docs(M290): V1_004 follow-up snapshot — 5 PRs shipped + root-cause diagnosis

[noahgift]

Summary

Consolidated view of the V1_004 follow-up chain after fixture 10's history inspection revealed 3 independent dense-vs-MoE gaps causing the M287 verbosity pattern.

Root cause diagnosis

Fixture 10 (oo__05-builder-pattern, turns=7) student output turn 2+ was repeated "Human: I need to see..." text. Three independent gaps:

1. No EOS stop_token in try_qwen3_moe_backend
2. No clean_chat_output post-decode cleanup
3. No few-shot examples in CODE_SYSTEM_PROMPT (model emits Markdown rust instead of <tool_call> JSON)

5 PRs shipped this session

PR	Status	What
#1832	MERGED	M32d KV cache, 19× speedup
#1837	MERGED	sampling contract
#1842	MERGED	sampling impl (+ #1844 squashed)
#1846	MERGED	3-knob HTTP wire-up
#1849	OPEN	few-shot prompt
#1852	OPEN	EOS stop_token + clean_chat_output

4-layer defense post-merge

• Layer 1 (prompt): #1849 shows examples + bans Markdown blocks
• Layer 2 (sampling): #1842 lets operator tune temperature/penalty
• Layer 3 (real-time stop): #1852's stop_tokens halts at EOS
• Layer 4 (post-process): #1852's clean_chat_output strips runaway

Revised dispatch sequence (sub-benches A/B/C)

Each ~10-15hr wall. Operator-coordinated. Acceptance: student_pass_rate > 0 in any sub-bench discharges V1_004.

Mechanical snapshot doc. M-counter NOT bumped.

🤖 Generated with Claude Code