fix: bound large-repo memory and traversal work

[danusha2345]

Summary

• bound resolver and C function-pointer synthesis caches by retained bytes, release pass-local facts early, and merge sequential synthesis results incrementally
• reject oversized package targets before UTF-8 decoding, preventing HarmonyOS .har archives and similar dependency blobs from exhausting V8 heap during resolution (Daemon OOMs (~4 GB heap) on small project during startup catch-up sync; stale daemon.pid + OS PID reuse then blocks all daemon restarts (unlock does not help) #1553)
• replace repeated source-prefix scans in dynamic-dispatch synthesis with linear line lookup (folded from fix: avoid quadratic synthesis line lookups #1580)
• recycle a loaded query-worker burst to one fresh worker after 60 seconds idle
• make shortest-path BFS linear and bound impact traversal at SQL level with explicit MCP truncation
• release tag-based CFML parse trees and prune Linux watchers for removed directory subtrees

Validation

• full suite: 186 test files, 3,131 passed, 9 skipped
• native kernel build and TypeScript build pass
• focused archive-read and quadratic-synthesis regressions pass
• 290 MB sparse .har probe: rejected before decode, 58 MB peak RSS, 0.05 s
• ArduPilot full index (13,633 files): peak PSS reduced from 4.09-4.26 GB to 3.316 GB with resolver workers disabled
• ArduPilot graph remained identical to baseline: 177,869 nodes / 474,074 edges
• 100,000-node shortest path probe: 0.14 s, ~119 MB max RSS
• focused query-pool, graph, CFML extraction, and Linux watcher regression tests pass

Commits are separated by defect for review.