processes/subtitles: STOPGAP-SUBS.md tracker for v3.5 → v4 cross-ref

Owner accepted Sassy the Sasquatch S01 v3.5 YouTube-auto-CC subs as
'85 % acceptable, fine, not great' but flagged them for v4 WhisperX
rebuild. Adds a single worklist file (STOPGAP-SUBS.md) so every show
that ships via the v3.5 path gets logged for the eventual v4 sweep
instead of being silently forgotten.

Sassy run log gets a STOP-GAP banner at the top pointing to the new
worklist. README.md gets a stop-gap-exception note alongside the
STYLE.md hard-prereq paragraph. ROADMAP H5 now points at the worklist
file as the canonical source of which shows v4 needs to regenerate.

ROADMAP.md

@@ -28,7 +28,7 @@ Last revised: **2026-05-08**
 | H2 | GPU transcode (nvidia driver kernel module + container toolkit + SecureBoot signing) | L | **owner sudo + reboot** |
 | H3 | Apply `bin/force-english-all-users.sh` (German Play button breaks UX for non-English browsers) | S | none — owner runs |
 | H4 | Backup `/home/docker/jellyfin/config/` off-host (no automated backup yet) | M | strategy decision |
-| H5 | **v4 subtitle path: WhisperX large-v3 on friend RTX 4080 node**. Regenerate Sassy + Big Lez Saga + Donny & Clarence + Mike Nolan with proper-noun prompts (replaces v3.5 YT auto-CC stop-gap). New helper at `processes/subtitles/lib/sub-whisperx-fetch.py`. WhisperX install on 100.64.0.3 (per memory `project_friend_gpu.md`, currently offline 2d); per-show prompt yaml at `processes/subtitles/prompts/<show>.yaml` (recurring proper nouns). Expected 4–6 % WER vs ~12 % for YT auto-CC; restores STYLE.md "best quality" bar. See `processes/subtitles/runs/sassy-the-sasquatch.md` for context. | M | friend node back online + WhisperX setup |
+| H5 | **v4 subtitle path: WhisperX large-v3 on friend RTX 4080 node**. Worklist = `processes/subtitles/STOPGAP-SUBS.md` (currently Sassy 5/5, will grow as more Big Lez universe shows ship via v3.5). Replaces v3.5 YT auto-CC stop-gap with proper-noun-prompted transcription. New helper at `processes/subtitles/lib/sub-whisperx-fetch.py`. WhisperX install on `100.64.0.3` (per memory `project_friend_gpu.md`, currently offline 2d); per-show prompt yaml at `processes/subtitles/prompts/<show>.yaml`. Expected 4–6 % WER vs ~12 % for YT auto-CC; restores STYLE.md "best quality" bar. | M | friend node back online + WhisperX setup |
 
 ## 🟨 Open — Medium value
 

processes/subtitles/README.md

@@ -13,6 +13,11 @@ how Jellyfin and the OpenSubtitles plugin work together lives in
 > AI / no Forced), best-quality release. The picker logic in v1/v2/v3
 > mirrors that bar; if a step would violate it, stop and ask before
 > downloading.
+>
+> Stop-gap exception: when the only available source is the v3.5 YouTube
+> auto-CC path (lowercase, censored, mangled names), ship the sub but
+> **add the show to [`STOPGAP-SUBS.md`](STOPGAP-SUBS.md)** so v4 WhisperX
+> picks it up later.
 
 ---
 

processes/subtitles/STOPGAP-SUBS.md

@@ -0,0 +1,46 @@
+# Stop-gap subs — pending Whisper cross-ref
+
+Shows whose current subtitles ship from a path that explicitly violates
+[`STYLE.md`](STYLE.md). Quality is "acceptable, not great" (~85 %). When
+v4 WhisperX (ROADMAP H5) lands on the friend RTX 4080 node, **regenerate
+every show on this list** with proper-noun-prompted transcription and
+replace the sidecars in place. Keep this file as the v4 worklist.
+
+## Active stop-gaps
+
+| Show | Eps subbed | Source path | Why stop-gap | Owner verdict | Logged |
+|---|---|---|---|---|---|
+| Sassy the Sasquatch (2022) | S01 5/5 | v3.5 YouTube auto-CC | lowercase, no punctuation, names mangled (`Sassy → sasha`), profanity = `[ __ ]` | "85 % the way there, acceptable, fine" — keep until v4 | 2026-05-10 |
+
+## When more Big Lez universe shows ship via v3.5
+
+Same channel hosts these — when subbed via the v3.5 yt-dlp path, append
+to the table above:
+
+- The Donny & Clarence Show (2024)
+- The Big Lez Saga (2022)
+- The Mike Nolan Show (2016) — but **try the YT "COMPLETE SEASON | SUBTITLES"
+  upload first** for hand-typed CCs before falling back to auto-CC
+
+## v4 WhisperX rebuild plan
+
+When the friend node (`100.64.0.3`, per memory `project_friend_gpu.md`) is
+back online:
+
+1. Install WhisperX on the node (CUDA 12 + cuDNN 9 + faster-whisper +
+   pyannote VAD).
+2. For each show in the table above, write
+   `processes/subtitles/prompts/<show>.yaml` with the recurring proper
+   nouns the YT auto-CC mangled.
+3. Run `lib/sub-whisperx-fetch.py` (TBD, ROADMAP H5) per show. Each
+   episode: pull mkv → ffmpeg extract 16k mono wav → WhisperX large-v3
+   with `--initial_prompt` from the yaml → SRT → SSH push to nullstone
+   with library filename, **overwriting the v3.5 sidecar in place**.
+4. Tick off the row from the table; move it to a "Cleared via v4" archive
+   section below this one (kept as record).
+5. Library scan; verify Jellyfin still reports 1 external eng sub stream
+   per ep (no dupes from v3.5 + v4 stacking).
+
+## Cleared via v4 (archive)
+
+(empty — populate as v4 rebuilds land)

processes/subtitles/runs/sassy-the-sasquatch.md

@@ -1,5 +1,9 @@
 # Subtitle run — `Sassy the Sasquatch (2022)`
 
+> ⚠ **STOP-GAP — needs v4 WhisperX cross-ref.** Owner accepted current
+> subs as "85 %, acceptable" but tracked for full rebuild when v4 lands
+> (ROADMAP H5). See [`STOPGAP-SUBS.md`](../STOPGAP-SUBS.md).
+
 Recipe version: v3.5 — YouTube auto-CC via yt-dlp + cleaner (v4 WhisperX planned, see ROADMAP)
 Run date: 2026-05-10
 Operator: Claude Code @ onyx session, ai-lab cwd