evidence + diagnosis: β-coop layer-sweep + 12L_3_47 survival-soak gate failure

[Natfii]

Summary

• Stage 2b 5-run survival soak at the 12L_3_47 β-coop arm (CUTE_PHASE_E_LAYERS=3,7,11,15,19,23,27,31,35,39,43,47, CUTE_WO_SPLIT=8) fails Gate 2b: silent quality collapse under sustained load (Runs 4-5 collapse mid-run; container stays alive; 0 errors logged; no docker-log corruption hits).
• Bundles the Task 0a/1a infrastructure the sweep depended on (sentinel-file workaround for CUTE_PHASE_E_* + CUTE_BETA_REGION_TIMING, [PHASE_E_DISPATCH] audit log, NVLLM_BIND_MOUNT_QWEN35=1) plus the full Stage 0a + 5-arm sweep + Stage 2b soak evidence.
• Adds docs/research/2026-05-09-beta-coop-layer-sweep-wo8/DIAGNOSIS.md opening a diagnosis arc: 6-leg bisection plan, leading hypothesis (persistent β-coop workspace/reset lifecycle), D3 per-request instrumentation design, side-hardening note (fail-closed CUTE_PHASE_E_LAYERS parsing).
• Memorializes 2L_3_7 stays the serve-cute.sh default via a header CAUTION block in scripts/serve-cute.sh.

Soak verdict

Run	Correct	Errors	Pass
1	48/50	0	yes
2	48/50	0	yes
3	48/50	0	yes
4	11/50	0	no (collapse at Q12, persistent through end of run)
5	37/50	0	no (inherited collapse Q0-Q13, sharp recovery at Q14)

gate_2b_pass=false, container_alive_at_end=true, docker_log_corruption_hits=0. The Q13→Q14 recovery (261s broken wall → 64.3s clean wall) is sharp/categorical, not analog drift — strong signal for a state flip (eviction / rotation / reset) rather than gradual numerical degradation.

Leading hypothesis (per friend code-review)

β-coop reads its workspace buffers from persistent impl attributes (wo_output, mlp_partial_fp32, counters, barriers) because host-side zeroing hung CUDA graph capture. Resets fire at multiple sites (captured memset op, pre-launch zero_(), inside-kernel MLP-partial reset). A missed reset → persistent garbage → silent quality collapse across many requests until a natural overwrite clears it. Matches the observed sustained-load + sharp-recovery shape better than generic "approximation drift."

See DIAGNOSIS.md for the full bisection table (D2.0–D2.5: wo1, lite, phaseE=0, layer-count, 2L survival control), the demotions (prefix caching off because hybrid models default-off already; eager mode broken on SM120), and the D3 instrumentation field list.

Test plan

• Stage 0a baseline (2L+wo8 GSM8K-50 + per-call β median): see stage0a/verdict.md
• Stage 1a runner harness: runner.sh, arms.csv, extract_dispatch_log.py, dispatch-audit gate
• Stage 1b-c 5-arm sweep (2L, 4L, 8L, 12L, 16L) with per-arm GSM8K-50 + dispatch audit: see sweep/summary.md
• Stage 2a dev baseline pick (12L_3_47)
• Stage 2b 5-run survival soak at 12L_3_47 (this PR's headline artifact): see soak/summary.md
• D2.1–D2.5 bisection legs (follow-up — each is a 10h-wall soak)
• D3 per-request instrumentation patch (follow-up)
• Side hardening: fail-closed CUTE_PHASE_E_LAYERS parsing (follow-up)

AI assistance disclosure

Claude Opus 4.7 was used for runner scripting, sweep analysis, summary aggregation, and diagnosis-plan authoring. All runtime evidence (5 soak runs, 5 sweep arms, Stage 0a baseline, dispatch audits) was produced on the human submitter's DGX Spark. Code changes were reviewed line-by-line by the submitter and an external friend whose code-review notes are credited inline in DIAGNOSIS.md (load-bearing citations verified live against the working tree before incorporation).

🤖 Generated with Claude Code

[Natfii]

Closing without merging.

The β-coop sustained-load collapse this PR's evidence was diagnosing was resolved 2026-05-15 via the SSM zero-on-realloc patch (PR #13). The cherry-pick bisection arc this PR documented (Stage 2b + D2.1–D2.6) turned out to be a dead end relative to the actual substrate (stale mamba recurrent state on slot recycle, not anything in the PR #10 tier-1 cherry-picks) — the bisection legs perturbed the shape but never fixed the magnitude, which is the canonical "substrate not in the bisect frame" signal (see memory:feedback_substrate_not_cherry_pick).

The diagnosis-arc receipts are preserved in:

• memory:project_beta_coop_sustained_collapse (full historical record of D2.x bisection + falsified hypotheses)
• memory:feedback_substrate_not_cherry_pick (the methodology lesson)
• PR feat: SSM zero-on-realloc guard + sentinel ablation harness #13 commit messages and benchmarks/nvllm/traces/ssm_zero_on_realloc/2026-05-15-sentinel-ablation/summary.md (the resolution + harness-validation evidence)

No code in this branch needs to land on main; the resolution path was different.