Release merge order - queue before next release

[Optic00]

Sequencing tracker for the open PR queue before the next release. Two files force a serial order; everything else is parallel. Check items off as they merge - rebasing the remaining main.js branches happens as each one above lands.

Update (25 Jul, morning): the queue is essentially drained. Overnight the maintainer merged #416, #391, #421, #363, #387, #393 and #395, which closes out Wave 1 and Wave 3 entirely and leaves only #398 from Wave 2. Four PRs remain open on our side, all green and review-blocked only: #398, #368, #422, #426 (the last two are new, see below).

#398 needed a third rebase - ruzin's #425 (auto-install updates) added hooks to GeneralTab.tsx at the same spot the record-hotkey hooks go. Additive, both sides kept; rebased, verified (typecheck + unit + full T1) and back to MERGEABLE.

Wave 1 - independent, green, merge in any order

• ci: run backend unittests and renderer typecheck per-PR (#376) #393 (CI gates)
• fix(processing): break the forever-spinner when a stop produced no job (#343) #368
• fix(parser): port reasoning-tag normalization to the JS detail parser (#346) #370
• refactor(cli): narrow best-effort except clusters in simple_recorder.py (#379) #394
• test(e2e): widen T2/T3 expect.poll timeout to de-flake real-backend polls #397
• test(e2e): assert real summarize-model output in a nightly tier (#382) #395
• fix(settings): reset to General on bare /settings so browser Back can't strand a stale tab #415
• fix(renderer): nav hover contrast, dock alignment, and tooltips #384 (external, @WilliamDrewett - maintainer review pending on the cubic findings; ball is with the author)
• fix(meetings): drop the delete confirm step now that Undo exists (#234) #421

Wave 2 - app/main.js chain (serial, rebase each onto the last)

• refactor(main): extract settings-toggle IPC into settings-ipc.js (RFC #327 Phase 2.2) #364 (settings-ipc extraction - the pivot)
• fix(org): skip /policy fetch + seed on steady-state auto-backup (#192) #388
• feat(settings): warn when a local model may exceed available memory (#248) #389
• feat(auto-detect): gate meeting auto-detect to macOS 14+ (#116) #390 · touches GeneralTab.tsx
• feat: make the global record shortcut toggleable + documented (#396) #398 · touches GeneralTab.tsx — rebased onto current main (3rd time); green + MERGEABLE. Last one left in this wave.
• feat(audio): drop the Screen Recording requirement for system audio, min macOS 14.4 (#116 Part 2) #416
• feat(meetings): re-transcribe an existing recording (#266) #392
• feat(meetings): undo / soft-delete for note deletion (#234) #391

Wave 3 - security + platform fixes

• fix(macos): stop the Keychain "Safe Storage" prompt at every launch #363 (Keychain "Safe Storage" prompt)
• fix(security): deny window.open + block foreign navigation on all BrowserWindows (#377) #387 (deny window.open + block foreign navigation)

New since this issue was opened

• perf(settings): make toggles respond instantly instead of after two backend calls #422 - perf(settings): settings toggles wrote, then spawned a second backend process to read the value back, so the switch only moved after both. Measured on the packaged binary: set 315 ms + get 311 ms = ~630 ms of visible lag per click. Now optimistic with rollback and no read-back. Renderer-only.
• fix(export): save the note that's on screen as PDF, not always the Standard one #426 - fix(export): with a generated template report open, "Save notes as PDF…" exported the Standard structured note instead of the report on screen. fix(meetings): Copy notes copies the open template report, not the Standard note #318 taught Copy notes to follow the screen; the PDF path never got the same treatment. Renderer-only.

Verified in a packaged build (25 Jul)

An integration build of main + all five open PRs (#398, #368, #422, #426 and the external #384) was packaged and ad-hoc signed locally. Typecheck clean, unit 255 node:test + 128 vitest, T1 55/55. Two findings worth recording:

• Gate auto-detect to macOS 14+; reconsider app's minimum macOS #116 Part 2 confirmed on real hardware. Launch + record produced exactly two macOS prompts - Microphone, then "System Audio Recording" - and no Screen Recording prompt. The packaged Info.plist carries only NSMicrophoneUsageDescription + NSAudioCaptureUsageDescription, and askForMediaAccess is called with 'microphone' only.
• UX nit from that test: the two prompts arrive back-to-back with no in-app context (mic and loopback are acquired in one sequence), which reads as a duplicate request. Worth explaining both in onboarding, or asking for system audio only when the toggle is enabled. Not release-critical.

Not this release

• chore(transfer): repoint all references to github.com/stenolabs/stenoai #424 (draft, maintainer) - the ruzin/stenoai → stenoai/stenoai reference sweep; merges only after the org transfer.
• feat(keywords): Custom Keywords backend — glossary-driven transcript + summary correction (draft) #367 (draft) - Custom Keywords backend, still in progress
• feat: Cloud transcription models and UI enhancements #349 (draft, external) - cloud transcription models; the source of the reverted feat(transcription): OpenAI-compatible cloud ASR engine (from #349) #401
• inject google oauth credentials at build time #154 (draft) - build-time OAuth secret; separate security discussion, do not merge as-is
• feat: enhance Chinese language support with variant handling and conversion #102 (external, changes requested + conflicting) - Chinese-language support, needs author follow-up
• feat: add configurable locally invoked CLI provider #420 (external) - Codex/Claude CLI providers. Correcting my earlier note here: I called this a flat ToS blocker, which overstated it. Headless mode is a documented, supported way to run both CLIs, the user brings their own login, and nothing is being resold. Two things worth resolving before it lands: the label reads "Local CLI (on-device)" while the description directly below says meeting text goes to that CLI's service (and that is where a usage warning belongs), and Codex's sandbox read-only stops writes but not tool calls - worth enforcing in config rather than asserting in the prompt, given this feature is fed other people's meeting transcripts. Maintainer's call.

Nothing is waiting on work from our side. Remaining must-haves: #398 (finishes Wave 2) and #368.

[Optic00]

Closing this tracker — v0.6.5 shipped on 26 Jul, so the release it was sequencing is done, and the checklist above has been stale since 25 Jul.

Final state of what it tracked:

• Wave 1 — all merged except fix(processing): break the forever-spinner when a stop produced no job (#343) #368 and fix(renderer): nav hover contrast, dock alignment, and tooltips #384 (both still green and MERGEABLE, review-blocked only).
• Wave 2 (app/main.js chain) — all merged except feat: make the global record shortcut toggleable + documented (#396) #398, which is green and MERGEABLE after its third rebase. The serial constraint the wave existed for is gone now that refactor(main): extract settings-toggle IPC into settings-ipc.js (RFC #327 Phase 2.2) #364 and the rest have landed.
• Wave 3 — complete.
• The release itself went out as v0.6.5 with chore(release): v0.6.5 #431 + fix(release): write the macOS floor into latest-mac.yml, in Darwin form (#432) #436 + docs: finish the 14.4 floor alignment (website FAQ, release template) #434, exactly the cut line agreed here.

Still open and not covered by any wave, all green and MERGEABLE, all waiting on maintainer review only: #422, #426, #428, #440, #441, #444.

One ordering constraint outlived this issue and is recorded on the PRs themselves rather than here: #440 should land before #441. They overlap in app/package.json and AboutTab.tsx, and #440 introduces user-facing English strings that #441 has never seen — a straight merge would render them untranslated inside the German UI, and the locale completeness gate cannot catch a string that was never extracted.

Nothing else needs a tracker; the remaining PRs are independent.