v2.4.0 — Automatic subtitle re-timing, on by default

Automatic subtitle re-timing, on by default (#359). subarr now post-processes every completed subtitle, extending cues that scroll past comfortable reading speed into the silent gap before the next line — so dialogue stays on screen long enough to actually read. Validated off-app across an 1801-subtitle corpus built from real subgen output: cues exceeding the critical 25 characters-per-second reading speed dropped from 22.9% to 5.2%, and unreadable sub-second "flash" cues fell 77%, with zero new overlaps (every extension is clamped to the available gap). It is idempotent and never shortens a cue or creates an overlap. On by default; opt out with SUBARR_RETIME_ENABLED=0.

Added

• Genuinely multilingual audio, represented honestly (#357). A film like As Bestas (Galician + Spanish + French in one track) is no longer flagged as a mislabelled English file. subarr recognises confident multi-language audio from per-chunk detection and marks it multilingual instead of raising a false "suspect" alarm. Non-linguistic tracks (constructed gibberish, silent-film scores) can be tagged zxx. Multilingual and zxx files skip the single-language override sent to subgen, so it self-detects per chunk.
• Correct or accept multilingual verdicts in Review (#406). Auto-classified multilingual files appear in Review with a 🌐 badge and their own filter tab: accept the detected language set as-is, or correct it with a multi-select picker. Settled verdicts never re-appear in the lane.
• subarr's own logs are visible (#157). The Logs page gains a subgen | subarr source switcher to tail subarr's internal log live, and the Health page shows a "Recent errors" panel of recent warnings and errors with expandable tracebacks. The opt-in library audio-audit now reports run-level crashes on the Health roster like the other background tasks. A new SUBARR_DEBUG=1 knob raises logging to DEBUG for local troubleshooting. All local-only — nothing new leaves your box.

Changed

• Audio languages are now canonical 2-letter ISO-639-1 (#358). The audio-language model and the Review picker use 2-letter codes end to end (e.g. en, ko), normalising on read and write so a mixed 2-/3-letter history resolves consistently.

Fixed

• Graceful degradation when an integration client is recycled mid-request (#392, #393). A Bazarr/Sonarr/Radarr/Plex/Ollama call that raced a client rebuild — or hit a client closed mid-flight — is now translated into a clean, retryable error instead of surfacing a raw RuntimeError.
• Manual "fill this gap" on a movie uploads straight to Bazarr (#394). The manual movie-gap action now uploads the generated subtitle directly to the owning Bazarr instance, matching the episode path, instead of only nudging its scan-disk task.
• Dependency bumps: nvidia-ml-py, actions/setup-python, @playwright/test (#398, #399, #400).