fix: reduce memory and disk usage from harness subprocess artifacts

[AbirAbbas]

Summary

• 7 GB RSS on Railway traced to Python heap fragmentation (pymalloc never returns freed arenas to OS) and unbounded evidence re-extraction in the coverage loop
• 2.7 GB disk waste from accumulated Claude Code session data (~1.14 GB in ~/.claude/projects/), V8 JIT .so files (~106 MB in /tmp), and pyright temp directories
• Fixes are all concurrency-safe — designed for multiple simultaneous reviews in the same container

Changes

Memory (heap)

• MALLOC_TRIM_THRESHOLD_=0 env var in Dockerfile so glibc aggressively returns freed pages
• malloc_trim(0) called via ctypes after each review's GC pass
• evidence_map and verification_map explicitly cleared after use in _run_review_layer
• Coverage loop now only extracts evidence for new gap findings instead of re-extracting for all accumulated findings each iteration (was quadratic, now linear)
• Explicit del gap_evidence between coverage iterations

Disk (session artifacts)

• Snapshot-based Claude session cleanup: snapshots ~/.claude/projects/ before review, deletes only sessions created during this review afterward (safe for concurrent reviews)
• Empty pyright-* temp dirs cleaned from /tmp
• Orphaned V8 .so files older than 60s removed from /tmp (age threshold protects concurrent harness calls)

Investigation findings

Component	Size	Type
Python main process heap	6.8 GB RSS (peak 14.1 GB)	RAM
.claude/projects/ session data	1.14 GB (49 sessions)	Disk
.opencode/bin/	135 MB	Disk
/tmp/*.so V8 artifacts	106 MB (24 files)	Disk
/tmp/pyright-* dirs	~40 empty dirs	Disk
.npm + .bun caches	140 MB	Disk

Test plan

• ruff check passes on all changed files
• Existing tests pass (test_staggered_gather.py)
• Python AST parse verification for both changed files
• Import verification with PYTHONPATH=src
• Deploy to Railway and verify RSS drops after reviews complete
• Run concurrent reviews and verify no cross-contamination of session cleanup

🤖 Generated with Claude Code