Epic: make CodeBurn's local interactive surfaces fast, truthful, and resilient

[avs-io]

Outcome

Make CodeBurn's local interactive surfaces feel fast, truthful, and resilient on both normal and heavy corpora.

A user should see useful, correctly labelled data quickly; changing period or provider must never leak stale totals; terminal and Desktop surfaces must remain responsive; packaged resident failures must recover without breaking the app.

Why these belong together

Each change remains an independently reviewable PR, but all slices came from the same 0.9.22 dogfood acceptance audit and share one user-facing quality bar. This issue is the acceptance ledger, not a request to merge them as one change.

CodeBurn Teams, plugin architecture, pricing semantics, sync policy, and feature-roadmap work are explicitly out of scope.

Merged slices

• fix(desktop): harden packaged resident CLI routing #1155 — packaged Desktop resident CLI routing and fallback reliability
• fix(tui): clear stale panels after terminal resize #1156 — terminal resize redraw / stale-panel cleanup
• perf(desktop): show exact totals while details refresh #1157 — Desktop exact totals first while detailed views refresh
• perf(tui): show cached periods before refreshing sources #1159 — TUI cached-period first paint and progressive source reconciliation
• fix(menubar): recover parked status item #1161 — Menu Bar parked status-item recovery

Review-ready draft slices

• perf(desktop): keep every destination useful across restarts #1163 — durable last-known-good Desktop destinations across restarts
• test: add repeatable release acceptance ledger #1164 — repeatable 60-case release acceptance ledger and runner
• fix(tui): clean cache locks on confirmed exit #1165 — TUI cache-lock cleanup on confirmed exit and interruption
• fix(desktop): reap CLI process trees on quit #1166 — Desktop resident/provider process-tree cleanup on real app quit

Shared acceptance gates

• cached values are explicitly labelled until reconciled with source
• no stale totals cross period or provider boundaries
• already-indexed standard period navigation settles within 1 second on the frozen heavy corpus
• exact-SHA heavy-corpus first-useful timings are recorded for affected surfaces
• settled terminal resize leaves no ghost panels or redraw artifacts
• packaged resident failure falls back cleanly without EPIPE or a dead surface
• packaged app quit leaves no owned CLI/provider process or cache lock
• focused automated suites, full regression suite, and build/type checks pass
• installed-artifact manual visual, navigation, exit, and recovery gates are recorded before closing this epic

Deliberately held, not hidden

The universal provider/period cache prototype is not in a PR. On the frozen heavy-corpus fixture its largest shard reached 109.6 MiB, JSON decoding took 231–312 ms, the measured event-loop stall reached 327.9 ms (over the 250 ms responsiveness gate), and filter/duplicate correctness was not yet proven. It stays a design input until the immutable provider-snapshot contract, compact manifest, bounded detail cache, and exact parity tests are implemented.

Follow-ups discovered

• provider-switch index reuse and heavy-corpus timing
• freshness policy while the user remains on heavy detail tabs
• per-phase cold discovery/fingerprint timing artifacts
• shared provider authentication/quota state across Desktop and Menu Bar
• any further optimization justified by the acceptance ledger rather than inferred from source alone

Close condition

Close only when each included PR is either merged or explicitly rejected with its evidence preserved, and the installed-product gates above have been run against the exact resulting candidate.

[avs-io]

Maintainer update: the epic now reflects the exact merged stack (#1155, #1156, #1157, #1159, #1161) and the four focused review slices now open as drafts (#1163 durable Desktop snapshots, #1164 repeatable acceptance ledger, #1165 TUI lock cleanup on exit, #1166 Desktop process-tree cleanup). Feedback on the earlier PRs has been incorporated or explicitly adjudicated in their threads/bodies. The broad universal provider-cache prototype is intentionally not being proposed: heavy-corpus testing measured a 327.9 ms event-loop stall, 109.6 MiB shard, and unresolved filter/duplicate correctness. I have added those hold reasons to the epic so the omitted code is visible rather than silently lost. CI reruns are in progress for two cache-lock lanes that showed a transient partial-JSON read / timeout; exact failed logs are being reconciled before any ready-for-review promotion.

[avs-io]

Quality stack update on exact current main 9a929c87ab94b55079e53ff6cb29292c0317402c:

• perf(desktop): keep every destination useful across restarts #1163 merged: durable, immediately useful Desktop destinations across restarts, with reviewer findings addressed before merge.
• test: add repeatable release acceptance ledger #1164 merged: repeatable 60-case release acceptance ledger plus a required Desktop typecheck/test/build CI job.
• fix(tui): clean cache locks on confirmed exit #1165 ready for review: TUI direct-exit cache-lock cleanup; exact head 38048078; all checks green.
• fix(desktop): reap CLI process trees on quit #1166 remains draft: Desktop process-tree cleanup; exact head 17f129c3; all source/CI/package/signing gates green, with only the final exact-head packaged Cmd-Q click held because the Mac locked.
• test(tui): cover exact-zero auto period #1168 ready for review: test-only rendered-TUI coverage for the already-shipped exact-zero Today -> 7 Days behavior; all checks green.

The universal provider-cache prototype remains intentionally unpublished. It improved parse time/memory, but a 109.6 MiB shard produced a ~328 ms event-loop stall and filter-correctness/cancellation gaps. That is useful research, not reviewable product code yet.

[avs-io]

Current exact-main correction (fa540325fb3bd56126dfd699d107ed184cb0b800):

• ✅ perf(desktop): keep every destination useful across restarts #1163 merged — persistent exact/stale desktop snapshots keep every destination useful while refresh continues.
• ✅ test: add repeatable release acceptance ledger #1164 merged — repeatable 60-case release acceptance ledger plus required Desktop CI gate.
• ✅ fix(tui): clean cache locks on confirmed exit #1165 merged — confirmed TUI exits remove owned cache locks and restore terminal state.
• ✅ fix(desktop): reap CLI process trees on quit #1166 merged — packaged Desktop quits reap the CLI process tree; source/CI/package checks passed. The final exact-head UI Cmd-Q replay remained explicitly disclosed as unperformed because the Mac was locked; a prior packaged artifact passed that interaction.
• ✅ perf(tui): show cached periods before refreshing sources #1159 already contains the user-facing exact-zero Today → 7 Days fallback.
• 🟢 test(tui): cover exact-zero auto period #1168 is ready for review — test-only rendered-Ink regression coverage for that exact-zero fallback, rebased onto this exact current main and locally green: focused 9/9, root 3,484 pass/5 skip, locks 35/35, TypeScript clean. CI is rerunning on the rebased head.

The larger universal provider-cache prototype remains deliberately unpublished. Its largest measured shard is ~109.6 MiB, JSON parsing is 231–312 ms, and the measured max event-loop stall is 327.9 ms; it also still has filter-correctness and cancellation gaps. Those are material product risks, so it is research rather than a reviewable PR.

This is the current authoritative status; it supersedes my earlier status comment in this issue.

[avs-io]

Final CI correction: all six required checks are green on #1168 head cdb4b03dc0beb0a31431bd0c8ce579e2f9e25023. It is mergeable and now waits only for the repository-required maintainer review.