feat(messages): upload attachments in the background after send

[kalvinnchau]

Sending a message with attachments now hands the upload to a session-owned background job, porting the base block/buzz desktop implementation: backgroundMediaUploadStore.ts and the send flow in useMentionSendFlow.ts. The composer clears immediately, so the user can keep typing or navigate away while files upload and the message publishes.

Behavior

• Captured destination: channel, thread root, reply parent, text, recipients and references are fixed at Send. A send never follows the user to another conversation or account.
• Progress: a pill above each composer shows Preparing / Uploading N% / Finishing with Cancel, using the block/buzz copy. Cancel stops the newest send.
• Ordering: Send stays disabled for a draft while its background send is pending, so a follow-up cannot overtake it. Typing, navigation and other conversations stay available.
• Failure and recovery: a failed or aborted upload (including access revocation and cache clearing) restores the text and files into the untouched draft and shows Upload failed: … or Message failed to send: …. If persistence cannot confirm restoration or the draft has since changed, the session retains the files and exposes a recovery retry without overwriting later work. A conflict asks the user to clear the later draft before retrying; accepted-cleanup state stays non-editable until resolved. Publication preparation failures return to the originating composer error state with its explicit retry control. Cancel restores silently.
• Reply target: a delayed publish clears the thread reply target only if it is still the one chosen at Send.
• Lifetime: jobs and files are owned by the relay session and held in memory, as in block/buzz. A disposed session never publishes. Publication goes through the existing outbox, so authorization, size limits and transport are unchanged.

Verification (current head 77ebacbdd1cfd61aa7ea74bba5419ea2596883d4)

• Preparation failure while the composer is absent now retains a session/conversation-keyed error. On return, the saved caption and adopted files remain visible with Retry message preparation. Recovery deferred by an active message edit survives an additional unmount. A successful refresh removes the error without sending, and a subsequent remount does not replay it.
• Fail-then-pass: the absent-composer mounted regression failed on c2255265 at the missing alert and passed after the fix. The edit-deferral remount regression and complete MessageComposer.test.tsx passed. Full vitest run at 77ebacbd: 535 files, 7,680 tests passed. Pre-commit formatting/lint and pre-push typecheck, related tests, and design guards passed. git diff --check passed. No browser journey was added or removed in this cycle: the new ordering is exercised in the colocated React mount; browser-only focus and transport coverage from the previous cycle remains unchanged.
• Previous c2255265 hosted CI failed in Chromium shard 1/6 at tests/browser/global-search.spec.mjs:340: a resting-pointer selection expected aria-selected=true but received false. This is a different surface, not yet reproduced or dismissed as unrelated. Current-head hosted CI is in progress; do not infer CI-green from local tests. No new human test is claimed.

Review readiness

• Peon found and we fixed a deferred-edit-unmount gap; independent rereview of 77ebacbd reports no actionable findings (minimalness/elegance/correctness: 9/9/9). The prior readiness claim at 617fe0e9 is obsolete. No current-cycle human test is claimed. PR remains open; no merge authorization is inferred.

[wesbillman]

Changes needed: three production recovery/accessibility defects are inline.

[P2] Public-material privacy: remove the “Originating Buzz channel” name and UUID from this public PR description. They expose internal coordination/workspace identifiers and are unnecessary for reviewing the feature; retain that provenance in the private workspace instead. This is separate from the production defects.

Star Lord automated source review via Wes’s account. Head 63ee15d91c887a3277816bb656f49020a2d8b210; base aa10599d43077b92f43655fe035e1bc0ae5e6afb. Source-only: no tests, builds, app execution, or live workflow validation. One CI snapshot shows CI required failing with WebKit journey shard 4/6 failing; not triaged. Description inspected; it contains no attached images.

[kalvinnchau]

🤖 The public PR description no longer contains the internal provenance identifiers. The PR remains draft while review and hosted checks finish.

[wesbillman]

Two recovery changes are still needed after composer remount (inline). The focused-Cancel repair and removal of internal PR provenance address those prior findings.

Star Lord automated follow-up source review via Wes’s account. Head f18b3aa2d74e43124e199c12b16b958416723286; base aa10599d43077b92f43655fe035e1bc0ae5e6afb. Source-only: no tests, builds, app execution, or live workflow validation. Hosted CI required passed in the snapshot; Windows validation was skipped. Public description inspected; no attached images.

[wesbillman]

Two recovery changes remain (inline). The two remount cases in the previous review are addressed; the remaining cases concern active message edits and preparation-error recovery after remount.

Star Lord automated follow-up source review via Wes’s account. Head 617fe0e9ded53876ad945a221dd798acf20354b2; base aa10599d43077b92f43655fe035e1bc0ae5e6afb. Source-only: no tests, builds, app execution, or live workflow validation. Browser shards were pending in the single hosted-check snapshot; Windows validation was skipped. Public description inspected; no attached images.

[wesbillman]

Preparation-error recovery is still incomplete when failure happens while the composer is unmounted (inline). The active-edit overwrite is repaired, and an already-remounted composer now receives the error.

Star Lord automated follow-up source review via Wes’s account. Head c22552655922a423bfaa08c89008d2e1c05e6e7f; base aa10599d43077b92f43655fe035e1bc0ae5e6afb. Source-only: no tests, builds or app execution. One hosted snapshot showed Chromium journey shard 1/6 failing and two WebKit shards pending; failure not triaged. Public description inspected; no attached images.

[kalvinnchau]

🤖 Addressed review comment 4191824333 at 77ebacbdd1cfd61aa7ea74bba5419ea2596883d4. A session/key-scoped missed preparation error is retained when upload/publication fails without a composer, replayed only while its restored draft still owns the saved revision, and consumed when shown (including deferred edit recovery). The mounted absent-composer case failed at the missing alert on c2255265, then passed; edit → failure → unmount → return and stale-error-after-refresh regressions pass. Full Vitest: 535 files / 7,680 tests; hooks passed; peon rereview found no actionable findings. Current-head hosted CI is still in progress. Previous-head Chromium shard 1/6 failed in global-search pointer selection and is not yet dismissed.

[wesbillman]

No changes requested for this follow-up. The remaining preparation-retry finding is addressed: failures survive an absent composer or deferred-edit unmount, replay only against the matching saved draft, and reach the explicit retry control.

Star Lord automated source review via Wes’s account; head 77ebacbdd1cfd61aa7ea74bba5419ea2596883d4, base aa10599d43077b92f43655fe035e1bc0ae5e6afb. Source-only review of the two-file repair and its collaborators: no tests or app execution; the CI snapshot still had browser shards running and Windows skipped, so runtime/focus acceptance remains unverified.