docs(M296): three-month-break closeout + M-counter M280→M296

[noahgift]

Summary

Operator-directed three-month break begins. V1_004 still open but empirically narrowed.

Headline finding from this session (M286-M295)

Through 12 PRs across the V1_004 chain, four candidate variables were tested as the load-bearing one behind 0%-tool_call emission on Phase 6 fixtures:

Variable	Test	Result
Inference stack quality	M286 KV cache + 3-knob + EOS + clean_chat_output	Necessary fix; not sufficient
Active params count	3B (30B-A3B-MoE) vs 7B (dense Qwen2.5-Coder-7B-Instruct)	Both 0 tool_calls — refuted
MoE vs dense	qwen3_moe vs qwen2 dense	Same pattern — refuted
Few-shot prompt	3 concrete <tool_call> examples	No shift — refuted
Qwen-Coder finetune family	Smoke against Qwen3-30B-A3B-Instruct-2507 (non-Coder)	Bare tool_call JSON in 20 tokens — confirmed at smoke level

Bench-level partial refutation

F1 of the non-Coder Instruct bench: driver_error at turn 8, tool_use_count=0, 8 Markdown turns. The smoke-vs-bench divergence surfaces a second-order constraint: apr code's multi-turn prompt context (rendered history with previous turn's Markdown + "### Continue:" suffix) self-recursively reinforces the Markdown distribution even on a finetune that emits tool_call JSON in 1-shot smoke.

Three resumption paths

Scoped in evidence/phase-6/m296-three-month-break-closeout-2026-05-22.md:

• (a) Investigate apr code's render_history + per-turn prompt construction
• (b) Post-decode Markdown→tool_call parser in apr code (unlocks Qwen-Coder family for V1_004 as written)
• (c) V1_005 against different model class on Lambda Labs GPU (Llama-3.3-70B, DeepSeek-V3, Qwen3-32B-Instruct dense)

What this PR ships

• evidence/phase-6/m296-three-month-break-closeout-2026-05-22.md — comprehensive resumption playbook
• Partial bench archives preserved in-repo for historical record (5 dirs, ~33MB total)
• M-counter M280→M296 bumped on all 5 surfaces (README, CONTRIBUTING, top spec, status-snapshots, milestones)
• M296 row + M286-M295 rollup row added to milestones-m101-m111.md
• .gitignore — exclude .claude/ runtime artifacts

Project handoff state

• ✅ No in-flight benches (all background processes killed)
• ✅ No orphan apr serve / apr code processes
• ✅ Book live at https://paiml.github.io/claude-code-parity-apr/
• ✅ Doc-drift detector: 17/17 classes clean
• ⚠️ V1_004 still open (operator-coordinated decisions remaining)

Test plan

• bash scripts/check-doc-drift.sh — 17/17 drift classes clean
• All 5 M-counter surfaces synchronized to M296
• No orphan processes verified via pgrep
• CI green

Cross-references

• M286 evidence/phase-6/m32d-shipped-2026-05-20.md
• M287 evidence/phase-6/m32d-bench-pattern-2026-05-20.md
• M291 evidence/phase-6/v1004-sub-bench-b-pattern-shift-2026-05-21.md
• M292 evidence/phase-6/v1004-agent-text-loop-detector-2026-05-21.md
• M295 README + book PR (docs(M295): professional README + mdBook companion at paiml.github.io/claude-code-parity-apr #263)
• aprender#1832, #1837, #1842, #1844, #1846, #1849, #1852, #1853 (all merged)

🤖 Generated with Claude Code