Skip to content

Official Codebase of 3DCodeVerse

Python

repo: one copy of the aurora reference photos (was four), docs/skills-attic removed, a test file named for what it tests

gaoypengdae6d284e85f
diff --git a/harness/bench/prompts/refs_v1_graphics.yaml b/harness/bench/prompts/refs_v1_graphics.yaml
index 98e7efae..c56cb2fd 100644
--- a/harness/bench/prompts/refs_v1_graphics.yaml
+++ b/harness/bench/prompts/refs_v1_graphics.yaml
@@ -5,6 +5,9 @@ description: Reference photos vs none on the same brief (aurora), 3 seeds per ar
the same LikenessJudge.
prompts:
- id: aurora_ref_1
+ references:
+ - ../refs/tsr_gfx_aurora_ridge/aurora_green_spiral_sea.png
+ - ../refs/tsr_gfx_aurora_ridge/aurora_purple_green_lake.png
prompt: 'an aurora over a snowy ridge: layered green and violet curtains with vertical ray structure rippling
across the sky, a dense star field behind them, a dark mountain-ridge silhouette along the bottom third and
a frozen lake reflecting the curtains with a soft shimmer
@@ -26,6 +29,9 @@ prompts:
- ref
- seed1
- id: aurora_ref_2
+ references:
+ - ../refs/tsr_gfx_aurora_ridge/aurora_green_spiral_sea.png
+ - ../refs/tsr_gfx_aurora_ridge/aurora_purple_green_lake.png
prompt: 'an aurora over a snowy ridge: layered green and violet curtains with vertical ray structure rippling
across the sky, a dense star field behind them, a dark mountain-ridge silhouette along the bottom third and
a frozen lake reflecting the curtains with a soft shimmer
@@ -47,6 +53,9 @@ prompts:
- ref
- seed2
- id: aurora_ref_3
+ references:
+ - ../refs/tsr_gfx_aurora_ridge/aurora_green_spiral_sea.png
+ - ../refs/tsr_gfx_aurora_ridge/aurora_purple_green_lake.png
prompt: 'an aurora over a snowy ridge: layered green and violet curtains with vertical ray structure rippling
across the sky, a dense star field behind them, a dark mountain-ridge silhouette along the bottom third and
a frozen lake reflecting the curtains with a soft shimmer
diff --git a/harness/bench/prompts/refs_v2_graphics.yaml b/harness/bench/prompts/refs_v2_graphics.yaml
index 5fe97833..33517a19 100644
--- a/harness/bench/prompts/refs_v2_graphics.yaml
+++ b/harness/bench/prompts/refs_v2_graphics.yaml
@@ -5,6 +5,9 @@ description: 'Second turn of the graphics loop: the same aurora brief, with/with
arm, judged in-loop by shader_v2 with the light-phenomena cookbook (aeb49c0) and brief-selected chapters (137541e).'
prompts:
- id: aurora_ref_1
+ references:
+ - ../refs/tsr_gfx_aurora_ridge/aurora_green_spiral_sea.png
+ - ../refs/tsr_gfx_aurora_ridge/aurora_purple_green_lake.png
prompt: 'an aurora over a snowy ridge: layered green and violet curtains with vertical ray structure rippling
across the sky, a dense star field behind them, a dark mountain-ridge silhouette along the bottom third and
a frozen lake reflecting the curtains with a soft shimmer
@@ -26,6 +29,9 @@ prompts:
- ref
- seed1
- id: aurora_ref_2
+ references:
+ - ../refs/tsr_gfx_aurora_ridge/aurora_green_spiral_sea.png
+ - ../refs/tsr_gfx_aurora_ridge/aurora_purple_green_lake.png
prompt: 'an aurora over a snowy ridge: layered green and violet curtains with vertical ray structure rippling
across the sky, a dense star field behind them, a dark mountain-ridge silhouette along the bottom third and
a frozen lake reflecting the curtains with a soft shimmer
@@ -47,6 +53,9 @@ prompts:
- ref
- seed2
- id: aurora_ref_3
+ references:
+ - ../refs/tsr_gfx_aurora_ridge/aurora_green_spiral_sea.png
+ - ../refs/tsr_gfx_aurora_ridge/aurora_purple_green_lake.png
prompt: 'an aurora over a snowy ridge: layered green and violet curtains with vertical ray structure rippling
across the sky, a dense star field behind them, a dark mountain-ridge silhouette along the bottom third and
a frozen lake reflecting the curtains with a soft shimmer
diff --git a/harness/bench/refs/README.md b/harness/bench/refs/README.md
index dfe0b26c..3a1bde6d 100644
--- a/harness/bench/refs/README.md
+++ b/harness/bench/refs/README.md
@@ -20,4 +20,5 @@ note the agent and the judge read. A prompt may also list `references:` paths i
So "with references" vs "without" is the same battery with the folder present or absent.
Keep photos you have the right to use; name the file by what it shows. First folder:
-`tsr_gfx_aurora_ridge/` (two aurora photographs, 2026-08-26).
+`tsr_gfx_aurora_ridge/` (two aurora photographs, 2026-08-26). A battery whose prompt ids differ from the folder
+name lists the files under `references:` instead of copying the folder (`prompts/refs_v{1,2}_graphics.yaml`).
diff --git a/harness/bench/refs/aurora_ref_1/aurora_green_spiral_sea.png b/harness/bench/refs/aurora_ref_1/aurora_green_spiral_sea.png
deleted file mode 100644
index 0af725b8..00000000
Binary files a/harness/bench/refs/aurora_ref_1/aurora_green_spiral_sea.png and /dev/null differ
diff --git a/harness/bench/refs/aurora_ref_1/aurora_purple_green_lake.png b/harness/bench/refs/aurora_ref_1/aurora_purple_green_lake.png
deleted file mode 100644
index 8075b10a..00000000
Binary files a/harness/bench/refs/aurora_ref_1/aurora_purple_green_lake.png and /dev/null differ
diff --git a/harness/bench/refs/aurora_ref_2/aurora_green_spiral_sea.png b/harness/bench/refs/aurora_ref_2/aurora_green_spiral_sea.png
deleted file mode 100644
index 0af725b8..00000000
Binary files a/harness/bench/refs/aurora_ref_2/aurora_green_spiral_sea.png and /dev/null differ
diff --git a/harness/bench/refs/aurora_ref_2/aurora_purple_green_lake.png b/harness/bench/refs/aurora_ref_2/aurora_purple_green_lake.png
deleted file mode 100644
index 8075b10a..00000000
Binary files a/harness/bench/refs/aurora_ref_2/aurora_purple_green_lake.png and /dev/null differ
diff --git a/harness/bench/refs/aurora_ref_3/aurora_green_spiral_sea.png b/harness/bench/refs/aurora_ref_3/aurora_green_spiral_sea.png
deleted file mode 100644
index 0af725b8..00000000
Binary files a/harness/bench/refs/aurora_ref_3/aurora_green_spiral_sea.png and /dev/null differ
diff --git a/harness/bench/refs/aurora_ref_3/aurora_purple_green_lake.png b/harness/bench/refs/aurora_ref_3/aurora_purple_green_lake.png
deleted file mode 100644
index 8075b10a..00000000
Binary files a/harness/bench/refs/aurora_ref_3/aurora_purple_green_lake.png and /dev/null differ
diff --git a/harness/codeverse/skills/registry.py b/harness/codeverse/skills/registry.py
index 6af25ba5..0a315d05 100644
--- a/harness/codeverse/skills/registry.py
+++ b/harness/codeverse/skills/registry.py
@@ -177,7 +177,7 @@ ROUTES: tuple[Route, ...] = (
Route("R4", "cv3d-bbox-contract", 95, kinds=("repair", "refine", "rebuild"), findings=("contract/*",),
why="the previous round's contract gate fired"),
# R5 was cv3d-form-manifest, retired 2026-08-25: read 2/19 (11%, CI [3%,31%]) before and
- # after its description was revised. See docs/skills-attic/cv3d-form-manifest/. The id is
+ # after its description was revised. See docs/SKILLS_LEDGER.md (the bundle's text is in git history). The id is
# not reused — a route id is how a ledger row and a run record refer to a routing decision.
Route("R6", "cv3d-repeats-and-mirrors", 55, tracks=OBJECT_TRACKS,
kinds=("baseline", "part", "detail", "refine", "rebuild"),
diff --git a/harness/docs/SKILLS_LEDGER.md b/harness/docs/SKILLS_LEDGER.md
index 59faf85a..5917e64c 100644
--- a/harness/docs/SKILLS_LEDGER.md
+++ b/harness/docs/SKILLS_LEDGER.md
@@ -75,7 +75,8 @@ own target metrics. The results, stated plainly:
A/B came back `underpowered`, and for several the underpowering is structural rather than
a matter of n. **So nothing was promoted to `measured`, and nothing ships on.**
* **One bundle was cut**: `cv3d-form-manifest`, read 2 of 19 times before *and* after its
- description was rewritten. Text and numbers in `docs/skills-attic/`.
+ description was rewritten. The numbers are in this ledger; the bundle's text is in git history
+ (`docs/skills-attic/`, removed 2026-09-21).
* **Four bundles were revised by fixing their delivery, not their prose.** The scene track
never handed its skills to the sessions that do the work. That is now fixed (§3).
* **`cadquery-forms` and `threejs-forms` stay `inherited-unverified`.** Their languages have
@@ -156,7 +157,7 @@ so every number below carries the snapshot it was read from.
| **cv3d-cadquery-forms** | inherited-unverified | `build_failure_rate` | down | 0.00, n=4 (all 4 built) | same | **4** | no variance to move | **routed OFF.** Needs 20 graded cadquery runs; has **4**. Read 2/2 when forced on |
| **cv3d-threejs-forms** | inherited-unverified | `missing_parts` | down | 0.00, n=3 | same | **3** | no variance to move | **routed OFF.** Needs 20 graded threejs runs; has 3. Read 2/2 when forced on |
-**Cut this wave:** `cv3d-form-manifest` → `docs/skills-attic/cv3d-form-manifest/`.
+**Cut this wave:** `cv3d-form-manifest` (text in git history; `docs/skills-attic/` was removed 2026-09-21).
### Read rate — the cheap filter, and the only thing that resolved
diff --git a/harness/docs/skills-attic/README.md b/harness/docs/skills-attic/README.md
deleted file mode 100644
index 674693b0..00000000
--- a/harness/docs/skills-attic/README.md
+++ /dev/null
@@ -1,18 +0,0 @@
-# skills attic
-
-Bundles the library **cut**, kept whole with the numbers that cut them.
-
-A bundle is retired here, not deleted, because the prose is usually fine — what failed was
-the bundle's claim on a session's attention, or its claim on a metric. If the delivery
-changes (a new read channel, a new gate, a language that finally has runs), the text is
-still here to route back in, and the numbers below say exactly what would have to be
-different for that to be worth doing.
-
-Nothing in this directory is loaded. `codeverse.skills.all_skills()` reads
-`codeverse/skills/*/SKILL.md` only, so an attic bundle costs no index line and no tokens.
-`3dcode skills validate` does not see it either — an attic bundle is a document, not a
-maintained artefact, and it is allowed to go stale.
-
-| bundle | cut | why | what would bring it back |
-|---|---|---|---|
-| `cv3d-form-manifest` | 2026-08-25 | read 2 of 19 times (11%, CI [3%, 31%]) — unread before AND after its description was rewritten to lead with the trigger. The only bundle whose read-rate interval excludes every other's. | a read channel that makes pre-geometry planning cheap for the agent to accept, or evidence that `feature_density` follows the judged `geometry_detail` more strongly than r=+0.19 |
diff --git a/harness/docs/skills-attic/cv3d-form-manifest/SKILL.md b/harness/docs/skills-attic/cv3d-form-manifest/SKILL.md
deleted file mode 100644
index f3521133..00000000
--- a/harness/docs/skills-attic/cv3d-form-manifest/SKILL.md
+++ /dev/null
@@ -1,214 +0,0 @@
-<!--
-RETIRED 2026-08-25 (curate wave). Not loaded; kept for the record.
-
-WHY IT WAS CUT — read rate, not prose quality.
- api-agent `read_skill` ground truth, both arms of the description A/B:
- baseline 1 open / 10 listings
- revised 1 open / 9 listings (first sentence rewritten to lead with the trigger)
- pooled 2 opens / 19 listings = 11%, 95% CI [3%, 31%]
- That interval excludes every other bundle's. The next-lowest, cv3d-repeats-and-mirrors,
- pooled 3/8 = 38%; the most-listed bundle, cv3d-bbox-contract, pooled 13/19 = 68%.
- It was listed in 19 of the 27 sessions that were actually offered the read tool — more
- than any bundle except bbox-contract — and cost ~1,917 body tokens plus an index line
- every one of those times.
-
- The standing rule is: a bundle that is unread AFTER a revision is cut. It was revised
- once, deliberately, on the cut that api-agent actually sees (`prompting.index_summary`
- takes the first sentence; `api_skills.SPEC_SUMMARY_CHARS` takes the first 150 chars),
- and it did not move.
-
- Plausible mechanism, recorded so the next attempt does not repeat it: the bundle asks
- the agent to do planning work BEFORE the geometry it was just told to write. That is
- the one instruction an eager coder skips, and no wording makes it cheaper to accept.
-
-WHAT ITS TARGET METRIC SAID — a second, independent reason.
- target: feature_density (spatial/complexity, recomputed from object.glb), direction UP
- corpus baseline: 2561.7 median / 8564.4 mean, n=175 graded object runs (2026-08-25)
- A/A floor on identical arms: delta +3904.7 (2095.7 -> 6000.5, a 2.9x swing),
- paired sd 3057.6, 2 SE +/-3530.6 — a noise band 1.4x the corpus median it measures.
- the one skills-on datapoint moved it the WRONG WAY: -1001.8 (direction is up).
- n needed to resolve a 25% move: ~92 paired prompts.
- And the metric is only a LEADING indicator: r = +0.19 against the judged
- geometry_detail (docs/COMPLEXITY.md 2.2, n=44). It was on ledger probation for that.
-
-WHAT WOULD BRING IT BACK.
- Either a read channel that makes pre-geometry planning cheap for an agent to accept, or
- evidence that feature_density tracks geometry_detail well enough to be worth 1,917
- tokens a session. Neither is a rewrite of the text below, which is why the text below
- is unchanged.
-
-Its FORM vocabulary outlived it: cv3d-blender-forms and cv3d-threejs-forms still key
-their technique tables on the same words, and those tables are self-contained.
--->
-
----
-name: cv3d-form-manifest
-description: "Use before calling any geometry API on a baseline, part, detail or rebuild session: name every shape, its dimensions and placement in FORM words. For the static_object and articulated_object tracks - whenever you are about to write geometry for a part you have not built yet. Turns the ENGINEERING BRIEF and the plan's part rows into a written element list the language skill's table then implements one line at a time, each with one countable refinement. Skip it on a repair round: a repair is not re-planning."
-license: Apache-2.0
-compatibility: "static_object and articulated_object tracks, all four object languages (blender, urdf_blender, cadquery, threejs). Needs the plan and, when the run enabled it, the ENGINEERING BRIEF block in your prompt."
-metadata:
- evidence: measured
- verified: "2026-08-25"
- corpus: "bench/out/*/runs/* graded records, recomputed 2026-08-25"
- pairs_with: "cv3d-blender-forms, cv3d-cadquery-forms, cv3d-threejs-forms, cv3d-part-contact"
- target_metric: "feature_density"
- target_direction: "up"
- target_unit: "small sharp edges per unit surface (spatial/complexity)"
- target_measurable: "true"
- target_baseline: "2551 median / 7946 mean; n=164 (bench/out, last gated round, 2026-08-25); right-skewed, read the median"
----
-
-# Form manifest — name the shapes before you name an API
-
-## When to use
-
-You are about to write geometry for a part you have not built yet. Spend three
-minutes writing what each element **is** as a shape. Then implement it.
-On a repair round, skip this and fix the finding.
-
-## Why this pays (measured on this harness)
-
-`geometry_detail` is the second-lowest judge criterion in both object tracks —
-**0.622** over 47 graded blender runs and **0.699** over 23 graded urdf runs
-(`bench/out/*/runs/*`, recomputed 2026-08-25). It does not improve by adding parts.
-Over 44 static runs (`docs/COMPLEXITY.md` §2.2, 2026-08-24):
-
-| what you add | correlation with `geometry_detail` | correlation with gate ERRORs |
-|---|--:|--:|
-| more plan parts | **−0.25** | −0.13 |
-| more built top-level parts | +0.01 | **+0.55** |
-| more triangles | −0.05 | +0.06 |
-| more `feature_density` (small sharp edges per unit surface) | **+0.19** | +0.03 |
-
-One axis out of four moves the criterion, and it is the one that measures modelled
-features, not part count. Adding parts instead is actively expensive: when one
-recorded toaster was rebuilt with more top-level parts its interpenetrating pairs
-went 4 → 8, max depth 3.5 mm → 9.0 mm, and overall 0.953 → 0.658
-(`docs/COMPLEXITY.md` §6).
-
-So: **fidelity, not count.** A drawer is five panels because a real one has five
-panels. Invented rivets bolted on as new parts are how floaters and
-interpenetration get made.
-
-## You are not starting from nothing
-
-Read these three before writing a word of the manifest — they are already in your
-prompt or your workspace:
-
-| source | what it hands you |
-|---|---|
-| the ENGINEERING BRIEF block | the real reference instance, real dimensions in metres, the sub-assemblies a fitter would name, `visible from outside`, `inside, NOT visible (do not model)`, `it does NOT have`, and 3–6 SIGNATURE FEATURES |
-| your part's row in the plan | `name` (PascalCase), `role`, `description`, `bbox`, `material`, `attach_to`, `symmetry`, `instances`, `children`, `detail_hint` |
-| the DETAIL BUDGET block | the triangle target for this object and per built object |
-
-Every SIGNATURE FEATURE must end up as an element in some part's manifest.
-Nothing on the `hidden inside` or `does NOT have` lists may appear at all.
-
-## Write it to `src/design/manifest_<snake>.md`
-
-`src/` and `public/` are the only directories you may create files in, and the
-language gates only read `.py` / `.js` / `.urdf`, so a Markdown note under
-`src/design/` is inert. One element per line, in this order:
-
-```
-Element — FORM — size (m) — placement — refinement — source
-```
-
-* **size** taken from the plan's bbox or the brief's dimensions, in the units
- the contract states; never a round number you liked the look of.
-* **placement** relative to your part's own bbox or to a named neighbour part
- ("top face of `Housing`", "20 mm in from the +X wall"), never an absolute
- guess.
-* **source** — which signature feature, `detail_hint` phrase or `children` row
- this element discharges. An element with no source is an element to cut.
-
-Stop when every signature feature and every `children` row is discharged. Six to
-twelve lines is a normal part. If your list is longer than the part's `children`
-list plus its `detail_hint`, you are padding.
-
-## The FORM words
-
-Language-neutral on purpose: the same word keys the table in
-`cv3d-blender-forms`, `cv3d-cadquery-forms` and `cv3d-threejs-forms`, so the
-manifest survives a language change and you look up exactly one row.
-
-| form | it means |
-|---|---|
-| revolved profile | a 2-D outline spun about an axis — bottles, knobs, feet, turned legs |
-| extruded outline | a closed 2-D outline given thickness — plates, brackets, panels |
-| tube swept along a path | a circular section carried along a curve — cables, handles, rails |
-| slab | a rectangular block whose thickness is much smaller than its face |
-| shaft | a long constant section — a rod, a bar, a post |
-| tapered solid | a section that changes size along its length — a draft, a cone, a wedge |
-| ring or band | an annulus or a torus — rims, ferrules, gaskets, trim |
-| carved cavity | a recess, pocket, slot, hole or notch taken out of a mass |
-| shell with wall thickness | a hollow body with a visible wall — a cup, a housing, a bin |
-| mirror pair | one element and its reflection across a plane |
-| radial array | N copies about an axis, angle derived from N |
-| layered or inset | a face broken into a proud or sunk sub-panel |
-| faceted solid | a flat-sided body no primitive covers — a wedge, a prism, a hull |
-| lofted transition | one section blending into a different section — a horn, a fairing |
-
-Write the form word, not the API call. Picking the API is the language skill's
-job and it changes per language; picking the shape is yours and it does not.
-
-## Every element carries at least one refinement kind
-
-`geometry_detail` is scored by counting **refinement kinds** visible in the
-renders — this is the rubric's own countable unit
-(`codeverse/judges/rubrics/static_object_v1.yaml`), not a heuristic:
-
-1. bevel / chamfer / fillet on an edge
-2. taper or draft along a part
-3. curve, sweep or compound surface where a primitive would be flat
-4. moulding or profile — a shaped edge, a cornice, a bead
-5. cut-out, hole, slot or notch
-6. visible wall thickness or hollow section — a rim, a lip, an open tube end
-7. surface relief — flutes, ribs, grooves, panelling, treads, weave, bark
-8. separately modelled hardware — hinge, knob, screw, bracket, strap, spoke
-
-Colour, material and shading are **not** refinement kinds and never raise this
-criterion. The anchors the judge actually uses:
-
-* 3–4 kinds on at least half the nameable parts → **0.85**
-* 5+ kinds with at least two thirds of parts carrying one, finest features at
- real hardware scale (a few mm) → **1.0**
-* many parts but no refinement on any of them → **0.25–0.4**
-
-Two or three kinds spread across the object is the difference between 0.55 and
-0.85. Pick them at manifest time so they are budgeted, not bolted on later.
-
-## Where a new element goes
-
-An element is a **new top-level part** only if a human would unbolt it as a
-separate object AND the plan already lists it. Otherwise it goes *inside* the
-part you are building.
-
-The plan's `children` are a planning device, not export nodes: the parent stays
-exactly **one** named object in the export, the contract gate measures the
-parent's bbox, and the connectivity gate loads one mesh per top-level node — so
-shapes inside a part are never pair-tested against each other, while still
-counting for `geometry_detail`. That is the cheapest detail on this harness.
-
-One caveat, so this does not become a licence to scatter: a sub-shape that
-floats free inside its parent is still caught, as a tiny-island warning on that
-part. Every element in your manifest has to touch something. Contact between
-parts, and its numbers, belong to `cv3d-part-contact`.
-
-## Before you write code
-
-1. Every signature feature appears in some manifest line.
-2. Nothing from `inside, NOT visible` or `does NOT have` appears in any line.
-3. At least three distinct refinement kinds across the object, and no element in
- your part is an unmodified primitive.
-4. Every line carries a real dimension and a placement anchored to a bbox or a
- named neighbour.
-
-Then implement the manifest top to bottom, one line at a time, looking up each
-FORM word in your language's table.
-
-## Depth
-
-`references/worked_manifest.md` — a full worked manifest for one part of a
-hand-crank coffee grinder, from brief fields to element lines, plus the two
-failure shapes (the padded manifest and the manifest with no refinements).
diff --git a/harness/docs/skills-attic/cv3d-form-manifest/references/worked_manifest.md b/harness/docs/skills-attic/cv3d-form-manifest/references/worked_manifest.md
deleted file mode 100644
index 03037cf9..00000000
--- a/harness/docs/skills-attic/cv3d-form-manifest/references/worked_manifest.md
+++ /dev/null
@@ -1,98 +0,0 @@
-# A manifest, worked end to end
-
-The session owns one part of a hand-crank coffee grinder: `BurrHousing`.
-Nothing below is invented — every line is traced to a field the harness already
-put in the prompt.
-
-## 1. What the prompt gave us
-
-From the ENGINEERING BRIEF block:
-
-```
-- reference instance: Hario Skerton Pro, ceramic conical burr, glass jar
-- real-world dimensions (m): body height 0.215, jar diameter 0.092,
- crank radius 0.075, burr diameter 0.038, lid diameter 0.058
-- sub-assemblies: Crank (handle, shaft, wing nut); BurrHousing (upper burr seat,
- ceramic cone, adjustment ring); Hopper (funnel, lid); Jar (glass body, thread)
-- visible from outside: the wing nut, the adjustment ring's knurl, the seam
- between hopper and jar, the crank shaft entering the lid, the pour spout
-- inside, NOT visible: the spring under the adjustment nut, the shaft bearing
-- it does NOT have: a motor, a power cord, a hopper window
-- SIGNATURE FEATURES: 1. the offset crank arm with a wooden knob;
- 2. the knurled adjustment ring under the burr; 3. the glass jar with a
- moulded shoulder; 4. the ceramic conical burr visible from above
-```
-
-From the plan's part row:
-
-```
-name: BurrHousing bbox: 0.062 x 0.062 x 0.058 m, centre (0, 0, 0.168)
-role: "carries the upper burr and the grind-size adjustment"
-symmetry: radial instances: 1
-children: [UpperBurrSeat, CeramicCone, AdjustmentRing]
-detail_hint: "carries the visible mechanism — model the burr teeth, the knurl
- and the seam where the hopper drops in"
-```
-
-Signature features 2 and 4 land here. 1 and 3 belong to `Crank` and `Jar`.
-
-## 2. `src/design/manifest_burr_housing.md`
-
-```
-HousingBody — shell with wall thickness — outer d 0.062, wall 0.003, h 0.034,
- base at z 0.151 — coaxial with Z, its top rim is where Hopper seats —
- visible wall thickness (kind 6) at the open top rim — children/UpperBurrSeat
-CeramicCone — tapered solid — d 0.038 at base tapering to 0.006 at the tip,
- h 0.026, base at z 0.155 — coaxial, tip up, inside HousingBody —
- taper (kind 2) — signature feature 4
-ConeTeeth — surface relief — 14 ribs 0.0012 proud, 0.0015 wide, running up the
- cone — radial array of 14 on the cone flank — surface relief (kind 7) —
- detail_hint "burr teeth"
-AdjustmentRing — ring or band — outer d 0.058, inner d 0.048, h 0.008,
- base at z 0.143 — coaxial, immediately under HousingBody — chamfer 0.0008
- on both outer edges (kind 1) — signature feature 2
-RingKnurl — surface relief — 36 flutes 0.0006 deep, full ring height —
- radial array of 36 on AdjustmentRing's outer face — surface relief (kind 7)
- — signature feature 2 "knurled"
-SeatLip — layered or inset — a 0.002 step sunk 0.0015 into the top rim of
- HousingBody, full circumference — the hopper drops into it —
- cut-out/notch (kind 5) — detail_hint "the seam where the hopper drops in"
-```
-
-Six lines. Four refinement kinds (2, 5, 6, 7) plus the chamfer (1) — five kinds
-in one part, which is the 1.0 anchor's requirement for the whole object.
-`ConeTeeth`, `RingKnurl` and `SeatLip` are **inside** `BurrHousing`; the export
-still has exactly one node called `BurrHousing`, so the connectivity gate sees
-one part and the penetration test never runs between the cone and its own teeth.
-
-Nothing models the spring or the bearing: they are on the `NOT visible` list.
-Nothing models a hopper window: it is on the `does NOT have` list.
-
-## 3. The two ways this goes wrong
-
-**Padded.** The manifest grows to nineteen lines because "a real grinder has
-screws". Eight of them become new top-level parts so they get names in the
-render. The measured consequence on this harness is not neutral: over 44 static
-runs part count correlates **+0.55** with gate ERRORs and **+0.01** with
-`geometry_detail` (`docs/COMPLEXITY.md` §2.2). The recorded toaster rebuild that
-did exactly this doubled its interpenetrating pairs (4 → 8), nearly tripled the
-depth (3.5 mm → 9.0 mm) and lost 0.30 of overall score (§6). If a screw matters,
-it is a refinement of kind 8 *inside* the part it fastens, sunk into the surface
-so it touches.
-
-**Refinement-free.** Six lines, correct dimensions, every element an
-unmodified cylinder or box. The rubric anchor for that is explicit: "Every part
-is an unmodified primitive; the only variation between parts is size, position
-and colour" → **0.25**. Correct dimensions do not buy detail. Each line needs a
-kind before you start typing geometry, because a bevel decided at manifest time
-costs one extra argument and a bevel decided after the build costs a rewrite.
-
-## 4. Reading it back
-
-Before you call the part done, read your manifest beside the render sheet and
-point at each line in the picture. A line you cannot point at was either never
-built or is too small to see. The harness renders objects at 768 x 768
-(`runtime_js/render_glb.mjs`), so on a 0.2 m object that fills the frame one
-millimetre is roughly three pixels and half a millimetre is not a feature at
-all. That is why the 1.0 anchor says "a few mm — screw heads, hinge leaves,
-bead mouldings": grow the line to hardware scale or delete it.
diff --git a/harness/tests/orchestrator_tracks/test_planner_geometry_reask.py b/harness/tests/orchestrator_tracks/test_planner_reask.py
similarity index 100%
rename from harness/tests/orchestrator_tracks/test_planner_geometry_reask.py
rename to harness/tests/orchestrator_tracks/test_planner_reask.py
diff --git a/harness/tests/skills/test_registry.py b/harness/tests/skills/test_registry.py
index 714ad00d..3bd76fb4 100644
--- a/harness/tests/skills/test_registry.py
+++ b/harness/tests/skills/test_registry.py
@@ -53,7 +53,7 @@ def test_finding_kinds_accepts_reports_findings_and_strings():
# --------------------------------------------------------------------------- the table
#: Retired ids remain reserved because run records refer to routing decisions by id.
-RETIRED_RULES = {"R5": "cv3d-form-manifest, cut 2026-08-25 (read 2/19); docs/skills-attic/"}
+RETIRED_RULES = {"R5": "cv3d-form-manifest, cut 2026-08-25 (read 2/19); docs/SKILLS_LEDGER.md"}
def test_rule_ids_are_unique_and_the_design_numbers_are_all_present():