main: bound long-running memory growth + heap-pressure watchdog

[Juliusolsson05]

Summary

Closes three concrete unbounded-growth paths in the Electron main process and adds a single-shot heap-snapshot watchdog so any remaining leak leaves a forensic artifact for follow-up.

Motivated by an OOM crash (Reached heap limit Allocation failed - JavaScript heap out of memory) at ~3.84 GB / 4.09 GB after 19h 51m of uptime. Stack frames _DPSNextEvent → -[NSApplication run] → ElectronMain confirm it's the main process, not a renderer. Default v8 main heap is 4 GB on macOS; we are NOT bumping that — bandaging the cap just postpones the crash. The fix is bounded growth.

What changed

• src/main/performance/heapWatchdog.ts (new): samples used_heap_size every 30 s and writes a .heapsnapshot to ~/.config/cc-shell/heap-snapshots/ the first time the process crosses 3 GB. Single-shot — additional samples in the same near-OOM window are no-ops. Wired in src/main/index.ts to app.whenReady/before-quit.
• src/main/storage/feedDebugLog.ts: the per-session feedDebugWriteQueues and lastWrittenFeedDebugId Maps used to grow forever. Now self-reap when the queue settles (with a === next successor check to avoid racing concurrent appends), and a new forgetFeedDebugSession() drops the cursor on session close. Wired into SessionManager.kill() AND both natural-exit paths (claude/codex agent exit + terminal exit).
• src/main/worktreeActivity/: the in-memory transcripts map is now an LRU bounded at 1000 entries with disk fallthrough on miss. Critically, refreshNow now loads the disk index ONCE per refresh when totalOnDisk > IN_MEMORY_MAX_ENTRIES, instead of reparsing the full ~30 MB JSON on every per-candidate LRU miss — the old shape would have churned ~120 GB of allocations per refresh on heavy users (5000 transcripts × 30 MB).
• src/main/storage/paths.ts: pruneStaleFeedDebugLogs() removes feed-debug/*.jsonl files older than 30 days at app startup. Matches observed disk pressure (149 GB / 5007 files / largest single 9.8 GB on the affected machine).

Also drops the loose scripts/test-*.ts harness scripts and the testing/regression-* / testing/ghost-demo bits — none of them are wired to CI and the project has been operating without them. The interactive testing/rendering/ harness stays (it's a dev tool, not a test suite).

Test plan

• npx tsc --noEmit -p tsconfig.node.json shows only pre-existing errors in unrelated files (claude-code-headless package types, dictation, pidusage). No errors cite the new/changed files.
• Manual: launch the app, lower the watchdog threshold to ~256 MiB temporarily, open enough panes to trip it, confirm a .heapsnapshot lands in ~/.config/cc-shell/heap-snapshots/. Restore the threshold before merging.
• Manual: with old feed-debug/*.jsonl present, launch the app, confirm a [feed-debug] pruned N stale logs line in the console.

Notes for reviewers

• 3 GB watchdog threshold leaves headroom for the snapshot writer itself (which allocates).
• LRU cap of 1000 was picked from observed working-set sizes; can tune up if heavy-user telemetry shows churn at that boundary.
• Independent of the title-turn PR (different subsystems; treat them as separate landings unless heap snapshots from this PR's watchdog point at something the title-turn fix would also address).

🤖 Generated with Claude Code