perf(mobile): reuse chat lists across navigation renders#27026

Timing evidence correction (October 30): The timing tables below used a validator that did not reject app frames marked as dropped by SurfaceFlinger. Their presentation-level percentiles are provisional FrameTimeline associations, pending a corrected audit and rerun. The new capture records each qualifying draw and requires an explicitly presented frame. Source tests, derivation counts, and viewport recordings remain separate evidence.

Problem

Unrelated navigation rebuilds the mobile chat list even when its threads, filters, and shelf state have not changed. The reported case involved hundreds of active threads and thousands of settled threads on Android.

Change

Memoize the shared mobile list derivation stages, retain queue-key identity while Home is hidden by Activity, and stabilize Home's selected-project scope references. Thread, filter, queue, shelf, and clock updates still invalidate the affected results.

Verification

• Follow-up 5f0c64207838 runs the existing 1,200-thread regression through the real settled-shelf hook and checks callback, layout, and item reuse. This follow-up changes only the test; measured production code remains the version at b238f41e8038.
• The large-list regression fails on the pinned base and passes with memoization. The actual queue atom's lifecycle regression fails before the retention correction and passes after it. The clock case independently verifies a snoozed thread waking.
• vp run check:changed 960d5761c13f68a2f10635bcd26489a5fa404ec1 passed. Focused queue and hook regressions passed 2/2.
• Rerun: all 73 focused regressions passed before native timing; all five sizes have 30 valid returns per path and revision, with persisted/client/logical cardinality checked on all 20 launches.
• Independent adversarial reviews: GPT-6.1 Sol and Claude Opus 5.5, both xhigh. Sol identified the selected-project gap; native testing identified queue disposal. A fresh Sol xhigh review found no correctness defect after both corrections.

Five-size Android latency (debug JavaScript)

Five-size load-gated rerun: median (p95), milliseconds. Thirty warm returns per path, revision, and size across two interleaved launches: 600/600 verified native observations. Baseline 960d5761c13f; corrected production b238f41e8038, unchanged by test-only follow-up 5f0c64207838. All-active fixtures verify N persisted eligible threads, N unique hook inputs, and N logical thread rows. This cohort is separate from the earlier measurements.

The endpoint is real native input UP → SurfaceFlinger presentation of the marked frame containing the correct leading row, with Chats selected and its native Screen/Fragment active. It can follow earlier partial pixels. All traces passed health and frame-association checks; no presentation fallback was used. Measurements use the same instrumented Android 15 emulator/debug APK, Hermes, React Compiler, development JS (__DEV__ true), working shelf enabled, and no recording during timing. Production and physical-device performance remain unmeasured.

All 100 blocks passed the declared lower-load starting gate (three five-second aggregate CPU samples below 40% busy and one-minute load below 14 cores). The shared host remained busy during collection/analysis windows: median CPU 61.2%, peak one-minute load 19.8. The public PR feed stayed live. These comparisons are descriptive and do not isolate host load or establish a production latency claim. Raw samples retain slow observations and minute changes. A live-log pacing failure caused the entire first partial 10-thread cohort to be archived and restarted; no endpoint or validity rule changed, and no primary sample needed the pacing allowance.

At 10,000 threads, the two corrected launch medians were 983/983 ms from PRs and 251/233 ms from Settings, compared with baseline 1757/1831 and 886/958 ms. Median hook CPU fell from about 86–90 ms to 0.056–0.070 ms. The corrected list skipped layout derivation in 291/293 clock-stable windows; two 10,000-thread PR returns rebuilt despite no exposed input reference change. Their internal invalidation cause remains unconfirmed, and both remain in the results.

The PR path still takes about a second at larger sizes, and its 10,000-thread p95 rose from 2,025 to 2,186 ms. One-thread Settings p95 also rose slightly. This confirms lower median latency and less derivation work, not uniformly improved tail latency or a complete fix for the original intermittent hang.

Full rerun report · Raw samples · Summary and confidence intervals · Reproduction harness and evidence

Earlier cohort report. Earlier observations and the user-paused partial cohort remain separate.

Follow-up: PR-return latency and the higher p95

A separate 10,000-thread experiment used minified production JavaScript (__DEV__ false) in the same instrumented debug native APK. Both revisions had steady chat-list clock inputs and forced one-minute advances as separate cases: 30 returns per cell, across two launches per revision in ABBA order. 240/240 native returns passed real-input, requested-clock React-commit, correct-row, presented-frame and trace-health validation; all 40 host gates passed. The clock control changes the list's clock output, not Date.now or every shared-clock subscriber.

Median (p95), milliseconds:

The higher corrected p95 did not recur in these cases. Both worst corrected PR returns in the earlier debug-JS cohort had minute changes; the baseline PR cohort had none. The newer experiment also changes JS mode, so it does not prove that clock imbalance alone caused the old tail or erase those observations. The corrected clock-advance Settings median is slightly higher; necessary clock recomputation remains costly.

Hermes sampling and saved scheduling traces identify substantial work reactivating Chats’ gesture/animation effects under Activity, amplified in development by Worklets stack capture. Settings largely avoids that lifecycle. No independent PR-list computation bottleneck was confirmed. The corrected steady PR median remains 351 ms versus Settings 153 ms; this reactivation cost remains an optimization target. All 60 corrected steady returns skipped layout derivation; all 60 forced-clock returns rebuilt. This follow-up changes no application code, and no full native release/physical-device claim is made. Host/background PR-feed variation and a validation pause are recorded in the evidence; no slow samples were removed.

Investigation report · Raw matched samples · Summary with intervals · Reproduction harness

Selected-project before recording · Selected-project after recording. Recordings are separate from the quantitative series.

Follow-up: minute-driven layout invalidation

5cfa1f95b6a1046541a47eebe8fd774ee31ee6de retains layout and move availability between snooze deadlines. Row labels/menu timestamps still refresh with the real minute. Search retains real time for recency ranking; crossing a snooze deadline rebuilds membership/order. Other thread, filter, queue, shelf and inbox invalidation remains unchanged.

This preserves Maria's main merge and additional real-shelf coverage at f1faee5d0082317ee46c2ee5d73927f24f6ee04b. The earlier tables above target earlier source. The independent viewport fix is #27393 and is absent from both arms below.

New native comparison: parent f1faee5d vs 5cfa1f95, 10,000 active threads, a controlled one-minute advance before every return. ABBA launches, 30 observations per revision/path. 120/120 passed input, requested-clock commit, unique input/logical row cardinality, correct active-visible content, frame presentation and trace-health checks; all 20 starting host gates passed; no fallback or discarded primary sample. Every observed row menu timestamp matches the requested minute.

Corrected launch medians were 469/468 ms from PRs and 347/348 ms from Settings; baseline launches were 568/535 and 444/384 ms. Median hook elapsed fell 81→26 ms and 87→29 ms respectively. Layout rebuilt on all 60 baseline returns and none of 60 corrected returns, while row items kept the fresh minute.

Same Android 15 emulator/instrumented debug native APK, Hermes/compiler, minified production JavaScript (__DEV__ false), working shelf on. Endpoint remains real input UP → first confirmed correct active/visible row frame presented by SurfaceFlinger. It may follow earlier partial pixels. This follow-up is at 10k only; it does not replace the earlier five-size cohorts or establish full release/physical-phone performance. Live public PR-feed and shared-host variation remain.

The corrected 705 ms tail still has zero layout derivations: hook elapsed 34 ms vs 23 ms in a typical 469 ms observation. Extra JS work and native/emulator graphics-buffer waits dominate the remaining difference; overlapping slices cannot be summed into a causal budget. This does not eliminate all hitches.

The hook regression failed before this change and passes after, including retained layout, real-minute item refresh, snooze crossing, search ticks and injected backwards timestamps. vp run check:changed f1faee5d0082317ee46c2ee5d73927f24f6ee04b passed (64 focused tests plus compiler checks). Fresh Sol and Opus xhigh reviews found no introduced defect. Backwards wall-clock clamping in the unchanged source clock hook is pre-existing and not fixed here.

Full protocol/results · Raw samples · Summary and descriptive intervals · Source and reproduction harness

Minute-return before recording · After recording. These are separate demonstrations, with the same fixture/build flags and a live prewarmed PR feed; their timings are outside the quantitative cohort.

Final CI integration: 318fdcab0829a6db948bd470c97a63bfb2c3f11e merges main 35c1210331a70fd3f460f9ad3425aed7d9a263cf and repairs three obsolete helper references in newly merged server tests (two provider helper names and one usage-cache writer call). All 107 PR-service tests and 10 usage-cache tests pass. vp run check:changed 35c1210331a70fd3f460f9ad3425aed7d9a263cf passes 183 focused tests, affected typechecks and compiler coverage. Both measured chat-cache source files remain byte-identical to 5cfa1f95; the native timing cohort remains pinned to that earlier exact runtime and was not rerun after the upstream integration.

A fresh independent Sol xhigh review of the nine-file integration diff at 318fdcab08 found no confirmed defect. Comment-only lint follow-up 0f885a862239e695e2b307d4bca561766ae61631 removes an issue-number reference from an upstream server comment; runtime behavior is unchanged. Its latest CI checks all passed or were skipped, and Macroscope approved this head. Final changed-file verification passed 440 focused tests plus compiler coverage. The native measurements remain pinned to 5cfa1f95 as recorded above.

Scope and approval

Alex requested this investigation and testing. This is a focused performance correction to an existing list, with no new setting or wire behavior. Web/desktop use a separate list; providers, MCP, and connection protocols do not change. iOS native performance, production builds, and streaming workloads remain unmeasured. Supplemental Android scenarios and their limits are recorded in the full report.

Model: GPT-6.1 Sol (xhigh reasoning) & Claude Opus 5.5 (xhigh effort, adversarial review) | Harness: Codex in T3 Code & Claude Code in T3 Code