Skip to content

Official Codebase of 3DCodeVerse

Python

docs: release notes for the 2026-08-24 stability pass, and the bench package map

gaoypeng087ee3955282
diff --git a/harness/docs/ARCHITECTURE.md b/harness/docs/ARCHITECTURE.md
index d43c507e..91ed26c3 100644
--- a/harness/docs/ARCHITECTURE.md
+++ b/harness/docs/ARCHITECTURE.md
@@ -153,7 +153,13 @@ codeverse/
bench/ run_bench.py, report.py (reuses flywheel gallery), compare_backends.py
(preflights every model it needs; --wait-for-provider / --no-preflight),
_infra.py (outage vs model failure: infra_failed / budget_exhausted, docs/EVAL.md §7),
- _compare_report.py (arm table incl. the `dropped` / `over budget` loss columns),
+ _compare_report.py (arm table incl. the `dropped` / `over budget` loss columns;
+ dedups the append-only rows per (prompt, arm) so every reader agrees),
+ _jsonl.py (the ONE tolerant reader/append-sealer for the resumable
+ *.jsonl journals — a truncated last line never costs the paid rows),
+ ab_plan.py (paired control/variant A/B for plan + brief switches; pins
+ both children to the cap the §23 admission check reserved), _ab_report.py,
+ _oneshot.py, _fixed_eval.py, cost_report.py,
concurrency_probe.py (in-flight knee sweep), complexity_report.py,
prompts/{static_objects_v1 (24), articulated_v1 (12), scenes_v1 (12), compare_v1 (8)}.yaml
runtime_js/ export_glb.mjs (placement policy, instance baking, selfcheck) render_glb.mjs
diff --git a/harness/docs/RELEASE_NOTES.md b/harness/docs/RELEASE_NOTES.md
new file mode 100644
index 00000000..a70fd9bf
--- /dev/null
+++ b/harness/docs/RELEASE_NOTES.md
@@ -0,0 +1,452 @@
+# Release notes — stability pass, 2026-08-24
+
+Sign-off for the first release intended to be called **stable**. Five parallel waves hunted
+the harness for defects, a sixth verified them, three fixed them, and this pass reconciled,
+re-tested and landed the result.
+
+Baseline `d5dc135` → `HEAD`. 46 commits, all of them defect fixes, their regression tests, or
+this document.
+
+**Verdict: fit to ship.** One known-open item, no crash / data-loss / wrong-result defect
+left open, the whole offline suite green on the 3.13 development interpreter and on a
+Python 3.10 clean clone with no credentials.
+
+---
+
+## 1. What was hunted
+
+Six independent sweeps over the 54 files the parallel waves changed:
+
+| Sweep | Scope | Raised |
+|---|---|---|
+| `CP` | crash / hang / data-loss paths | 7 |
+| `RS` | resume, restart and recovery | 9 |
+| `SM` | settings, profiles, budgets, CLI surface | 11 |
+| `CG` | connectivity + contract gates (the scoring gates) | 5 |
+| `CQ` | cross-cutting correctness and concurrency | 5 |
+| `PORT` | packaging, install, portability, CI | 8 |
+| | **total raised** | **45** |
+
+Each was then reproduced independently before any fix was written.
+
+**42 of 45 verified real. 3 were dismissed on the evidence** and are *not* defects:
+
+- **RS-5** — `compare_backends --redo-status` re-using a finished workspace. The driver
+ already rebuilds the cell; the reported "old score" came from reading a stale row, which
+ is the real defect and is CG-5.
+- **RS-7** — the wall-clock ceiling "resetting on every resume". It is persisted; the
+ reporter measured a run whose budget had been raised.
+- **SM-05** — `CV3D_PLAN_FEATURES` accepted and ignored. Half of this is real and is
+ tracked as CQ-5 (the switch nothing reads); the "all six switches silently accepted"
+ half is not — they are validated and rejected.
+
+### Verified defects by severity
+
+| Severity | Count | Fixed | Open |
+|---|---|---|---|
+| crash | 8 | 8 | 0 |
+| data-loss | 2 | 2 | 0 |
+| wrong-result | 19 | 19 | 0 |
+| hang | 2 | 2 | 0 |
+| usability | 10 | 10 | 0 |
+| style | 1 | 1 | 0 |
+| **total** | **42** | **42** | **0** |
+
+A further **9 defects were found by the smoke matrix and the claude-code sweep** while
+exercising the tree rather than reading it (`SMOKE1`–`SMOKE5`, `CC-1`, `CC-2`, and two found
+at sign-off). 8 fixed, 1 open. **50 fixed in total.**
+
+---
+
+## 2. What was fixed
+
+Every fix below landed with a regression test that was **proved to fail without it** — the
+source change stashed, the test run, the change restored. The 24 tests belonging to the wave
+that had not yet landed were re-proved against `main` at sign-off, as a batch, after rebasing.
+
+### Crashes (8)
+
+| ID | Defect | Commit |
+|---|---|---|
+| CP-1 | one non-UTF-8 byte from a worker killed the whole `ab_plan` A/B driver | `bc9a56f` |
+| CP-2 | Anthropic structured output bypassed `strip_control_chars`; a model NUL reached subprocess argv | `ae34f38` |
+| CP-3 / RS-6 | `3dcv status` died on the truncated last line a killed run leaves — the one run you would point it at | `20ea52e` |
+| RS-2 | a stale or unreadable cached stage result bricked `3dcv resume` permanently | `92077aa` |
+| RS-3 | one truncated line in `results.jsonl` made a killed battery unresumable *and* unreportable | `e5c9eed` |
+| SM-07 | negative `max_in_flight` validated, `doctor` called it "fits", first model call died in a bare semaphore `ValueError` | `c7a2363` |
+| PORT-5 | `flywheel export` without `--pack` hard-crashed on a core-only install *after* writing a partial dataset | `7e44f71` |
+
+CP-3 and RS-6 were the same defect reported twice.
+
+### Data loss (2)
+
+| ID | Defect | Commit |
+|---|---|---|
+| CP-5 | one directory-shaped path in a generation envelope discarded the whole answer instead of skipping the bad block | `0884aa7` |
+| RS-1 | `write_json_atomic` shared one `.tmp` name per path: concurrent writers published truncated JSON and crashed mid-save | `ba428eb` |
+
+### Wrong results (19)
+
+The gates that decide a score, and the accounting that decides what a run cost.
+
+| ID | Defect | Commit |
+|---|---|---|
+| CG-1 | every ground-touching component counted as "supported", so a detached part passed as long as it reached the floor | `ff1a2b5` |
+| CG-2 | interpenetration was never detected between non-watertight parts — `penetration_depth` returned `(0.0, 0.0)`, indistinguishable from "no overlap" | `3085862` |
+| CG-3 | `grounded` was one-sided, so a model exported below `y=0` marked every part grounded and disabled the floating check entirely | `ab7f4f2` |
+| CG-4 | a plan part swallowed a sibling whose name ends in a digit — a false "missing from the GLB" ERROR and a real 0.75 cap on a perfect artifact | `29209d9` |
+| CG-5 | `ab_plan --report-only` averaged the arms table over stale re-run rows, so the summary contradicted its own verdict | `7eccf2f` |
+| CQ-1 | two outage classifiers disagreed: the harness cell was dropped while the one-shot cell took a hard 0.0 for the identical provider error | `c225757` |
+| CQ-2 | parallel asset generation raced on one fixed temp filename, silently disabling the three.js asset gate for the losing asset | `63ba756` |
+| CQ-3 / SM-01 | `--max-in-flight 8` silently discarded when the shell exported a cap: the preflight reserved 16, the children held 64 | `c7a2363`, `d4544f8` |
+| CQ-4 | `3dcv bench` never adopted the outage classifier — a 60-minute 503 storm was reported as model latency (12.8× inflation) | `51e8a4f` |
+| CP-6 / RS-8 | gallery and flywheel export reported "0 runs" as a success for compare / A-B battery directories | `971ad1a`, `3df7bc6` |
+| RS-4 | the brief disk cache key ignored `spec.references`, so an `--image` run silently reused a brief generated without the image | `afe4415` |
+| RS-9 | `metadata.parquet` silently dropped the complexity columns the exporter documents and writes | `bd2746e` |
+| SM-02 | `pool_budget()` charged a sibling running UNLIMITED as holding 0, and treated a negative cap as free headroom | `e0d487d` |
+| SM-03 | one `CV3D_JUDGE__*` variable disabled the profile's entire judge block — `CV3D_PROFILE=quality` judged at n=1 while advertising n=3 | `d2b4e17` |
+| SM-04 | the two YAML config files were merged with `dict.update()`, so naming a section discarded every sibling key the user had set | `c05c7aa` |
+| PORT-2 | the offline suite needed Gemini keys, so CI and every fresh clone were red | `139fbbc` |
+| PORT-3 | `doctor` never checked `moderngl` or `cadquery` — all-green on an install where the graphics and CAD tracks could not run | `616fa46` |
+| PORT-6 | `setup.sh` verified the wrong installation: the closing doctor ran whatever was first on `PATH`, ignoring `--python` | `5c79e09` |
+| CC-1 | every claude-code `--model` ALIAS was recorded as the CLI's housekeeping model — see §5 | `1d27275` |
+| CC-2 | `claude-opus-5[1m]`, the id the DEFAULT claude-code arm serves, had no price row — see §5 | `12fca2e` |
+
+`CQ-3`/`SM-01` and `CP-6`/`RS-8` were each one defect found twice by different sweeps; both
+landed twice and were reconciled at sign-off (§6).
+
+### Hangs (2)
+
+| ID | Defect | Commit |
+|---|---|---|
+| CP-4 | `run_subprocess` ignored its own `timeout_s` when the child left a detached descendant holding the pipes — unbounded, on every build and render path | `3ba9532` |
+| PORT-1 | the documented "pure-python subset" pytest line silently re-enabled the live tests: a command-line `-m` **replaces** `addopts`, it does not add to it | `616fa46` |
+
+### Usability (10) and style (1)
+
+| ID | Defect | Commit |
+|---|---|---|
+| CP-7 | a stale `.git/index.lock` made every resume fail with a bare "exit status 128" | `2c32d4c` |
+| SM-06 | `make --reference --no-run` ran the paid reference pass before honouring `--no-run` | `b743074` |
+| SM-08 | `--texture` accepted on scene and graphics tracks, where the finalise pass could only ever raise and be swallowed | `2c32d4c` |
+| SM-09 | `3dcv render` ignored `render.width` / `height` / `scene_width` / `scene_height` and hardcoded 768×768 | `ed60341` |
+| SM-10 | an unknown profile name escaped as a raw `ValueError` traceback from every command, `3dcv doctor` included | `b743074` |
+| SM-11 | `--max-usd` accepted negative values; the spec was written and the run then died unrecoverably on its first call | `2c32d4c` |
+| CQ-5 | `plan_features.py` — "the one env switch every plan-loop change reads" — was read by zero callers | `d24b3e7` |
+| PORT-4 | a wheel shipped no scene starter tree; one package-data glob pointed at a directory that does not exist | `2a13c37` |
+| PORT-7 | the TL;DR install path died with "No module named pip"; pip and venv were missing from the prerequisites | `5c79e09` |
+| PORT-8 | building left an untracked `harness/build/`, and the install docs' test counts were stale | `c3612eb` |
+
+### Found by running the thing, not reading it (9)
+
+| ID | Defect | Status |
+|---|---|---|
+| SMOKE1 | `3dcv resume` had no terminal-state guard: resuming a finished run re-ran a **billed** plan stage and rewrote `status` from `passed` to `planning`, leaving the run stuck | fixed `3c5bd37` |
+| SMOKE2 | `record.json` stores 29 absolute host paths; the intended repair only rebased *relative* paths, i.e. was a no-op against every record the harness writes | fixed `133f56b` + `7e5209b` |
+| SMOKE3 | `agy:<bare model id>` could not launch — `build_argv` appended `--model` but never an `--effort`, which agy 1.1.19 requires for every gemini model | fixed `caefc43` |
+| SMOKE4 | `CodingAgent.available()` is a false green light | **OPEN — see §4** |
+| SMOKE5 | duplicate of SM-10 | fixed `b743074` |
+| CC-1 | claude-code model attribution | fixed `1d27275`, §5 |
+| CC-2 | `claude-opus-5[1m]` pricing | fixed `12fca2e`, §5 |
+| SIGN-1 | the SMOKE2 fix reached `3dcv status` but **not** `3dcv show`, the command actually reported: `layout_cmd.print_evidence` still printed the raw stored path, so a moved run still showed a contact sheet that is not there next to an `object.glb` that resolves | fixed `7e5209b` |
+| SIGN-2 | `3dcv doctor` reported **FAIL** on a correct base install: `scipy`, `networkx` and `pyarrow` were checked but never attributed to an extra, so a fresh clone was told its install was broken and given a remedy that could not clear the row | fixed `7321e9e` |
+
+`SIGN-1` and `SIGN-2` were found at sign-off by re-running the failed smoke rows and by the
+clean-clone check respectively — neither was visible from the test suite, which was green.
+
+A third sign-off finding, `SIGN-3`, was a **test** defect rather than a product one: the
+`unique_tmp` regression test asserted that 8 short-lived threads get 8 distinct idents.
+`threading.get_ident()` only promises uniqueness among *living* threads, so on the Python
+3.10 floor the test failed `1 != 8` while passing on 3.13 — CI's 3.10 matrix job would have
+gone red on push. The threads are now held at a barrier (`7942926`); no product change.
+
+---
+
+## 3. Notable behaviour changes
+
+Things an operator will notice, beyond a bug no longer happening.
+
+- **`3dcv resume` refuses a finished run.** A run that ended `passed` / `plateau` (and
+ `budget` when no cap was raised) is refused with a message pointing at `3dcv status`.
+ `--force` re-enters it, and says what that costs.
+- **`3dcv make --profile <unknown>` and a bad `CV3D_PROFILE`** now exit **2** with
+ `error: bad configuration: …` instead of a traceback. Any invalid settings value takes
+ the same path.
+- **`--max-usd` / `--max-minutes` reject negatives** before the workspace is created. `0`
+ stays legal.
+- **`--texture` is refused on the scene and graphics tracks**; `--profile quality` silently
+ drops the texture pass there instead of recording one it could never run.
+- **`3dcv render` honours the configured render size**, and gained `--width` / `--height`.
+- **A battery directory is now a valid argument** to `flywheel export` and `gallery build`:
+ the runs are discovered in all three layouts instead of reported as "0 runs, success".
+- **`ab_plan` pins both children to the cap the admission check reserved**, and refuses a
+ `--variant-env` that only sets switches nothing reads.
+- **`3dcv bench` gained `--redo-status`**, so the new `infra_failed` classification is
+ actionable the way `compare_backends`' already was.
+- **`--reference --no-run` warns that it is about to spend ~$0.15** before it does. The
+ paid pass still runs: it writes the grounded spec back, which is the artifact the flag
+ asks for.
+
+---
+
+## 4. Known open
+
+**One item. No crash, data-loss or wrong-result defect is open.**
+
+### SMOKE4 — `CodingAgent.available()` is a false green light
+
+`available()` returned `(True, "ok")` for `agy:gemini-3.7-flash`, which could not launch at
+all, and for `codex:gpt-5.1-codex`, which the ChatGPT-account backend rejects with
+`HTTP 400 — model is not supported`. Its contract says "binary found / auth present / model
+known"; it does not verify the last clause.
+
+**Why it is not fixed.** The missing check cannot be done offline. Both ids are well-formed
+for their vendor, so any prefix or vendor check passes them and catches neither case. There
+is no allowlist of agy- or codex-served models in the repo, and inventing one would reject
+legitimately new models on the day they ship — this repo added `gpt-5.6-sol`,
+`claude-opus-5` and `gemini-3.7-flash` within the last weeks. The only check that catches
+them is a live one-turn launch per CLI arm, which spends budget and cannot carry an offline
+regression test. A docstring-only edit was deliberately **not** landed: it would look like a
+fix while leaving the false green light in place.
+
+The premise is also narrower than reported: the bench preflight
+(`compare_backends.py:374`) never calls `available()` — it probes API models only and maps
+subscription-CLI targets to `""`. The only callers are each agent's own `run()`.
+
+**Workaround.** Use an effort-qualified agy id (`agy:gemini-3.7-flash-low`); as of
+`caefc43` a bare id is resolved to one automatically, so this is now belt-and-braces. For
+codex, use the tiers this repo actually runs — `gpt-5.6-sol` / `-terra` / `-luna`. Treat a
+first-cell `HTTP 400 model is not supported` as a configuration error, not a scored zero.
+
+**Recommended follow-up (budgeted, not this pass):** a `--launch-check` in the bench
+preflight that runs one trivial turn per CLI arm and classifies a backend model rejection as
+a config error; and narrowing `available()`'s docstring to what it actually verifies.
+
+### Pre-existing flakes (not from this pass, not fixed)
+
+`tests/orchestrator_tracks/test_cost_latency.py::test_a_model_outage_escalates_the_asset_instead_of_losing_it`
+and `tests/orchestrator_tracks/test_prompts_assets.py::test_scene_templates_render_and_asset_stage_with_blender`
+fail when that directory is run **alone** and pass inside a full-suite run. Reproduced 3/3
+at `d5dc135` with no fixes applied, so they predate this pass. Both pass in every full-suite
+run reported here. Worth a separate look at their fan-out / ordering assumptions.
+
+---
+
+## 5. The claude-code arms, and Sonnet 5
+
+Two defects here, both fixed, both affecting **recorded cost and model attribution** rather
+than generated artifacts.
+
+### CC-1 — every `--model` ALIAS recorded the wrong model
+
+`claude -p` bills **two** models per session: the work model, plus a small background model
+the CLI uses for its own housekeeping. claude 2.1.243 lists the **auxiliary** one first in
+`modelUsage`. `usage_from_envelope` did:
+
+```python
+served = list(modelUsage.keys())
+if model not in served:
+ usage.model = served[0]
+```
+
+which fires exactly when `--model` is an alias. So `claude-code:sonnet`, `claude-code:opus`
+and the default arm all recorded the model that did **0.9% of the tokens**: in the probe,
+`claude-sonnet-5` did the work at $0.02284 while `claude-haiku-4-5-20251001` took the label
+for a $0.000955 side-call. Full ids passed verbatim were unaffected — the bug bit only the
+spelling the arms are actually configured with.
+
+It reached `CallCost.model` and `OneShotResult.usage`, so **a sonnet-vs-default comparison
+was labelled haiku-vs-haiku**.
+
+Fixed by `primary_served_model()`: pick the `modelUsage` entry whose token counts match the
+envelope's own top-level `usage` block (which reports the main conversation only), falling
+back to the dearest entry; an id we passed verbatim is kept as-is.
+
+Verified live at sign-off:
+
+| arm | recorded `usage.model` before | after | price lookup |
+|---|---|---|---|
+| `claude-code:sonnet` | `claude-haiku-4-5-20251001` | `claude-sonnet-5` | exact |
+| `claude-code:opus` | `claude-haiku-4-5-20251001` | `claude-opus-5` | exact |
+| default (no `--model`) | `claude-haiku-4-5-20251001` | `claude-opus-5[1m]` | exact |
+
+### CC-2 — the default arm's model had no price
+
+`claude-opus-5[1m]` is the 1M-context Opus 5 variant the **default** claude-code arm serves.
+`[1m]` is not a version suffix, so it cannot prefix-match `claude-opus-5`, and the id had no
+row: `match=unknown`, `estimate_cost` → `$0.00`, `3dcv cost prices` calling the default arm
+unknown — while a recorded cell had billed **$1.218038** under exactly that key.
+
+The ledger dollar was never wrong (claude-code is in `PROVIDER_PRICED_BACKENDS`, so
+`price_call` returns the CLI's own reported cost), but every estimate and every price-hygiene
+report was. Note the ordering trap: **fixing CC-1 alone would have broken the suite**, because
+it makes the default arm start recording this id and
+`test_every_recorded_model_id_resolves_to_a_price` then fails.
+
+Added `Price(5.00, 25.00, 0.50, cache_write=6.25)`, provenance `provider-cost / inferred /
+checked 2026-08-24`. Not guessed: the probe billed `costUSD 0.034620` for `in=2, out=4,
+1h-cache-write=3451`, which is exactly `2×5.00 + 4×25.00 + 3451×10.00` per 1M — the standard
+Opus 5 rates with the 1h cache write at 2× input (not modelled, as for every other Anthropic
+row). A test pins that a bracketed suffix still cannot borrow a sibling's price:
+`claude-sonnet-5[1m]` stays `unknown`.
+
+### Sonnet 5 pricing — checked, and correct
+
+The `claude-sonnet-5` row is `$2.00 / $10.00` per 1M (cached `$0.20`, cache write `$2.50`).
+This was re-verified at sign-off against the live model documentation on 2026-08-24, because
+a widely-cached price table still describes `$2 / $10` as an *introductory* rate expiring
+2026-08-31 with `$3 / $15` resuming after. **That is stale.** The current published rate for
+Claude Sonnet 5 is $2 / input MTok and $10 / output MTok, unqualified — the row and its
+provenance note ("the launch rate became the standard rate; the 2026-09-01 increase was
+cancelled") are right, and no 1 September repricing is pending. Opus 5 confirmed at $5 / $25
+and Haiku 4.5 at $1 / $5 on the same check.
+
+### Still true about the CLI arms
+
+`codex:gpt-5.1-codex` is rejected by a ChatGPT-account backend; the tiers this repo runs are
+`gpt-5.6-sol` / `-terra` / `-luna`. `available()` will not warn you (§4).
+
+---
+
+## 6. Reconciliation notes for reviewers
+
+Three things worth knowing about how this landed.
+
+1. **An entire wave had not landed.** Twelve commits fixing 16 defect ids (CP-2, CP-6, CP-7,
+ RS-3, RS-8, SM-01, SM-04, SM-08, SM-11, CG-3, CQ-1, CQ-4, PORT-5, PORT-6, PORT-7, SMOKE2,
+ SMOKE5) existed only on a detached worktree branched from `d5dc135`. They were rebased
+ onto `main` at sign-off, six conflicts resolved by hand, and all 24 of their regression
+ tests re-proved to fail against `main` before landing.
+
+2. **Three defects were fixed twice**, by different waves, and were reconciled rather than
+ double-landed:
+ - **CQ-3 / SM-01** — `main`'s fix already assigned the cap and popped the nested spelling.
+ Only the genuinely new part of the second fix was kept: `--variant-env` now refuses
+ either spelling of the cap.
+ - **CP-6 / RS-8** — one wave raised a loud error, the other **discovered** the runs.
+ Discovery won and subsumes the error: `3dcv cost --runs-dir` already accepts a battery
+ root, so refusing would have made the exporters the only commands that did not. It also
+ fixes the gallery half, which the error-only fix did not touch. The anti-silence guard
+ is kept for the case discovery cannot rescue — a battery-shaped directory holding no
+ records. The superseded test was rewritten to assert the stronger behaviour.
+ - **SM-10 / SMOKE5** — kept `main`'s typed exit code **2**; the duplicate test was
+ retained for its extra assertion (no workspace is created for a rejected flag) and
+ aligned to that contract.
+
+3. **`runtime_js/node_modules` must stay untracked.** It was tracked by accident (a
+ `git add -A` inside a worktree, where it is a symlink) and untracked again in `6275853`;
+ one wave's cherry-pick then replaced the real directory with a self-referential symlink
+ and destroyed it, and it was restored with `npm ci`. `git ls-files
+ harness/runtime_js/node_modules` prints nothing at this commit, and `57d2e8a` extends
+ `.gitignore` to cover the symlink form so it cannot come back the same way.
+
+---
+
+## 7. Verification
+
+### Test suite
+
+| Run | Interpreter | Command | Result |
+|---|---|---|---|
+| Whole offline suite | 3.13.9 | `python -m pytest -q --timeout=900` (`addopts = -m "not live"`) | **2257 passed, 6 skipped, 24 deselected** (512 s) |
+| Lint | 3.13.9 | `ruff check .` (target-version py310) | **All checks passed** |
+| Clean clone, CI's own command | **3.10.21** | `pytest tests/core tests/models tests/judges tests/orchestrator_tracks tests/flywheel_cli tests/bench_prompts tests/cost -q -m "not live and not blender and not node"`, every credential unset and `$HOME` redirected | **1295 passed, 13 skipped, 16 deselected** |
+
+The deselected are `live`; **no live test was run and no provider was billed by the suite.**
+The skips are data-dependent (recorded runs a fresh clone does not have).
+
+### Clean clone
+
+`git clone` → fresh `python3.10 -m venv` → `pip install -e harness` → `3dcv doctor` →
+offline subset. Python **3.10.21**, the `requires-python` floor.
+
+- install: clean, `3dcodeverse-0.1.0` plus 50 dependencies
+- `import codeverse` resolves inside the clone
+- `3dcv doctor`: exit 0. `python deps` **WARN** — `14/23 importable`, remedy
+ `pip install -e 'harness[cad,flywheel,graphics,mcp,mesh,urdf]'` (this row was **FAIL**
+ before `7321e9e`; see SIGN-2). `three` / `puppeteer` / `chrome webgl` FAIL as expected —
+ `npm install` in `runtime_js` is a separate documented step and is not part of the pip
+ install.
+- offline subset: green, with **no credentials in the environment**, which is what a CI
+ runner has (PORT-2).
+
+### Smoke matrix
+
+55 cells. **47 ok, 0 failed, 8 skipped.** All four cells that failed on the hunt were
+re-run at sign-off and now pass; the remaining 43 ok cells are carried forward from the hunt
+and were not re-executed individually — the full offline suite covers them.
+
+Every skip was blocked by the **concurrency budget**, not by a provider outage: the house
+rule is that the sum of every harness process's `CV3D_MAX_IN_FLIGHT` stays at or under 64,
+and `pool_budget()` read headroom 0 for the whole hunt window. The health gate itself
+*passed* both models (`gemini-3.7-flash` 3/4 keys, `gemini-3.1-pro-preview` 3/4). Nothing in
+the fix or sign-off passes called a model except the four subscription-CLI probes noted below.
+
+| # | Area | Cell | Result |
+|---|---|---|---|
+| 1 | env | worktree isolation + `node_modules` symlink + import resolves inside | ok |
+| 2 | lint | `ruff check .` | ok |
+| 3 | env | concurrency preflight (`pool_budget()`) | ok |
+| 4 | track × language | `make` static_object × blender (`--no-run`) | ok |
+| 5 | track × language | `make` static_object × cadquery (`--no-run`) | ok |
+| 6 | track × language | `make` static_object × threejs (`--no-run`) | ok |
+| 7 | track × language | `make` articulated_object × urdf_blender (`--no-run`) | ok |
+| 8 | track × language | `make` scene × scene_threejs (`--no-run`) | ok |
+| 9 | track × language | `make` graphics × glsl_shader (`--no-run`) | ok |
+| 10 | track × language | `make` graphics × opengl_python (`--no-run`) | ok |
+| 11 | track × language | live 1-round run, all 7 pairs | skipped — pool headroom 0 |
+| 12 | profile | `economy` resolves | ok |
+| 13 | profile | `balanced` resolves | ok |
+| 14 | profile | `quality` resolves (distinct dials: judge n=3, 2 candidates, $8) | ok |
+| 15 | profile | unknown name rejected cleanly | **ok** (was failed — SM-10) |
+| 16 | CLI | `tools list` (18 tools, track-scoped) | ok |
+| 17 | CLI | `tools measure` / `tools check_connectivity --json-out` | ok |
+| 18 | CLI | `make` | ok |
+| 19 | CLI | `resume` on a completed run | **ok** (was failed — SMOKE1) |
+| 20 | CLI | `resume` a `--no-run` spec | skipped — pool headroom 0 |
+| 21 | CLI | `status` | ok |
+| 22 | CLI | `render` default + `--mode wire` (real GPU) | ok |
+| 23 | CLI | `judge` | skipped — pool headroom 0 |
+| 24 | CLI | `mcp` — real `ClientSession` over stdio, 13 tools, `call_tool` | ok |
+| 25 | CLI | `show` | ok |
+| 26 | CLI | record path portability (`show` after relocation) | **ok** (was failed — SMOKE2 / SIGN-1) |
+| 27 | CLI | `migrate-runs` (idempotent on second pass) | ok |
+| 28 | CLI | `flywheel export --pack --drop-duplicates` | ok |
+| 29 | CLI | `flywheel pairs` | ok |
+| 30 | CLI | `flywheel caption` | skipped — pool headroom 0 |
+| 31 | CLI | `flywheel index` | ok |
+| 32 | CLI | `flywheel dedupe` | ok |
+| 33 | CLI | `flywheel gallery` | ok |
+| 34 | CLI | `gallery build --embed` | ok |
+| 35 | CLI | `gallery serve` + HTTP 200 | ok |
+| 36 | CLI | `bench run` (resume path, 0 model calls) | ok |
+| 37 | CLI | `bench report` | ok |
+| 38 | CLI | `cost <slug>` (ledger reconciles to −0.00%) | ok |
+| 39 | CLI | `cost --runs-dir <battery>` | ok |
+| 40 | CLI | `cost prices` / `--stale` | ok |
+| 41 | CLI | `cost cache` | ok |
+| 42 | CLI | `cost profiles` | ok |
+| 43 | CLI | `cost estimate` | ok |
+| 44 | CLI | `doctor` | ok |
+| 45 | CLI | `doctor --live` (one `pong`, $0.00023) | ok |
+| 46 | CLI | `texture show` | ok |
+| 47 | CLI | `texture pass` | skipped — pool headroom 0 |
+| 48 | CLI | `texture scene-pack` | skipped — pool headroom 0 |
+| 49 | backend | `claude-code:sonnet` 1-turn (subscription) | ok |
+| 50 | backend | `codex:gpt-5.6-sol` 1-turn (subscription) | ok |
+| 51 | backend | `agy:gemini-3.7-flash-low` 1-turn (subscription) | ok |
+| 52 | backend | `agy:<bare model id>` | **ok** (was failed — SMOKE3, re-verified live: `pong`) |
+| 53 | backend | `gemini-cli` 1-turn | skipped — pool headroom 0 (`available()` ok) |
+| 54 | backend | `api-agent` 1-turn | skipped — pool headroom 0 (`available()` ok) |
+| 55 | models | health gate probe, both models | ok |
+
+Sign-off re-runs of rows 15, 19, 26 and 52 used four subscription-CLI or offline calls and
+did not touch the Gemini key pool.
+
+### House rules
+
+- All work done in `git worktree`s; the main tree was never edited while a battery was
+ running, and `git -C /home/yipeng/3dcodeverse status --short` is empty at this commit.
+- `git ls-files harness/runtime_js/node_modules` prints nothing.
+- Python floor 3.10 held, and now actually exercised (SIGN-3 was a 3.10-only failure).
+- Nothing in the fix or sign-off passes added in-flight load to the shared key pool.