Govern output and thrashing context

[soolaugust]

Summary

• add synchronous output working-set governance for long/incomplete assistant responses
• proactively hard-reclaim transcripts when users report input context-window 400s
• convert thrashing critical handling from manual compact/clear guidance to automatic working-set/reclaim pressure governance
• add regression coverage for output paging, reported 400 reclaim, and posttool thrashing accounting

Validation

• python -m py_compile hooks/thrashing_detector.py hooks/posttool_guard.py hooks/output_governor.py hooks/prompt_budget_guard.py
• python -m pytest tests/test_posttool_guard_thrashing.py tests/test_output_governor.py tests/hooks/test_prompt_budget_guard.py -q
• python -m pytest tests/test_output_governor.py tests/test_transcript_reclaimer.py tests/test_posttool_guard_thrashing.py tests/test_compact_epoch_guards.py tests/hooks/test_prompt_budget_guard.py tests/test_production_benchmark.py -q
• python benchmarks/production/runner.py --suite smoke --fail-under 85 --report /tmp/vmem-benchmark.json --markdown /tmp/vmem-benchmark.md

🤖 Generated with Claude Code