feat(m2): doom-loop pruning for BFTSLiteStrategy

[suzuke]

Summary

Stacked on #3 (M1b SearchStrategy). First chunk of M2 — wires should_prune() into the search seam so BFTS stops re-expanding futile branches.

What's new

Doom-loop pruning

• BFTSLiteStrategy.should_prune(ctx, candidate_id) — a kept node is pruned when its trailing prune_threshold direct children all fail to produce a strict metric improvement.
• BFTSLiteStrategy.decide() filters pruned candidates BEFORE max-metric selection, so a high-scoring but doom-looped branch cannot starve a viable lower-scoring one.
• All-pruned → Stop(reason="all kept nodes pruned (doom-loop)"), not Continue/Restart — those are separate recovery policies.

Failure semantics (matches reviewer round 1)

• discard outcome → counts toward streak (always, regardless of metric)
• crash / violation / skip → counts (no metric to compare; still a failed expansion attempt)
• keep outcome with non-improving metric (incl. equal) → counts
• Strict improvement at any tail position → resets the streak
• Candidate without metric → should_prune returns False (defensive, bootstrap-safe)

Config

• search.prune_threshold: int = 3 — validated as positive int in _build_search.
• __post_init__ guard on BFTSLiteStrategy rejects prune_threshold < 1 for direct construction (tests/scripts that bypass config).
• Other strategies (greedy/restart) ignore the kwarg.

Reviewer trail

Round	Verdict	Findings
1	NEEDS_TWEAK	trailing-only counter / crash semantics / equal-metric / bootstrap-safe / all-pruned semantics / filter-before-select / config typing
2	VERIFIED	All 7 addressed; one non-blocking note on prune_threshold=0 direct-construction → fixed in 833cb06

Stats

• 2 commits, ~85 LOC strategy + ~12 LOC config + ~200 LOC tests
• 2,397 passed / 4 skipped, 0 regressions on top of M1b baseline
• 13 new unit tests in test_strategy.py (35 → 36 with __post_init__ test) covering reviewer's 10-case checklist + factory threading + bootstrap defence

Test plan

• Trailing 3 failures → prune at threshold
• Below threshold (2 failures) → not pruned
• Strict improvement resets streak
• Equal metric kept counts as non-improvement
• Discard with worse metric counts (regardless of metric)
• Crash / violation / skip count without metric comparison
• Unrelated siblings don't affect candidate streak
• Candidate without metric → return False (defence)
• Minimize direction strict improvement
• All pruned → Stop("doom-loop")
• decide() filters pruned, picks next-best
• BranchFrom to next-best when metric-best pruned
• Factory threads prune_threshold correctly; greedy/restart accept-and-ignore
• __post_init__ rejects 0 / negative
• Real-agent demo gate — 30 iter on optimize-compress, see docs/M2-DEMO-GATE.md. Doom-loop pruning explicitly fired on n3, n21, n20, n19 (each accumulated 3 trailing failures and was removed from BFTS's candidate set).

Real-agent demo gate (M2 30-iter)

	Iters	Best compression_ratio	Stop reason
Greedy	9	2.2528	5 consecutive failures, hard-stop
BFTS-lite	30	2.5013	clean max_iterations=30

Greedy hit the legacy max_retries=5 wall at iter 9 — exactly the failure mode this PR's doom-loop pruning is designed to escape. BFTS-lite reached iter 30 with 6 BranchFrom events and 4 nodes explicitly pruned by should_prune, finding a deeper kept path (n1→…→n21, 10 levels) that beat greedy's local optimum by +11%. Total cost $2.05, ~55 min wall (parallel runs).

Known limitations (M2 PR 11+)

• Pruning is trailing-only; windowed (3-of-5) is a v2 follow-up if domains show oscillating bad branches.
• Pruning policy lives on BFTSLiteStrategy. If multiple future strategies want to share it, extract a Pruner interface then (YAGNI now).

🤖 Generated with Claude Code