dev: STATE.md + log — v0.29.2 released, §4a gate completed, @v0 = 5f74d74; round-4 arm

.dev/STATE.md

@@ -7,11 +7,22 @@ Roadmap detail lives in the work-plan tracker **#257**, not here (PLAN.md predat
 
 ## In flight
 
-- **v0.29.2 release cut 2026-09-21** (payload #317 `62504c1`, on top of #315 `cd41558`) — patch:
-  the scope sentence on the `ml` exercise rule that v0.29.1's failed gate called for, carrying
-  with it the editor's round-3 answers that v0.29.1 never delivered. §4a gate status is
-  recorded on the release PR; scenario 17 on the `.ml` lane is the one to read first. After the
-  floating tags move, round 4 (`numpy`, lecture-python-programming.ml#23) is regenerated at `@v0`.
+- **v0.29.2 released 2026-09-21** (release PR #318 `5f74d74`; payload #317 `62504c1`, on top of
+  #315 `cd41558`; §4a gate **completed**: 84/84 sync runs, 28/28 delivery + 28/28
+  `engineVersion: 0.29.2` verdicts per lane, scenario 17 on `.ml` delivered; `v0.29` = `v0` =
+  `5f74d74`; GitHub release published; `@v0` smoke `engineRef: v0` on all three lanes — tally
+  on #318) — patch: the scope sentence on the `ml` exercise rule that v0.29.1's failed gate
+  called for, carrying the editor's round-3 answers that v0.29.1 never delivered. **Round 4
+  (`numpy`, lecture-python-programming.ml#23) regenerated at `@v0` the same day**: three draws,
+  draw 1 sent (bare endings 1 / 37 / 2), one `ml_repair.py` comma applied and disclosed — the
+  first repaired seed; arm `experiments/ml-benchmark/arms/2026-09-21-round4-numpy-v0.29.2/`.
+  Harness note: the reset left one PR from the previous gate open on `.ml` (closed by hand).
+  W1 (#259) still targets v0.30.0. **2026-09-23 close-out**: the reset survivor's cause was
+  `gh pr list` with no `--limit` — fixed with a post-reset check on **#322** (closes #321); the
+  gate's one-draw blindness (a ~40% defect passed three gates) is **#320**, decided as a local
+  N-draw rate check — `tool-test-action-on-github/rate-check.sh`, release step 4b, on **#323**
+  (v0.29.2 reads 0/12, v0.29.0 4/12). **Next engine-side: the rule-2 arm** (rewrite around the
+  editor's everyday-speech test, judged held-out before it ships).
 - **v0.29.1 is tagged but NOT released (2026-09-21)** — its §4a gate came back 83/84: scenario 17
   on the `.ml` lane failed twice (the model wrapped a plain `## Exercises` list in
   `{exercise-start}`; structural parity refused the file). `v0` = `v0.29` = `a6fda54` (v0.29.0)

.dev/log/2026-09-21-v0291-gate-scenario17.md

@@ -24,3 +24,7 @@ The first six draws per arm read 1/6 against 5/6 and looked like a clean regress
 **Merged, release cut.** #317 merged as `62504c1` (Copilot: two release-status wording comments — STATE entries reworded, CHANGELOG marks 0.29.1 `[YANKED]` with a note, `cc7c6ab`); release PR for v0.29.2 opened on `release-v0.29.2`.
 
 **Next (as written before the merge)**: PR → merge → **v0.29.2** (v0.29.1 stays a tag that was never released; the CHANGELOG says so) → §4a gate on the new tag → floating tags → smoke → release → regenerate round 4 at `@v0`.
+
+**Addendum — v0.29.2 released.** #318 merged `5f74d74`; tagged `v0.29.2`; §4a gate on the tag: census 9/9 at `@v0.29.2`, 28 source PRs, **84/84 sync success**, 28/28 delivery and **28/28 verdicts at `engineVersion: 0.29.2` per lane**; scenario 17 on `.ml` delivered (test-translation-sync.ml#242). The `.ml` lane briefly read 29 PRs: the reset had left test-translation-sync.ml#192 open from the v0.29.1 gate — closed by hand, and worth a look in the script's close-all step. `v0.29` and `v0` → `5f74d74`; `@v0` smoke (scenario 01) 3/3, `engineRef: v0` on all lanes; GitHub release published (title = tag; notes cover #315 and #317 and say not to pin `v0.29.1`). Tally on #318.
+
+**Round 4 regenerated at `@v0`.** Three draws of `numpy`, draw 1 chosen by lint (bare endings 1 / 37 / 2), `ml_repair.py` applied — one comma, `Columns-ഉം, rows-ഉം` — and disclosed; force-updated lecture-python-programming.ml#23 (`889d001`), retitled, description rewritten, editor told it is ready. He had not started on the v0.29.0 draft. Arm: `experiments/ml-benchmark/arms/2026-09-21-round4-numpy-v0.29.2/`.

experiments/ml-benchmark/arms/2026-09-21-round4-numpy-v0.29.2/README.md

@@ -0,0 +1,35 @@
+# Arm: round 4 (`numpy`) regenerated — three draws at v0.29.2, chosen by lint, 2026-09-21
+
+The round-4 calibration seed for lecture-python-programming.ml was first generated at v0.29.0
+(`../2026-09-18-round4-numpy-v0.29.0/`, draw 3 sent). The editor then answered the round-3
+questions (lecture-python-programming.ml#22) before starting on it, so the seed was regenerated
+at the release that carries his answers — the round-3 precedent at v0.28.0. Three draws at
+v0.29.2 (`5f74d74` = `@v0`) from `lecture-python-programming@b0b0b56` (`numpy.md` unchanged
+since `1706cea`), `init -f numpy.md --localize none -m claude-sonnet-5`.
+
+| Draw | bare endings before a cell/list | paragraph without punctuation | lowercase-initial | banned | hortative watch | `-ഉം` pair, no comma | *For example* → ഉദാഹരണത്തിന് | `provide ചെയ്യ…` | exercise blocks | headings / code cells |
+|---|---|---|---|---|---|---|---|---|---|---|
+| 1 (**sent**, after repair) | **1** | 0 | 0 | 1 | 1 | 1 → 0 | 0 | 0 | 8/8 | identical |
+| 2 | **37** | 1 | 0 | 2 | 0 | 1 | 6 | 0 | 8/8 | identical |
+| 3 | **2** | 8 | 2 | 2 | 2 | 1 | 0 | 0 | 8/8 | identical |
+| v0.29.0 draw 3 (first sent) | 3 | 0 | 0 | 3 | 1 | 2 | 0 | 6 | 8/8 | identical |
+
+`ml_repair.py` was applied to draw 1 and **disclosed on the PR** — the first seed to go to the
+editor repaired. It made one change: `Columns-ഉം rows-ഉം` → `Columns-ഉം, rows-ഉം` (line 280).
+Every draw here leaves exactly one hyphenated pair without its comma — that pair in draws 1 and
+3, `Step 1-ഉം 2-ഉം` in draw 2 — and those are the same two lines the 2026-09-21 answers arm
+missed: the prompt rule does not reach them, the script does. Left as
+generated and disclosed: line 878 (ends on a closing bracket before a cell), line 898 (*In
+fact* → വാസ്തവത്തിൽ, a banned rendering), line 1092 (a genuine future statement the hortative
+watch flags). Whether he touches the repaired line is the first field evidence for #260.
+
+Terminal punctuation remains the draw-dependent class (1 / 37 / 2 here; 37 / 34 / 3 at
+v0.29.0): best-of-N by lint is still what keeps it away from the editor.
+
+One more observation, raised in review of the PR that archived this arm: draw 2, line 177,
+carries a garbage token — `2x2 array undegerbestellen ഉണ്ടാക്കാൻ` — a Latin-script string that
+is neither English, code nor a glossary term, emitted by the model and absent from the other
+two draws and from the draft sent. It is left exactly as generated: the draws are the record.
+It is also a lint gap — nothing in `ml_metrics.py` flags a Latin token that is not English,
+code or a pinned term (#301) — and draw 2 was already the one not chosen, on bare endings.
+