities, would save most of the parse. Low to medium risk, no visible change.
  • The kits (about 255 + 220 ms): bake the kit meshes and tree matrices from their inputs (settlements, localities, population, routes, unit clearings, the ?kitscale flags fall back). About 300-400 ms, medium risk (many inputs). Without a bake, the hill scatter's per-cell checks could be reordered for 10-20 ms with the same trees.
  • heightmap decoded (about 190 ms): smaller, untested.

#321 round 2: the units' step (T47)

The units placed span #529 uncovered, split by four more ?bootlog markers. Branch build with the markers only, one --exclusive run of 3 cold loads at --cpu 4, median ms: route paths read 127, route corridors 8, footprints 3, unit data read 51, units placed (now the unit spots alone: UnitGround and UnitPlacer.PlaceAll) 184.

  • Route paths (127 ms): YamlDotNet over route_paths_1948.yaml's 1,544 points. tools/route_paths_bake.py writes content/map/route_paths_1948.tsv (29 KB: one row per route, its points in one field, the version, the YAML's FNV-1a check, the counts of routes and points, and a last body_check row, FNV-1a over the route rows), and RoutePaths.ParseRuntime reads it with the YAML reader's checks. The YAML still ships (WebGL now fetches both: 29 KB more). A runtime file that is missing, of another version, made from another YAML, cut anywhere (even inside a number: the review of #547 found that version 1's counts missed that case, so version 2 adds body_check) or edited by hand is refused with a warning and the YAML is read. CI: test_route_paths_bake.py fails a stale file and pins a fingerprint of the C# readers; Unity's RoutePathsRuntimeTests compares every route and point bit for bit.
  • Unit spots (184 ms): UnitPlacer.PlaceAll made every hex's candidate spots (19 a hex in the compact rounds, with their trigonometry) for each anchor and round, then dropped all but the seven hexes within its one-hex limit. It now lists those seven first and makes only their candidates, in the same order. UnitPlacerSpeedTests keeps the old algorithm verbatim and gets exactly its spots and skips on four grounds, from all clear to none clear.
  • Not changed: unit data read (about 35-51 ms; the unit YAML is also read again by GameRoot), the corridors and the footprints.

Base abe2263 against the branch, ABBA under one --exclusive hold (machine load 3.7-4.6), median of 3 each:

base round 1 branch round 1 branch round 2 base round 2
map_build_ms 2507 2214 2196 2379
units' span (base: units placed; branch: its five phases) 315 90 91 291
start screen shown 576 427 426 528
first_frame_ms 9165 8249 8217 8351

The units' span goes from about 300 to 90 ms. The map build is at least 165 ms shorter: the closest base and branch rounds differ by 165 ms, and the two base rounds spread by 128 ms, so the mean difference (about 240 ms) is not a firm figure. start screen shown falls by about 120 ms, since the units layer places the tokens again there.

#321 round 3: the units file parsed once (T52)

units_1948.yaml was parsed at boot by the map (unit data read, for the unit spots) and again by GameRoot for the units layer. Three ?bootlog markers round the layer (screens wired, units layer data read, units layer built), one --exclusive run of 3 cold loads at --cpu 4 on main 4875d0e plus the markers, median ms: the map's parse 31, the layer's parse 25, the layer's build 62 (its own run of the unit placer included), and screens wired 272.

The map now keeps its parse (MapScene.UnitEntries) and GameRoot takes it through MapScene.UnitEntriesFor, which hands back the same list when the text is the one the map read and parses otherwise. UnitDataOnceTests checks that the layer's list is the map's (it failed with the old second parse), that it holds the same units as a fresh parse, field for field, and that a rebuild or other text never hands on a stale parse. New Game replaces only the session; the units are content, read once for the layer's life as before.

Base 4875d0e against the branch, base, branch, branch, base under one --exclusive hold (load 2.8-11.9; the first base round stalled after one cold load when another process loaded the machine, so the base has four loads):

base branch round 1 branch round 2
units layer data read (25, the split) 1 1
map built to start screen shown, per load 437, 444, 445, 456 417, 422, 434 415, 416, 419
map_build_ms (median) 2221 2206 2214

About 25 ms off the time from the map to the start screen; the map build is unchanged, since the parse that stays is the map's. Not done: the layer's second run of the unit placer (about 40 ms of units layer built) could take the map's spots too, and screens wired (about 270 ms) is not split yet.

#321 round 4: after the regions bake (#601), T66

One clean measurement of main 7a4617ed (after #601's region files from baked runtime forms) against the 2720 ms budget: a -compression disabled WebGL build, tools/headed.py boot --runs 3 --cpu 4 under one machine_lock --exclusive hold (load 3.4 at the start; the job itself 28 s), median of 3 cold loads:

run map_build_ms first_frame_ms
1 2283 8223
2 2273 7983
3 2246 7964
median 2273 7983
  • Headroom: 447 ms under the 2720 ms budget, which meets #321's target of at least 300 ms. first_frame_ms is 7983 against its 10445 ms budget.
  • The largest phase is still terrain heights set (TerrainData.SetHeights): 1437, 1429 and 1413 ms. Next come kit buildings built (about 185 ms) and kits built (the trees, about 162 ms).
  • The kits-and-trees bake stays parked (#321, T66 comment of 15 October): its trigger, headroom below about 200 ms, is not met.
  • Main has moved since (HUD round 2: #584, #602; #598's ?mvp flag). They are expected to run after map built, and so outside this span, but that was not checked and they were not re-measured.

The art guards mask the HUD (T74, #388)

Promo 03 has its own HUD, so the guards leave the same screen regions out of both images rather than hide ours. tools/art_compare.py's HUD_MASK, as (x0, y0, x1, y1) fractions of the frame, is the top bar (0, 0, 1, 0.07), the date plate hanging below it (0, 0, 0.13, 0.10) and the left button column (0, 0.10, 0.045, 0.44). metrics() and histogram_distance() apply it by default; mask=() gives the whole frame. A mask that leaves less than half the frame raises. art_compare.py's art-log rows are named <checkpoint-name>-hudmask (as art_checkpoint.sh's -units rows), since masking alone moves main's gameplay03 palette distance by about 0.027, more than the main + 0.02 tolerance: compare masked rows only with masked rows.

  • Why. #624 alone moved diorama02's whole-frame warmth 0.1409 -> 0.1465 and gameplay03's palette distance 0.3505 -> 0.3681 (988ad417 with only #624 cherry-picked, 3 WebGL ?shot captures each). About half is the shorter navy bar; the rest is the camera reframing to the new bar height (TopBarInset -> CameraRig.TopInset): with the top 10% and the left column left out of both frames, diorama02 warmth still moves 0.1643 -> 0.1667.
  • diorama02 warmth band: [0.087, 0.151] (was [0.087, 0.145]). The top is promo 02's warmth measured the same masked way, metrics(promo/final/02-map-diorama.png, mask=HUD_MASK) = 0.1509 (whole frame 0.1454, the old 0.145). Promo 02 has no HUD: on it the boxes are geometric, matching where our HUD sits in our diorama02 frame, and they remove a dark backdrop band, part of the title plaque and some sea. The floor 0.087 has no recorded derivation (first in 9d28c722); possibly 0.6 × the old whole-frame top. It is kept.
  • The other bars are unchanged: gameplay03 palette distance to promo 03 <= 0.44 and <= main + 0.02, diorama02 saturation <= 0.48, now all measured masked.

diorama02 warmth per material (#695, T77)

The owner ruled on #695: "Measure each material". Whole-frame warmth compares geography as much as colour: masked, promo 02 is about a quarter sea and 6% dark wood, our diorama02 an eighth sea and a third table, and that mix alone puts our frame 0.028 warmer while our materials' colours put it 0.015 cooler (T77's split of the 0.0133 gap). So diorama02's warmth guard now compares each material with the same material in promo 02.

  • Classifier (tools/art_compare.py, classify()): neutral (chroma (max - min) / max under 0.08), sea (B > R), green (G > R), then the warm pixels by luma: wood under 0.30, land 0.30-0.55, sand from 0.55. The cuts sit between hand-picked region means (wood 0.13-0.26, our table 0.18-0.20; hills 0.35-0.50; Negev sand 0.70-0.71), not in gaps of the pixel histograms: ours has a valley at 0.25-0.50, promo 02's runs flat there and 0.55 sits on its ramp. "Sand" is therefore bright warm ground: the Negev, the beaches, and sunlit hills and stone towns in promo 02. Sensitivity of the sand verdict on main: FAIL for a sand cut of 0.45, 0.50, 0.55 and 0.60 (ours 0.287-0.307 against floors 0.288-0.313); at 0.65 it passes by 0.001 (0.3206 against 0.3197).
  • Mask. The material guard uses MATERIAL_MASK: HUD_MASK plus the minimap (0.86, 0.62, 1, 1), the "Click the map" plate (0, 0.92, 0.17, 1) and the map-layer legend (0.76, 0.94, 0.87, 1). Their navy is B > R, so unmasked they read as sea: about 16% of main's sea pixels, which made it -0.1729 instead of -0.2000. The boxes apply to promo 02 alike; there they cover open sand, a corner of the Dead Sea and some hills, no UI (without them on promo 02: sand 0.3811, sea -0.2598, within 0.002). HUD_MASK and the other guards are unchanged.
  • Bands. The gates are sand and sea. Each band is promo 02's mean warmth (R - B) for the material, masked with MATERIAL_MASK, plus or minus 20% of its size: the old whole-frame band [0.087, 0.145] was 40% of its promo anchor wide, here centred on the promo. Sand [0.3037, 0.4556] (promo 0.3796), sea [-0.3119, -0.2080] (promo -0.2599). No band is fitted to our output. Known limitation: promo 02's "sea" also holds its glass towers and solar fields (B > R): 9% of its sea pixels (chroma under 0.2, warmth -0.048); without them its sea reads -0.281, so the sea band's centre is about 0.02 warmer than promo 02's water. A simple chroma cut does not separate them cleanly (it also drops shore and salt pixels and keeps the bluest panels), so they stay in. The sand gate is the class mean. As a check, a pure-sand crop of the open Negev west of Beersheba (ours (0.22, 0.80, 0.34, 0.93), promo 02 (0.08, 0.86, 0.28, 0.97), both over 99% sand) reads 0.344 against promo 02's 0.468: the same direction, larger.
  • Wood is printed, not gated. It is not like-for-like: promo 02's dark backdrop lies mostly under HUD_MASK's top-bar box, so promo 02's "wood" is mostly the frame's walnut and terrain shadow (6% of the frame), while ours is mostly the table (a third of it). Main reads 0.1126 against promo 02's 0.0908.
  • Which image is gated. Only the WebGL ?shot (directly under the checkpoint folder): it is what the owner sees. The editor renders (editor/PC-diorama02.png, editor/Mobile-diorama02.png) are printed as information. Every MATERIAL line names its file.
  • Fail-closed. A gating material covering under 2% of the measured pixels of either image (about 37,000 pixels at 1920x1080) cannot be measured. A gate out of band is a failure, not a note: art_compare.py exits 3 and prints GUARD FAIL: <material> <value> vs <band> for each; it exits 4 when a gating material cannot be measured (1 when no screenshot matched a pose, 2 on a blank capture). art_checkpoint.sh stops on 1, 2 and 4 as on any error; on 3 it finishes the checkpoint (its images are what the owner reviews), lists every GUARD FAIL line and exits 3.
  • Whole-frame warmth is still printed, as information only; [0.087, 0.151] (T74) no longer decides.
  • Where main stands (d031af36, 3 WebGL ?shot captures, spread 0.0000): sand 0.2952 (-0.085, paler and less orange) and sea -0.2000 (+0.060, less blue; 0.008 short of the band's edge): both outside their bands, so art_checkpoint.sh exits 3 on main until the art follow-up on #695 lands. This section only sets the measure.

Issue #391 QA record, 19 October 2026

Technical checks on main 9e57134201f74412bf3d7513e26b600a9c760f5c, macOS 27.0.1 arm64, Unity 6000.3.25f1. No grade, shader, threshold or asset changed. This is an agent-run QA record, not the person's judgement of judea30 requested in #391; that remains pending before closing the issue.

The three existing GradeTests passed: ColourBlindPlayersStillTellCoastAndSelectionApart, ColourBlindPlayersTellTheBorderStylesApart and ColourBlindPlayersTellTheNeutralBandFromBothSides. EditMode XML: 3 total, 3 passed, 0 failed, 0 skipped; exit 0 and 0 C# warnings/errors. These are the existing simulated-vision checks at their sampled points, not a human accessibility or full-suite sign-off. MapCapture.Promo -extraPose judea30:35.10:31.85:30000:10 also completed: six 1920x1080 editor renders (gameplay03, diorama02, judea30, each Mobile and PC).

WebGL reused the unmodified local buildB from the Firefox #437 check, built from the same 9e source; every file was checked against its original SHA256 manifest. Two web_check.py --query golden&shot attempts exited 1: gameplay03 received 624 of 625 chunks, so the collector rejected SHOT done. Both failed logs and partial capture sets were kept. Collecting the console through the existing headed.Browser CDP connection, headless, instead of stderr then completed all 13 shots through the same web_check.ShotCollector: all PNG SHA256s and 1920x1080 sizes verified, all 8 manifest goldens matched, exit 0. The standalone transport harness and exact commands are included in the contribution's PR description. No repository tool changed; the successful CDP run does not turn the two failed stderr runs into passes or establish their exact cause.

Current art_compare.py completed with exit 0. Its WebGL diorama02 material guards, using the existing MATERIAL_MASK and bands from T77 above:

Material Measured warmth Promo band Result
Sand 0.3130 [0.3037, 0.4556] PASS
Sea -0.2337 [-0.3119, -0.2080] PASS

The HUD-masked diorama02 saturation was 0.4192, below 0.48. Whole-frame warmth was 0.1486 (HUD-masked 0.1664), above the issue's old 0.145 top (and T74's masked 0.151 top). As T77 specifies, those whole-frame warmth bands are information only, not the current material gate. Neither the old band nor the current bands were adjusted.

For a reproducible diagnostic, dark non-blue pixels (B<=R, Rec.709 luma 0.06–0.30 on sRGB/255, outside HUD_MASK) gave judea30 median-colour hues 27.5 degrees on Mobile and 26.4 on PC. This selection also includes wood, props and their shadows; it is not terrain segmentation or the older bare-terrain probe. These numbers cannot replace a person's look, and no new hue tolerance or human acceptance is inferred. The full measurement script, counts, RGB medians and image hashes accompany the PR; the older before/after results in #391 remain attributed to their original authors and heads.

Captures and colour-blind comparison images were retained locally with hashes. Left panels of comparisons are in-engine captures at the source above; right panels are the existing AI-generated promo references, not gameplay. Pending: a person's look at judea30 and the transport anomaly noted above. No simulation hash, human review flag, licence declaration or release gate changed.

Issue #2431: brighter, glowing borders at the gameplay zoom

Branch visual/2431, on #2427's lighting. The band's width was not the gap: at gameplay03 A2 draws about 10 px against the promo's 8. The colour was: the blue stripe's saturation was 0.44 against the promo's 0.59 (72,114,176), and nothing marked the band as lit.

  • What changed. On A2, B2 and C2 (the stronger tier; Legacy, A, B and C are as they were): a brighter orange edge stripe, _ScBorderStripe2 (sRGB 0.92, 0.37, 0.14), while the blue stripe stays the owner's ruled blue (Primary's colour; C2 keeps its own blue; see "The owner's blue" below), and a halo along the inside of the light stripe (_ScBorderGlow: pale blue 0.66, 0.78, 0.97 and warm 1.02, 0.57, 0.22, opacity 0.72 at the stripe, falling as a square over 320 m, at least 6 px). Grey does not glow, so contested ground stays flat and dashed. The tint and wash keep Style.Colour.
  • Why a split colour. A brighter Colour also brightens the fill. At gameplay03 the fill at the new colours gave rose +4 -> +7, land warmth 0.232 -> 0.248, saturation 0.426 -> 0.435 and palette distance 0.308 -> 0.345, though its hue came nearer the promo (27.9 -> 26.5). With the stripe apart from the fill, the land's numbers do not move. B2's rim is drawn from the stripe colour, so on the orange side it follows the brighter stripe.
  • Why in the shader. Bloom is a full-screen pass and #2427 kept it off. The halo is drawn in the wash's room: no fetch (TheBordersReadOneTexelAPixel), no wider reach (it is never wider than the wash, GlowWidthPx, so the early-out stays exact, TheEarlyOutIsExactInEveryVariant, and a far hex whose wash has given way has no glow). Not under the stripes or a rim.
  • The tint is not the golden gap. Editor frames at gameplay03, composited with the UI, one session, the same state, the projection at 16:9 (the rig sets cam.aspect: #2882); the crop as in "Issue #38: grade retune". #2427's hypothesis was that the tint explains much of what is left. It does not:
Land at gameplay03 hue rose saturation palette distance
promo 26.6 +5 0.364
as played (tint 0.12, wash 0.45) 27.9 +4 0.426 0.308
tint 0.06 28.2 +4 0.424 0.305
tint 0 28.4 +3 0.423 0.303
tint 0, wash 0.25 28.5 +3 0.423 0.305
borders off 32.9 -5 0.407 0.309

(The issue's 0.40 is the Primary style's own strength; A2 draws 0.45, StrongWashStrength.) Taking the tint away moves the hue half a degree, away from the promo, takes 0.003 off the saturation and 0.005 off the distance. With the borders off the land is further from the promo on hue, rose and distance, and nearer on saturation (0.407): the territory colour helps the hue and costs some saturation; the distance barely moves. The tint and wash stay as they were; the rest of the gap is not the borders (hypothesis: the land and town art).

  • Rendered (TheStripesGlowAtGameplay03, proof edge on the coastal plain). Old colours against the first fit (blue stripe 0.26, 0.40, 0.61, since withdrawn): blue stripe saturation 0.45 -> 0.52, orange stripe value 0.79 -> 0.91, the glow's lift over the far wash -0.003 -> 0.130 (blue) and -0.015 -> 0.074 (orange). The two halos stand 38.3 apart under simulated deuteranopia and protanopia at worst (bar 15), and ColourBlindPlayersTellTheStylesApartInEveryVariant passes on every variant, Neutral included.
  • The owner's blue. The first fit's brighter blue stripe took ThePrimaryStripeIsTheDeeperBlueOfTheRuling from 2.6 to 8.8 deltaE76, past the owner's #519 bound ("bring the stripe back to about 3 deltaE76 of the ruled blue or less", above). Retuned on the owner's ruling of 7 Nov 2026 (relayed by the coordinator): the blue stripe is Primary's ruled colour again on every tier and the halo carries the Israeli side's glow; the test now asserts the bound (md <= 3) for the default, stronger tier. Expected about #2427's 2.6, since the stripe colour is #2427's again: to be measured (the halo's first pixel may still count in the stripe's read). With it, TheStripesGlowAtGameplay03's blue saturation bar is the ruled blue's: 89,119,156 has HSV saturation (156-89)/156 = 0.429, so the bar is at least 0.42 (the tightest under it), in place of 0.50; the bars on the blue and orange halo lifts, the halos' colour-blind distance and the orange stripe's value are unchanged. The numbers in "Rendered" above and in "WebGL ?shot" below are the first fit's, to be measured again.
  • WebGL ?shot, against #2427's final values. Gameplay03: land mean (145, 115, 87), hue 28.9, saturation 0.422, whole-frame warmth 0.091 and palette distance 0.32 on both sides. Overview: sand 0.347 -> 0.348, sea -0.237 -> -0.239, both inside their bands; Negev clipped 0.088% on both. art_checkpoint.sh passed: goldens, shots and material guards.
  • Cost: NOT MEASURED. ?vprofile needs machine_lock.py --exclusive, which the lanes may not take. From the source: on every land pixel of a drawn region, two colour selects, a max and a lerp; inside the band's branch, the halo (a division, a min, a max, a square). No fetch, pass or variant. That is an argument, not a number. Boot: NOT MEASURED. Web build, uncompressed: 567,916 KB by du (wasm 65,096 KB, data 36,072 KB); only a shader and C# changed, no asset.
  • Colour-blind copies. In the deuteranopia copy of gameplay03 the blue reads clearly; the orange reads as a pale yellow band, near the land's own hue on the plain and carried by its contrast. The rendered bars above pass; a player's eye on it is the owner's call.
  • Readability. Looked at in composited frames, before and after: label pills, the map legend, English and Hebrew; a Move pick (#2201's lit hexes) over the Jordan Valley front: its fill and outline draw over the bands, as before. The sea and the Negev are unchanged by eye and in the numbers (TheSunlitDesertDoesNotClip: 0.00%).

Issue #2427: second lighting and grade pass (warm golden land, sea held)

Branch visual/2427. The issue called the land flat khaki next to promo 03. Measured on main at the gameplay pose, in WebGL, the land's means are already near the promo's. What differs is the spread: the promo's lit ground runs wider (hue 15-40 degrees) and its shade is greyer, ours sits in a narrow terracotta band with a pink cast. Much of that band in the frame is the territory tint of the 15 May 1948 control layer (_ScBorderTint), which the editor's MapCapture.Promo does not draw: orange-red over the hills, blue over the Israeli plain, which turns its fields lilac-grey. That tint is the borders lane's (#2431), not the grade's. Hypothesis, not tested: most of what is left between this pass and the promo's hills is the tint and the land's own colours.

  • What changed. MapLighting: sun 1.6 -> 1.8, colour (1.00, 0.84, 0.64) -> (1.00, 0.83, 0.63), the fill about a tenth lower, as colours (0.47, 0.54, 0.65), (0.43, 0.415, 0.38), (0.235, 0.20, 0.155). MapGrade: contrast 18 -> 20, and a ColorCurves with three LUT curves: HueVsHue (terracotta up, greens down, to gold), HueVsSat (the land, and a narrow band at the sea's own hue), LumVsSat (dark tones 9% less saturated). Post exposure 0.45, vignette 0.26 / 0.45, temperature, tint, split toning and global saturation are untouched. The sun's elevation and azimuth are unchanged (27 degrees changed the picture little; the land bake's bush shadows follow azimuth 285).
  • Why hue curves. #38 found one global grade cannot warm the land and keep the sea. No hue moves at the sea or the stripes, so that limit does not apply. The ambient does not reach the water (ScWaterShade reads the sun, not the SH). The sun does: a stronger sun lights the sea with its colour, and the first version of this pass took gameplay03's open sea from warmth -0.202 to -0.187, the #38 failure (found in review; the changelog then wrongly said the sea stayed). The fix is a saturation band at the sea's own hue, 8% at 0.57-0.592 ending at 0.606, short of the ruled stripe, with the sun's blue back near main's (0.63) and less dark-tone desaturation. A band to 0.615 pushed the stripe to 3.1 deltaE76. All three curves are inputs to the same LUT: no pass and no shader variant added. Keys are flat (zero tangents): smoothed tangents overshot into the sea's hue and took 0.011 off its luma.
  • Bloom. Not brought back. In the editor, LDR bloom 0.2 above 0.8 moved the gameplay land's luma by 0.003 and the Negev's mean by about 3/255 against the same grade without it: nothing a player would see, so no cost run was needed. SSAO was not tried (Tier decisions: metre radius cannot reach km relief).

WebGL ?shot, 1920x1080, main 5f8fa9fcd against this grade on 14fcb0c3b plus these values; crops as in "Issue #38: grade retune":

promo main this grade
g03 land mean colour (130, 104, 84) (143, 114, 89) (145, 115, 87)
g03 land luma / p5 / contrast 0.425 / 0.145 / 0.187 0.464 / 0.148 / 0.179 0.468 / 0.147 / 0.188
g03 land saturation / warmth 0.364 / 0.181 0.404 / 0.212 0.422 / 0.230
g03 land hue, rose index 26.6, +5 27.2, +5 28.9, +2
g03 open sea luma / p5 / warmth 0.351 / 0.295 / -0.248 0.326 / 0.273 / -0.202 0.329 / 0.273 / -0.206
whole frame saturation / contrast / warmth, g03 0.37 / 0.196 / 0.078 0.40 / 0.174 / 0.080 0.42 / 0.179 / 0.091
palette distance g03 / d02 0.32 / 0.53 0.32 / 0.49
d02 sand / sea warmth (bands [0.304, 0.456] / [-0.312, -0.208]) 0.380 / -0.260 0.319 / -0.225 0.347 / -0.237
d02 Negev mean colour, pixels clipped in the crop (190, 156, 128) (201, 169, 113), 0.016% (205, 170, 111), 0.088%
d02 table mean colour (59, 44, 31) (58, 44, 34)
  • What held it back. A stronger first fit (sun 1.8, fill a quarter lower, land saturation 0.476 in the editor, against 0.401 on main) read more golden but failed five existing tests. No test or threshold was changed for this; the values were brought back until they pass: olive crown saturation 0.431 (bar 0.37), TheBeachIsWide... luma 0.685 (bar 0.72), TheBakesCarveTheRelief ratio 1.06 (bar 1.08), a neutral band's colour-blind distance 14.6 (bar 15) and the A2 wash 1.29 (bar 1.3). The beach's luma floor is why the land stays as bright as main's, 0.468 against the promo's 0.425. If the owner wants the stronger look, these are the bars to revisit: that is a design call, not a tuning step.
  • Readability. Editor renders at the gameplay pose: ThePrimaryStripeIsTheDeeperBlueOfTheRuling 2.6 deltaE76 (main 2.1, the #519 bound about 3), TheSunlitDesertDoesNotClip 0.00% in the Negev and at gameplay03, olive crown saturation 0.356 (bar 0.37). TheLandReadsGoldenAndTheSeaHolds guards the sea at gameplay03 in the editor: luma at least main's 0.368 and warmth at most main's -0.218 (it fails on the first version, -0.208). The editor read the second attempt as level with main while WebGL still read 0.005 greyer, so that guard is a floor, not proof: the WebGL number above is the one to trust.
  • Board mode (?hex-tiles). HexLook multiplies the sun by 1.22 (2.196 now, 1.952 on main) and sets saturation +30 over these curves. HexBoardCapture.Run -board tel_aviv -boardOnly, 1920x1080, in the editor on both sides: no pixel of the board at or over 0.995 in any channel on either side; mean board colour (136, 151, 57) -> (160, 154, 38), so its greens go more yellow, saturation 0.776 -> 0.802. The local ground textures are absent here, so this is the flat-colour fallback, not the textured board, and HexLook was not changed.
  • HUD. The top bar and date plate are pixel-identical between the two WebGL shots at gameplay03 (mean difference 0.00-0.01/255); the translucent help bar and legend move by 0.3 and 0.85/255 on average, and the minimap's frame by about 0.8/255 where map shows through. English and Hebrew composited frames were looked at, before and after. Label pills were looked at, not measured.
  • Cost: NOT MEASURED. ?vprofile needs machine_lock.py --exclusive and the lanes may not take it. By construction only LUT inputs changed (as in #38), but that is an argument, not a number; the sun and ambient are uniforms. Boot: NOT MEASURED. The web build, uncompressed, is 555 MB on both sides by du; wasm 65,092 KB and data 36,072 KB on this branch.
  • Seen, not changed. Close views (hero06negev) show contour bands on the terrain before and after; not caused by this pass.

Issue #2428: one shared atlas for the town kits

The kits drew vertex colours and a procedural window pattern. They now also sample one shared atlas, unity/Assets/Art/Kits/kit_atlas.png, painted by asset-sources/kit-atlas/make_kit_atlas.py (uv run; no input files, no image model, same bytes every run). It is a 4 x 4 sheet of 512 px cells imported as one Texture2DArray (16 layers), so a cell tiles and has its own mipmaps with no bleeding. The cells: five facades (White City, stone, hill village, mud plaster, moshava), plaster, ashlar, rubble, roof tile, flat roof, ribbon glass, balcony rail, corrugated iron, planks, and two spare white ones.

  • Mesh: the vertex layout, the meshes and the one material are as before, so no town gains a draw (KITS ... draws=65 before and after). The colour's alpha is now the surface, KitMesh.Facade: the layer + 1, times 17. uv0 is in cells; on a facade v counts storeys, and Kit.shader lays out one ground storey under three repeating ones. A facade wall counts whole bays, so a corner never cuts a window, and starts at a phase hashed from its position, so no random stream moves. Plain (trees, boats, tents, the ma'abara's patched sheet) has no cell.
  • Light: the atlas is albedo only. Stains and street dust are dirt in the colour, not shadow; KitAtlasTests checks that a solid cell has no brighter side. The shader's old contact darkening at a wall's foot stays.
  • Far off: mipmaps replace the window pattern's box filter; a wall fades to its mean with no moire. The atlas is sampled with explicit gradients taken outside the branch, because fwidth inside it is undefined on WebGL 2.
  • Ribbon windows start one storey up, so the ground storey shows the facade's doors and shop.
  • Size: the build's .data gzips (level 6, the player built with -compression disabled) 1,316,744 bytes larger, about 3.2% of the download, so the PR carries a nightly size budget line.
  • Frame time, not measured to this document's standard: ?vprofile needs the exclusive lock, which this lane did not take. A run under the shared lock at load 11-16, main against the branch, read gameplay03 23.53 -> 22.12 ms (spreads 2.93, 3.93) and diorama02 25.78 -> 28.03 ms (spreads 1.56, 2.98): a noisy reading, not a cost.

Issue #2877: a plowed field or drained swamp on its hex's ground

The owner's answer on #2812 was "Ground change only", so a built field or a drained swamp has no kit and no pill and the hex's ground shows it. SiteGroundLayer drapes one mesh per site over the terrain (about 2,100 vertices), drawn by Statecraft/SiteGround after the ground and before the roads, in the order of HulaPapyrusLayer: GameRoot calls it on every Changed, it reads State.Sites, and it builds a mesh only for a site that changed, so a loaded save draws its own. A repaint of _ScBiomeTex was rejected: it is one map-wide upload, which would fall on End turn. A zero-area mesh draws from the first frame, meant to keep the shader's first draw off the End turn that finishes the first site (measured below). The look is a design value (no field layout or crop is sourced): seeded parcels of ploughed earth, young crop and stubble with pale tracks; a drained swamp is dark earth with straight canals. The patch keeps off water, off ground that rises over 45 m (world) round a vertex and off the inner 70% of every locality's footprint, and it fades out short of the hex's edge, so the grid and the region borders stay on top. A hex with both the swamp and its field shows the field.

Cost, on a WebGL build (Build.Web -compression disabled), ?vprofile under the shared machine lock on a busy machine (load 4-9), so these are not the "How to measure" idle numbers. A throwaway patch, never committed, built seven sites and limited the run to all against -siteground at the poses below, 3 interleaved runs of 300 frames each:

Pose all ms (spread) -siteground ms (spread) cost ms
Beit Kama, 60 km 21.94 (3.43) 21.66 (0.31) +0.28
Beit Kama, 15 km 19.89 (0.47) 20.14 (0.16) -0.25
Beit Kama, 6 km 16.14 (0.24) 16.36 (0.16) -0.22
Hula, 6 km 15.27 (0.52) 15.91 (0.19) -0.64
gameplay03 25.27 (0.20) 25.41 (0.29) -0.14

All are inside the 2.5 ms budget and within this busy machine's noise, so they say "no measurable cost here", not a number. The first run at the first pose (24.85 ms, p95 48) is the 3.43 spread; the two runs after it were 21.42 and 21.94 ms. Not measured: the same run under the exclusive lock.

Time on the main thread, on a WebGL build at 220a7d0a9 with a throwaway timing patch, never committed, under the shared lock (load about 6-9), six page loads. Each field was left one turn from done in State.Sites, and the sim finished it in a real SimBridge.EndTurn, with the camera on the fields at 31 km:

Step ShowAt ms
The first field (one mesh plus PlacementLayer.CampGround) 2.5-5.3
Each later field (one mesh) 0.9-1.5
SimBridge.Replace with seven built fields, four of them new meshes 3.4-4.3

So one mesh is about 1 ms; seven new at once would be about 7 ms by that rate, not measured. The End turn around it took 668-2,253 ms with or without a field (EndTurn alone, on controls and on field turns alike), so that freeze is not this layer's. The load was the Replace path with a session whose sites were set by hand, not a save file read from disk.

The warm mesh: three runs with it and three without (sgnowarm in the throwaway patch) gave the same frames after the first field's End turn, the worst of the next 12 being 416-441 ms with it and 417-478 ms without, on this Mac's Chrome (ANGLE over Metal). Two controls without a field had a 601-621 ms frame in five of the six runs, so that spike is not traced to the field. The later fields' worst frame was 74-127 ms against about 50 ms on the control turn, with or without the warm mesh and outside ShowAt; the cause is not traced. So the warm mesh shows no gain here; its use on other GPU drivers is not measured.

With #2884 (the border glow) merged locally, the 60 km stills of Beit Kama and Mefalsim show the blue and white stripes whole: the patch stops inside them. It does cover the wash inside the hex where a parcel lies.

Issue #2878: ground detail and low 3D scrub between the towns

Survey (main 93cd2f174). Between the towns the only ground is landuse.png, 79 x 105 m a texel, so it reads as smooth colour at the 31 km opening and blurs at 12 km. The trees are 1,267 to 2,562 instances in view at 31 km and 779 to 1,464 at 12 km: large and far apart. Promo 03's land between its towns is covered with small bushes and worn ground.

Why not in the terrain shader. That function sits at the cliff recorded under issue #391, and land_bake.py is outside this lane. So the paint is a pass of its own, GroundDetail.shader: a Blend DstColor SrcColor mesh in the queue right after the terrain, with one tiling texture (grain, bushes as a height field, their shadows, stones) built at boot.

Standing scrub. GroundScrub.shader: one mesh of 1,089 bushes covers a 15 km tile and is drawn once for each tile in view, a few tile instances a frame, not one instance a bush. Its vertex shader stands each bush on a half-float height texture and shrinks to a point any bush whose rank is over the density. Its bushes stop at 55 m, under the 60 m lip of the smallest unit base (a compact token's), so a moved token covers them.

Where it grows (StatecraftGroundDetail.hlsl, shared). By biome (hexes.json), the land-use weight (rainfall and slope from land_bake.py), the bake's curvature (thin in gullies) and a broad clump noise; along the land bake's painted hedges; off the towns, the routes and the units' start clearings (a keep-out texture); off the beaches and dune coasts; off drawn region borders. Biome scrub shares are design values. It adds no paths: the land round 2 ruling keeps tracks to the 1946 Survey's roads. The shader-side bush shadow assumes MapLighting.SunAzimuth 285 degrees: if #2427 turns the sun, the offset in GroundDetail.FillTexels turns with it. A warmer or darker grade changes the multiply's result in proportion and is not expected to change its balance (hypothesis).

Tests that now switch the layers off. BorderTests.Open and TerrainShaderTests.TheBakesCarveTheRelief measure bare stripes and bakes, so they switch both new toggles off, as they do the kits and trees. Their thresholds are unchanged.

Issue #2917: furrows on the worked fields

What draws them. land_bake.py lays its parcels on a wobbled 5.2 km block grid, each block turned by a hashed angle, with 1,000 m strips along its rows. It now writes two more textures. furrows.png (1024 x 2048, R8) holds the furrows' strength: 1.0 on ploughland, 0.6 on green crops, 0.35 on stubble, none on fallow, grass or groves, and none on the rims, hedges and tracks. fieldblocks.png (34 x 85, RGBA8, read texel by texel) holds each block's angle and strip offset as 16 bits. The bake's angle comes from a 64-bit hash that a shader cannot repeat, but the wobble and the row coordinate can be repeated exactly. So GroundDetail.shader recomputes them from the world position and reads only the block's two values, and the furrows run along each parcel's own rows. They stop exactly at strip and block edges. Where one parcel ends and the next begins along a strip, the shader has no edge to stop at (the bake places those ends with its hash); the furrows fade there through furrows.png's 158 x 211 m texels, so they can reach about one texel into a fallow neighbour. test_the_block_table_rebuilds_the_bakes_rows checks the Python mirror of that formula against the bake at 200,000 points, to within 0.5 m. test_the_shader_holds_the_bakes_constants reads the shader's own copies (the block and strip size, the angle range, the pad, the wobble's numbers and the table's 16-bit decoding) and fails if they leave the bake's. As the scrub does, the furrows give way to the keep-out mask (_ScDetailKeepOutTex: the towns, the routes and the units' clearings), so no lines show through a unit token's clearing or along a route corridor.

Why in the ground detail pass. As with #2878, the terrain function sits at the frame-time cliff of issue #391. The furrows are a variant of GroundDetail.shader (SC_FURROWS) behind the furrows visual toggle, so the profile prices them alone. GroundDetailTests.BothShadersCompileForWebGL2 compiles that variant for GLES3 too.

Low tier. GroundDetail is not one of QualityFallback's steps, so the furrows have no fallback: they were priced on the desktop headless build only. If a mobile profile shows them costly, the furrows toggle is the hook a fallback step would turn off.

Symbolic size. At gameplay03 a pixel covers 35 to 46 m of ground, so real furrows (under 1 m apart) cannot show. The lines are 240 m apart, with a finer set at 80 m weighted 0.6, gain 0.30, and both fade out once a line is under about 4 pixels apart. Each is an odd wave around zero, so the ground keeps its brightness and colour; GradeTests' warmth margin (0.266 against a bar of 0.265) is unchanged.

Measured (EditMode, GroundDetailTests.TheFurrowsLineTheFieldsAndKeepTheirColour, on against off, kits and trees off). Mean gradient +3.2% on the coastal plain at 12 km, +3.2% at 31 km, +1.45% at gameplay03 (the frame is mostly towns, trees and hills); brightness and saturation within 0.0003; no change on the unfarmed Judean hills or at diorama02. Before the keep-out mask (the PR's first round) the same frames gave +4.6%, +4.7% and +2.3%: the difference is the furrows that lay inside town discs, clearings and route corridors, most of which the kits cover in play.

Frame time (headless WebGL ?vprofile, 1920x1080, DPR 1, machine_lock.py --exclusive --wait-unity). The full round of 10 November 2026: gameplay03 all 22.46 ms, -furrows 23.04 ms, so furrows cost -0.58 ms (spreads 0.59 and 0.94 ms); diorama02 all 23.33 ms, -furrows 23.20 ms, +0.13 ms (spreads 0.20 and 0.80 ms). An earlier round on 8 November gave +0.62 ms at gameplay03 (load average 7.1, spread 1.53 ms) and stopped partway through diorama02 on a res=1920x880 line. So the cost is within the noise, at most about 0.6 ms, under the 2.5 ms budget. With the keep-out sample added (review of #3070, 10 November 2026, a new build): gameplay03 all 23.20 ms, -furrows 23.28 ms, -0.08 ms (spreads 1.68 and 0.74 ms, load average about 5); diorama02 24.07 and 24.02 ms, +0.05 ms (spreads 0.94 and 0.76 ms). A round just before it, which started at load average 42, gave +0.64 ms at gameplay03 with spreads of 4.6 and 5.6 ms (noise) and +0.12 ms at diorama02 (spreads 0.13 and 0.10 ms).

Boot (?bootlog, screens ready, three headless boots each, alternated, exclusive window, 10 November 2026): 4997, 5033 and 5067 ms against 5069, 5054 and 5103 ms for the v0.3.9 release (gunzipped to load as this build does), medians 5033 and 5069 ms. v0.3.9 is this branch's base plus 16 commits, so it is a near neighbour, not an exact A/B.

Size. furrows.png adds 2.8 MB to the data file uncompressed (R8 with mips), about 0.33 MB gzipped; fieldblocks.png adds 11.6 KB. The build's data, wasm and framework files gzip to about 42.7 MB, so the nightly's 2% tolerance (about 0.85 MB) covers it without a budget line.

Colour (tools/art_compare.py, HUD masked, editor renders at 1920x1080 with furrows off and on, 10 November 2026). gameplay03 against promo 03: saturation 0.4350 both, contrast 0.1585 both, warmth 0.0945 and 0.0944, palette distance 0.3775 and 0.3773. diorama02: identical to four places (0.4249, 0.2237, 0.1680, 0.5220); sand and sea warmth pass their bands. The web build after: gameplay03 0.4220, 0.1765, 0.0884, 0.3050; diorama02 0.4203, 0.2341, 0.1707, 0.4923. The furrows add lines, not colour.