Cancel export and media download strategy improvement

[TimRl]

Cancel export and media download strategy improvement

Frontier: genesis-ai-dev/frontier-authentication#34

Pinned Project: 1050-cancel-export-and-media-download-strategy-improvement

#997

Summary

Adds the ability to cancel an in-progress project export (cleaning up any
partial output), makes whole-project audio export resilient to transient
LFS/network failures, and overhauls how unresolved/missing audio is surfaced
to the user — including a categorized, expandable, clickable issues list, a
targeted "retry failed" action, and clearer real-time progress.

Also tightens the media-download-strategy switch (clear "Calculating…" feedback
and diagnostic logging into why some media is retained).

Changes

Export cancellation

• Cancel button on the export progress view; threads a CancellationToken
  through every exporter (audio, plaintext, USFM, HTML, VTT/SRT, XLIFF, CSV/TSV,
  rebuild, backtranslations).
• Aborts in-flight LFS downloads via an AbortSignal bridged from the token.
• Deletes the partial (unique) export folder on cancel and shows a terminal
  "Cancelled — partial output removed" state with a Close button.
• Re-entrancy guard prevents starting a second export while one is running;
  closing the export tab aborts the run.

Resilient audio downloads

• New downloadLfsWithRetry: up to 4 attempts with exponential backoff + jitter.
• Retries only transient errors (5xx, 429, timeouts, ECONNRESET, fetch failed,
  etc.); fails fast on permanent ones (4xx, corrupt pointer) and never retries a
  user-initiated abort.

Export issues UX

• "X audio file(s) exported, Y could not be resolved" now expands into a
  categorized, collapsible list with human-readable labels/descriptions and an
  "Expand all / Collapse all" toggle.
• Each issue row is clickable and deep-links to the exact cell in the editor
  (cell id + codex path threaded through the reporter chain).
• New "selected take missing but other takes available" warning category
  (recoverable) split out from hard "could not be resolved" errors.
• "Retry failed" action re-runs only the recoverable failed cells and merges
  recovered audio back into the existing export folder.
• Unified amber accent across all issue groups.

Pre-flight (Step 1) + real-time stats

• Cells whose selected take is flagged isMissing are surfaced as "possibly
  missing"; cells whose only remaining takes are all missing are reclassified as
  "without audio" (consistently in pre-flight and export).
• Audio stat pills update in real time via a debounced .codex file watcher;
  the cell-list popover stays open when navigating to a cell.

Progress display

• Count shown once (in the title); the secondary line shows the actual item
  (cell label while writing) instead of repeating the .codex filename during
  concurrent downloads.

Media download strategy switch

• Disabled "Calculating…" state while counting media files before the prompt.
• Diagnostic tally in replaceFilesWithPointers explaining why each file was
  retained (persisted / unsynced / already-pointer) without changing behavior.

Testing Checklist

Export cancellation

• Start an audio export, cancel mid-download → run stops, partial folder is deleted, "Cancelled — partial output removed" state shows with a Close button.
• Close the export tab mid-export → the run aborts.
• Attempt to start a second export while one is running → blocked.
• Cancel during a retry backoff → aborts promptly (doesn't wait out the delay) and surfaces as cancellation, not an error.

Resilient audio downloads (retry/backoff)

• Export against a flaky/500-prone server → transient failures (5xx, 429, timeouts, ECONNRESET, fetch failed) auto-recover via retry.
• Permanent errors (404/401/403, corrupt pointer) fail fast without retrying.

Export issues UX

• Trigger unresolved cells → issues list groups them by category with human-readable labels/descriptions.
• "Expand all / Collapse all" toggle works and stays readable in hover/selected states.
• Issue rows deep-link to the correct cell in the editor.
• "Selected take missing but other takes available" appears as a recoverable warning, separate from hard "could not be resolved" errors.
• "Retry failed" → recovered cells drop off, still-failing ones remain, recovered audio lands in the original export folder.
• All issue groups share the unified amber accent.

Pre-flight (Step 1) + real-time stats

• Cells with a selected isMissing take show as "possibly missing".
• Cells whose only remaining takes are all missing are classified as "without audio" (consistent in pre-flight and export).
• Edit a cell's audio on Step 1 → stat pills update live; the cell-list popover stays open when navigating to a cell.

Progress display

• Count shows once (in the title); the secondary line shows the actual item (cell label while writing), not the .codex filename during concurrent downloads.

Media download strategies — audio pointerization (replaceFilesWithPointers)

• Default stream-only run reverts a synced audio take (pointers/ holds the pointer, files/ holds real bytes) to a pointer.
• Default stream-only run preserves an unsynced audio take (real bytes in both files/ and pointers/) — bytes untouched.
• restrictToAudio pointerizes audio only, leaving video untouched.
• restrictToVideos pointerizes video only, leaving audio untouched.
• Persisted (user-saved) audio is protected on automatic cleanup; ignorePersisted frees it.
• Mixed audio + video, default run: both synced files pointerized, both unsynced files preserved (regression for the "always compute videoRefs" refactor).

Media download strategies — countSyncedDeletableAudioFiles

• Counts synced audio (pointer in pointers/, real bytes in files/).
• Excludes unsynced audio (no/real-bytes pointer).
• Excludes persisted (allowlisted) audio.
• Excludes files that are already pointer stubs.
• Excludes video files.
• Counts ambiguous .webm recordings as audio (not referenced by a notebook videoUrl).

Media download strategies — counting & stubs

• countDownloadedMediaFiles counts real downloaded media, ignores pointer stubs.
• removeFilesPointerStubs (less-restrictive switch, audio): removes audio pointer stubs so they re-download; preserves real downloaded audio bytes.

Media download strategies — applyMediaStrategy transition matrix

• auto-download → stream-only (default): synced media freed, unsynced preserved.
• auto-download → stream-only (keep-video): local videos allowlisted and kept; rest freed.
• auto-download → stream-only (free-all): everything freed incl. saved videos; video allowlist cleared.
• auto-download → stream-and-save (granular keepVideo/keepAudio combinations).
• auto-download → stream-and-save (keepFilesOnStreamAndSave === false): explicit free incl. saved videos.
• stream-only → auto-download: pointer stubs removed + pending swap downloads processed.
• stream-only → stream-and-save: preserve-videos behavior.
• stream-and-save → stream-only / auto-download: restrictive vs less-restrictive outcomes.
• Same strategy (no change) is a no-op; forceApply re-runs anyway.

Media download strategies — "Calculating…" UX (project.mediaStrategyCalculating)

• Switch media strategy on a large project → disabled "Calculating…" state shows while counting.
• Message sent (calculating: true) before counting media files.
• Cleared (calculating: false) on cancel and before persisting the result.

Frontier-authentication — AbortSignal for LFS downloads

• Cancel an export while LFS objects are downloading → downloads abort promptly and surface as cancellation, not errors.
• Two callers request the same OID; one aborts → the other still completes (refcounted abort).
• All waiters for an OID abort → the shared fetch is actually cancelled (no lingering bandwidth use).
• An already-aborted signal never joins/starts a shared download.
• Normal (uncancelled) downloads and the dedup fast-path still work.
• Aborted requests propagate AbortError untouched (callers can distinguish cancel vs failure).

[TimRl]

/build

[github-actions]

Pre-release: 0.28.0-pr1050-548374d https://github.com/genesis-ai-dev/codex-editor/releases/tag/0.28.0-pr1050-548374d

[LeviXIII]

Works well for me. There was one issue that was caught unrelated to the PR but that was made into a separate issue.