inspector-audio

Audio subsystem skill. Standalone — references below split by layer so the skill can grow new branches (per-codec, per-feature, per-platform) without bloating one doc.

Spans two arcs: the runtime path (bytes-on-disk → <audio>) and the upstream handoff (contributor source links → a reviewable reciters/<slug>/ folder). The native align pipeline (Requests-tab Align, incl. online playlist intake) or offline Katana extraction writes the bucket content; auto_detect reconciles it into the lifecycle. See references/extraction-intake.md.

Two corrections to hold (the docs used to lie about both)

1. The CDN tier is a same-origin 200/206 stream, not a 302. audio_source.resolve is three tiers — local Path → in-mem bytes → CDN — and the CDN tier is served by _stream_cdn same-origin with Access-Control-Allow-Origin: *. The old 302 was removed because it silenced <audio crossorigin> + the Web Audio kill-switch. There is no disk-cache tier.
2. No prefetch worker, no GC sweeper, _done.json not read at runtime. Bucket audio + peaks are written once — by the align pipeline's acquire_audio / split_audio HF jobs or by Katana extraction — and only read by the serving path. Nothing warms the bucket, nothing GCs it. The reconciler keys on the DB state row (AWAITING_ALIGNMENT), not the sentinel. "Audio missing on the bucket" is an extraction/upload problem.

Topology

[upstream]  intake/edit request (DB requests row)
              ALIGN  = delivery_states.state == 'awaiting_alignment'
              INTAKE = requests status='pending' AND slug IS NULL  ─► Requests tab: /api/admin/intake/<rid>/plan → /align
                       (plan → mint reciter+delivery+slug, seed AWAITING_ALIGNMENT → align run)   [extraction-intake.md]
        │
        ▼
chapter URL (CDN, in catalog/audio_manifest/<slug>.json)
        │
        ▼
[writers]  align pipeline HF jobs (acquire_audio / split_audio) or Katana extraction (audio_persist.py + upload_to_bucket.py)
                                   ──►  bucket: reciters/<slug>/audio/<ch>.mp3      (Xing TOC injected if VBR)
                                   ──►  bucket: reciters/<slug>/peaks/<ch>.json.gz  (slim int8, schema v3)
                                   ──►  bucket: reciters/<slug>/audio/_done.json    (written last; offline audit only — NOT read at runtime)
        │                                       (read-only at runtime — no fetch worker, no GC sweeper)
        ▼
[reconcile] auto_detect: reciters/<slug>/ appears for an AWAITING_ALIGNMENT slug (keys on the DB state row)
                                   ──►  reciter.alignment_completed  →  AWAITING_REVIEW      [extraction-intake.md]
        │
        ▼
[backend]  audio_source.resolve(reciter, url)            (services/audio/audio_source.py — 3 tiers, no disk cache)
              ├─► local Path  (bucket mount)         ─► send_file (Range/ETag/304/sendfile)
              ├─► in-mem bytes (local-dev no-mount)  ─► send_file(BytesIO)
              └─► cdn_url                            ─► _stream_cdn: same-origin 200/206 stream + ACAO:* (NOT a 302)
        │
        ▼
[wire]    /api/seg/audio-proxy/<reciter>?url=…                                  (chapter MP3)           routes/audio/proxy.py
          ↳ FE only wraps when the CDN host fails the CORS+Range probe OR the file's first frame carries a `Xing` tag OR an untagged file's sniffed frames miss the nominal byte rate (lib/playback/play-url.ts + mp3-header.ts — Chrome TOC-seeks `Xing` files ±seconds; bucket remux is `Info`); other CORS-ok CDN files are played DIRECT
          /api/seg/segment-clip/<reciter>?url=…&start_ms=…&end_ms=…            (VBR fallback, ffmpeg -ss/-t -vn)  routes/audio/clip.py
          /api/seg/peaks/<reciter>?chapters=…&h=…                              (slim int8 envelopes)   routes/segments/peaks.py
          /api/audio/surahs/<cat>/<src>/<slug>                                 (dashboard player {url,duration_ms})  routes/audio/metadata.py
        │
        ▼
[frontend]  AudioPort → <audio> → (optional) MediaElementAudioSourceNode → GainNode → ctx.destination
                          (per-tab port; segPort for Segments, shared dashPort for Dashboard + Timestamps)
                          (file-absolute ms outside, clip-relative inside; element-pool gapless via shadow-audio.ts)

Mode matrix

Knob	Dev (python3 inspector/app.py)	Deployed (HF Space, gunicorn)
Audio source	bucket if mounted, else CDN stream-through every play	bucket NFS-mounted, sendfile via Path
Bucket	hetchyy/quranic-inspector-bucket-dev	bucket-dev (dev Space) / bucket (prod Space)
ffmpeg HTTPS reachability	full network	full network — image compiled with --enable-openssl + file,pipe,http,https,tcp,tls
Web Audio kill-switch	only fires once ctx.state === 'running' (post-warmup)	same

VBR routing fork is per-chapter, not per-reciter. Decided by audio_meta.is_vbr_for_url server-side and the FE-shipped reciter_vbr_chapters list client-side. Same reciter can be CBR for chapter 1 and VBR for chapter 36.

Reference index

Reference	When to read	Key files
references/extraction-intake.md	The upstream handoff: align pipeline / Katana write reciters/<slug>/ + audio-manifest sidecar, auto_detect fires reciter.alignment_completed → AWAITING_REVIEW, the three request kinds, ALIGN / INTAKE queues, the online intake plan → mint → align flow and the intake.ingest() mint	services/segments/auto_detect.py, services/admin/intake.py, services/admin/intake_plan/, routes/admin/intake_plan.py, services/db/repo_requests.py, services/state/catalog.py, qua_shared/schemas/{intake_requests,intake_plan,catalog,state}.py
references/backend.md	Proxy/clip/metadata routes, the 3-tier audio-source resolver (no disk cache), manifest sidecar + reverse index, storage paths, MIME, config tunables, the /api/audio/surahs route, chapter_bitrate_kbps_for_reciter	routes/audio/{proxy,clip,metadata}.py, services/audio/{audio_source,audio_meta,audio_fetch}.py, services/storage/storage_paths.py, config.py
references/prefetch.md	Bucket audio + peaks read-only at runtime: sole offline writer (Katana), read primitives, what's gone (removed prefetch worker + GC sweeper), and the FE-side warmups that replaced the deleted prefetch util	services/audio/audio_fetch.py, routes/audio/proxy.py, routes/segments/peaks.py
references/peaks.md	Slim int8 v3 envelope, pack_slim/unpack_slim_envelope, route fan-out + LRU response cache (NOT evicted on save), shared b64ToInt8 decoder, peaks-view.ts shape adapter, history-peaks (now int8), backfill/audit	services/audio/{peaks,peaks_slim,op_peaks,peaks_history}.py, routes/segments/peaks.py, lib/utils/{peaks-view,peaks-decode}.ts
references/vbr.md	VBR-specific behavior — why Xing matters, Katana _ensure_xing, segment-clip fallback (-vn is load-bearing), AudioPort VBR reuse rule, VBR-only bug shapes	.local/extraction/segments/audio_persist.py::_ensure_xing, routes/audio/clip.py, lib/playback/audio-port.ts (VBR branch)
references/frontend.md	AudioPort (incl. adoptElement/prewarm/covers), AudioGraph kill-switch, warmup (in main.ts), AudioRange, shadow-audio element-pool gapless, the shared dashPort, peaks rendering, coordinate contract, cross-origin gotcha	lib/playback/, lib/utils/audio-warmup.ts, tabs/segments/utils/playback/, tabs/segments/stores/playback.ts
references/bugs.md	Common bug shapes — symptom → root → first probe, indexed by area	spans the whole stack
references/probes.md	Terminal recipes (ffprobe / ffmpeg / curl / hf bucket) + browser recipes (DevTools, Playwright)	—

Conventions

• File-absolute milliseconds outside AudioPort, clip-relative inside. Any caller writing el.currentTime directly is a bug.
• Bucket audio + peaks are written offline (Katana extraction audio_persist.py + upload_to_bucket.py) and only read at runtime — no in-Space fetch worker, no GC sweeper. Audio + peaks persist indefinitely. (FE-side warmup.ts / shadow-audio.ts are browser warmups — HTTP Range / element-pool — unrelated to the removed backend worker.)
• Manifest sidecar catalog/audio_manifest/<slug>.json is the single source of truth for VBR routing and chapter ↔ URL reverse lookup (cached as _audio_manifest + an O(1) _audio_manifest_url_index). Built offline by scripts/audio/probe_audio_meta.py. Never resolve chapter URLs through detailed.json — its per-entry audio field is "" post-migration-#5.
• Style across references: terse, table-first, file-path-anchored. When the live filesystem drifts, fix the matching reference — don't add a new layer.
• This skill is the ground truth for audio — by design there is no docs/reference/audio.md (docs/reference/README.md carves audio out to here). The reference docs only touch audio as thin pointers: the route map in architecture.md, the playback stores in frontend.md, audio manifests in catalog.md. Keep those thin and consistent with this skill; the depth lives here.

Bucket layout (audio-relevant)

catalog/audio_manifest/<slug>.json    # per-chapter URL + size + duration + bitrate_kbps + bitrate_mode
reciters/<slug>/audio/<chapter>.mp3   # Katana-written, Xing-injected if source is VBR
reciters/<slug>/audio/_done.json      # written last by extraction; offline audit/upload artifact only — NOT read at runtime
reciters/<slug>/peaks/<chapter>.json.gz   # slim int8 packed gzip (schema v3) — see references/peaks.md

Chapter keys: "1".."114" for by_surah, "<surah>:<ayah>" for by_ayah. Audio + peaks persist indefinitely (no GC).

What this subsystem deliberately doesn't do

No gapless / crossfade beyond the element-pool adopt (shadow-audio.ts, which is best-effort look-ahead, not sample-aligned crossfade). No HLS/DASH/adaptive bitrate. No DSP / EQ / loudness normalization (Web Audio gain is a kill-switch only). No on-the-fly transcoding beyond the Katana Xing inject + the segment-clip re-encode. No in-Space audio prefetch / re-bake / GC. rAF-bound boundary enforcement (~16 ms), not audio-clock-locked. Acceptable for review; not for sub-frame timing edits.