[N=3] Qwen3.5-397B-A17B firmed results + P2a guard (#28 merged without it)

[Lightheartdevs]

Draft / WIP. Stacked on #28 (the N=1 + tooling/grader-fix PR) — base is that branch, so this PR's diff is only the N=3 delta. Merge after #28.

Firms the 397B entry from N=1 → N=3, graded with the fixed phase1_grade.py from #28.

Status

• ✅ No-think arm complete (N=3): 23/36 = 8/12 by majority — N=1 headline holds.
• ⏳ Think arm running — column fills on completion.

What N=3 already shows (no-think)

Replicates expose that two N=1 verdicts were single-draw luck:

• p3_market: N=1 ✓ → 1/3 (high variance, hand-graded citation dimension).
• p3_pm: N=1 ✗ → 2/3 (intermittent under-recall, not systematic).
• p3_doc: N=1 ✓ → 2/3 (occasional 700-word-limit trip).

Zero-variance (all 3 reps): p1_bugfix, the grounded mid-tier (p2_extract/ci/hallucination/triage), p3_business — and the consistent fails (testwrite/refactor/writing). The grader-fix-sensitive cells (bugfix 3/3, testwrite 0/3, refactor 0/3) are perfectly consistent → the P1 fix is stable.

Remaining before un-drafting

• Think N=3 column + think-vs-no-think at N=3 (does "thinking net −1" hold, or was the p3_doc flip also draw-luck?).
• Think-arm per-cell variance (esp. p3_market runaway behavior across reps).
• Fold N=3 into findings.md / README / manifest; retire the "N=1 provisional" framing.

🤖 Generated with Claude Code

[Lightheartdevs]

Retargeted base main (was stacked on #28's branch). Heads-up: #28 merged at 500c423, i.e. without the P2a --thinking guard (7ba3650) — the merge landed a moment before that push made it in. P2a is carried here instead (it's in this PR's history), so merging #30 brings both the guard and the N=3 results into main. Diff is clean: run_microbench.sh (P2a) + the 4 entry docs folded to N=3.

[Lightheartdevs]

Self-audit (against the grade/summary data)

Verified the published claims against ground truth. One real correction:

• 🔴 Fixed an overclaim: "all 72 cells done_signal / never ran away" was false — 69/72 done_signal. The 3 non-clean exits are all no-think: p3_market v2/v3 hit the 500-iter stuck threshold and p3_pm v1 model_stopped. Zero max_tokens/length runaways is still true (that's the Flash failure mode), so reframed as: 397B's pathology is stalling, not runaway — and thinking clears the market stall (think p3_market 3/3 done_signal vs no-think 2-stuck/1-pass). Corrected in findings/README/QUALITATIVE/manifest (4809534).

Verified accurate (no change): scorecard vs grade.json (23/36, 22/36 and every cell match); testwrite-think cov=99 / 153 passing, fails on logalyzer_unchanged; bugfix ruff 2→0, bench 11.2s→0.537s; redistribution deltas (market +2 / pm −2 / doc −1); internal-doc consistency; links resolve; no run-logs/agent-pilot in the diff; P2a guard present and carried to main.