perf(mobile): bound the parsed review cache

[t3dotgg]

Mobile review keeps every parsed diff and its native rows until the app registry resets. Opening more sections keeps adding memory.

The cache now keeps up to eight entries and 4,194,304 source characters. Nearby prewarming uses the same budget and reserves room for the selected section. Oversized diffs are not cached or prewarmed. Current view references and row IDs stay valid.

In a Node source proof with 64 sections, retained parsed diffs and prepared native row sets each fall from 64 to 8. The retained heap delta falls from 47.9 MB to 10.7 MB. In the 10-section prewarming proof, switching between the first two sections adds no parses after warmup. These are not device measurements.

M6 in the performance audit remains partial. Raw patches, drafts, comments, and preferences are unchanged. The source-character limit is not an exact heap-byte limit.

Verified with 31 focused tests, mobile typecheck, targeted lint, and real-source GC and prewarming proofs. No native code changed. No browser, device, or native build verification ran.

Created with GPT-6 Astra (preview) in Codex.

---

Note

Medium Risk
Changes review diff caching and idle prewarming semantics; wrong eviction or budget accounting could cause extra parses or stale assumptions, but behavior is heavily tested and scoped to mobile review performance.

Overview
Bounds mobile review parsed-diff memory by replacing unbounded per-section atom caches with a single LRU store capped at 8 sections and 4 MiB of raw diff source (not parsed/native heap).

getCachedReviewParsedDiff now tracks source length, evicts oldest entries when limits are hit, skips caching oversized patches (still returns a fresh parse), and clears on appAtomRegistry.reset. Whitespace-trimmed hits refresh LRU order; content changes reclaim budget for that section.

Prewarming no longer warms every other loaded section. getReviewDiffPrewarmSections picks nearest neighbors that fit the shared entry and character budget (reserving the selected section, including when cached source was retained with trailing whitespace). prewarmReviewDiffSection bails on null or oversized diffs without parsing.

New unit tests cover LRU across threads, source-budget eviction/reclaim, oversized behavior, registry reset, and prewarm selection/reuse.

^{Reviewed by Cursor Bugbot for commit 0369853. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Bound parsed review diff cache to 8 entries and 4 MiB source budget

• Replaces the per-section atom-family cache with a single registry-owned LRU map in reviewState.ts, enforcing an 8-entry maximum and a 4 MiB aggregate source-character limit across all threads
• Adds getReviewParsedDiffSourceCharacterCount so prewarming callers account for the full retained source length, including whitespace absent from an equivalent normalized input
• Reworks useReviewDiffPrewarming to plan prewarming of at most 7 loaded neighboring sections ordered by distance from the selection, skipping neighbors that exceed the remaining source budget
• Direct prewarming now returns early for oversized sources without invoking parsed-diff construction
• Risk: getCachedReviewParsedDiff evicts oldest entries globally across thread keys when either the entry count or source budget is exceeded; callers expecting unbounded retention will see parsed results rebuilt on cache misses

^{Macroscope summarized 0369853.}

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	13.5 KiB	13.7 KiB	+240 B (+1.7%)	15.1 KiB	✅
Codex	Thread snapshot wire	7.0 KiB	7.0 KiB	+7 B (+0.1%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.5 KiB	6.7 KiB	+233 B (+3.5%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	57.0 KiB	58.5 KiB	+1.5 KiB (+2.6%)	66.4 KiB	✅
Codex	Live turn messages	8	10	+2 (+25.0%)	21	✅
Claude	Total thread wire	13.7 KiB	13.8 KiB	+16 B (+0.1%)	15.1 KiB	✅
Claude	Thread snapshot wire	7.0 KiB	7.0 KiB	+3 B (+0.0%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.7 KiB	6.7 KiB	+13 B (+0.2%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	59.3 KiB	59.3 KiB	0 B (0.0%)	66.4 KiB	✅
Claude	Live turn messages	10	10	0 (0.0%)	21	✅

Baseline: c8f77e0 · PR result: 0369853 · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 113.8 KiB
• Claude decoded thread snapshot: 114.5 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.

[macroscopeapp]

Approvability

Verdict: Approved at 0369853

Macroscope's review found this PR approvable — This is a contained mobile performance change that bounds parsed review-diff retention and optional prewarming without changing the rendered review results or introducing a new capability. Focused tests cover LRU eviction, source budgets, reset behavior, and reuse of parsed/native results.

^{You can add or adjust custom eligibility rules. Learn more.}