xenarathon/DubTitlerr
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

V2: Models, Ops & Remaining Polish (llama.cpp backend, phonetic match, CUDA-gating fix, signs/font audit) #2

Merged
xenarathon merged 45 commits from feat/v2-models-ops into main 2026-07-24 18:58:29 -04:00
Conversation 0 Commits 45 Files changed 42 +2027 -103
xenarathon commented 2026-07-24 18:51:57 -04:00
Owner
Copy link

Summary

Implements the V2: Models, Ops & Remaining Polish spec (specs/v2-models-ops/) — the ~35
review items deferred from V1. Executed and reviewed unit-by-unit (6 units, one whole-branch review),
grouped by the plan's Phase A–D. Depends on V1 (merged to main first).

Phase A — Models & Accuracy (repair.py, glossary.py, generate.py)

  • llm() now dispatches on REPAIR_BACKEND (ollama | llamacpp); a llamacpp backend posts to
    REPAIR_LLAMACPP_URL (/completion schema). Split connect/read timeouts (REPAIR_TIMEOUT_CONNECT
    /_READ) via stdlib http.client; per-line latency_ms in the repair CSV; repair-summary.json.
  • Two-pass repair via REPAIR_MODEL_SECONDARY (no-op when equal to primary).
  • Tier-4 phonetic glossary matching (jellyfish.metaphone), gated on non-English tokens, degrades
    gracefully if jellyfish is absent.
  • Per-word word_probs in dubtitles.conf.json (optional) + a low-prob-word targeting gate in is_target().
  • WHISPER_AUDIO_FILTER highpass+compand pre-filter in extract_wav(). See "⚠️ Deploy decision" below.
  • A9 (large-v3-turbo) deferred — needs a real GTX 1060 bench; WHISPER_MODEL is env-configurable,
    default stays large-v3.

Phase B — Shell & Ops (shell scripts, data/, common.py, plex_refresh.py)

  • Deprecation headers on the legacy orchestrators; set -e in gen_loop.sh; removed the self-healing
    apt-get from merge_pass.sh (fail loud instead).
  • Consolidated the EXTRA_DIRS exclusion list into a single data/extras.txt, read by both
    common.load_extras() and shell/lib.sh::extras_grep_pattern (all 4 consumers), with inline fallbacks.
  • .gitignore pipeline artifacts; plex_refresh.py env hardening (also cleared the last 2 tree-wide ruff errors).

Phase C — Python Polish (mux.py, repair.py, generate.py, glossary_verify.py, ordering.py, reflow.py, …)

  • mux: partners() inode cache, explicit HL_ROOTS default, dropped redundant samefile(), removed
    the verify() half-size heuristic (duration-tolerance is the sole truncation canary — a truncation
    test was added), identify() reuse.
  • repair: prompt-injection XML-wrap of the fansub reference; whole-term glossary-cap (no mid-name truncation).
  • generate: per-show lastrun.json; CUDA error gating fixed — gates on isinstance(RuntimeError)
    instead of substring-matching "cuda", so a non-GPU error no longer permanently poisons an episode
    (it clears .fail and logs crash.json for retry).
  • glossary_verify.adjudicate() parallelized (VERIFY_WORKERS, default 4); ordering.read_start()
    default fixed; anime_library.sh --dry-run; COMMON/BLOCKLIST moved to data/ with inline fallback;
    cross-file os.chown-failure logging; Authorization sections filled in the older specs; reflow readability.

Phase D — Signs/songs + font audit (dub_signs_merge.py, mux.py)

  • Diagnostic logging: style-name collisions, WrapStyle differences, resolution mismatch (WARN-ONLY —
    no track drop / no coordinate transform, deferred to V3), and a forced ScaledBorderAndShadow: yes.
  • mux.verify() font-attachment audit — compares source vs muxed font counts (returns
    font-count-mismatch), strictly additive (never a false "ok").

Notable decisions

  • ordering.read_start() (C4) — the literal task said "no priority file → return 0", but that would
    have silently dropped the existing SEASON_START env override. Kept SEASON_START working and made
    the "watch-order disabled" log fire only when the result truly resolves to 0. (User-adjudicated.)
  • verify() half-size removal (C16) — confirmed the duration-tolerance check runs unconditionally on
    the sole success path before removing the size proxy; added a truncated-output regression test.
  • D2 font audit — the task's literal code targeted the wrong mkvmerge -J field (tracks w/
    type=="attachments"); corrected to the real top-level attachments array (else it'd be a permanent no-op).
  • Fixed a pre-existing V1 bug: Dockerfile.builder never COPY'd common.py (added in V1) → the
    container would ImportError at startup. Now copies common.py, data/, and shell/. A rebuild of
    the builder image is required on deploy regardless of this PR.

⚠️ Deploy decision required — WHISPER_AUDIO_FILTER

The A8 default applies a highpass=f=80,compand=... filter to every new transcription going forward
(already-stamped episodes are untouched). This is the spec's intent but is unverified on real audio.
Please make a go/no-go before rebuilding. To opt out with zero code change, set WHISPER_AUDIO_FILTER="".

Deploy notes

  • Rebuild Dockerfile.builder (closes the V1 common.py ImportError; ships data/+shell/+jellyfish).
  • New env vars (all default to prior behavior except WHISPER_AUDIO_FILTER) need DockHand per-stack
    wiring only if you want to change defaults: REPAIR_BACKEND, REPAIR_LLAMACPP_URL,
    REPAIR_MODEL_SECONDARY, REPAIR_TIMEOUT_CONNECT/READ, VERIFY_WORKERS, WHISPER_AUDIO_FILTER,
    SEASON_PRIORITY_FILE, HARDLINK_ROOTS.
  • Hardware-gated items PENDING manual verification (not tested here): A9 large-v3-turbo bench, the live
    llama.cpp backend (192.168.1.232:8080), and a real container_run.sh end-to-end sweep.

Test plan

  • pytest -q → 193 passed (baseline 122 at V2 start; +71 across the 6 units). Run via
    rtk proxy python -m pytest tests/ (this shell's RTK hook otherwise mangles pytest output).
  • ruff check . → 0 errors whole-tree (V1 had left 2 in plex_refresh.py; B11 cleared them).
  • LLM/HTTP paths tested with mocked transport (routing + parsing); verify() font/truncation paths and
    the CUDA-gating retry-vs-poison branches have dedicated tests.

Known minor follow-ups (non-blocking)

mux.partners()/HL_ROOTS/DELETE_BROKEN are dead code on the current hardlink-safe path (polished but
unused); llm_llamacpp ignores its model arg, so two-pass under REPAIR_BACKEND=llamacpp re-runs the same
model (wasted compute, not incorrect). Both documented for a later pass.

🤖 Generated with Claude Code

## Summary Implements the **V2: Models, Ops & Remaining Polish** spec (`specs/v2-models-ops/`) — the ~35 review items deferred from V1. Executed and reviewed unit-by-unit (6 units, one whole-branch review), grouped by the plan's Phase A–D. Depends on V1 (merged to main first). **Phase A — Models & Accuracy** (repair.py, glossary.py, generate.py) - `llm()` now dispatches on `REPAIR_BACKEND` (`ollama` | `llamacpp`); a `llamacpp` backend posts to `REPAIR_LLAMACPP_URL` (`/completion` schema). Split connect/read timeouts (`REPAIR_TIMEOUT_CONNECT` /`_READ`) via stdlib `http.client`; per-line `latency_ms` in the repair CSV; `repair-summary.json`. - Two-pass repair via `REPAIR_MODEL_SECONDARY` (no-op when equal to primary). - Tier-4 phonetic glossary matching (`jellyfish.metaphone`), gated on non-English tokens, degrades gracefully if `jellyfish` is absent. - Per-word `word_probs` in `dubtitles.conf.json` (optional) + a low-prob-word targeting gate in `is_target()`. - `WHISPER_AUDIO_FILTER` highpass+compand pre-filter in `extract_wav()`. **See "⚠️ Deploy decision" below.** - A9 (large-v3-turbo) **deferred** — needs a real GTX 1060 bench; `WHISPER_MODEL` is env-configurable, default stays `large-v3`. **Phase B — Shell & Ops** (shell scripts, data/, common.py, plex_refresh.py) - Deprecation headers on the legacy orchestrators; `set -e` in `gen_loop.sh`; removed the self-healing `apt-get` from `merge_pass.sh` (fail loud instead). - Consolidated the EXTRA_DIRS exclusion list into a single `data/extras.txt`, read by both `common.load_extras()` and `shell/lib.sh::extras_grep_pattern` (all 4 consumers), with inline fallbacks. - `.gitignore` pipeline artifacts; `plex_refresh.py` env hardening (also cleared the last 2 tree-wide ruff errors). **Phase C — Python Polish** (mux.py, repair.py, generate.py, glossary_verify.py, ordering.py, reflow.py, …) - mux: `partners()` inode cache, explicit `HL_ROOTS` default, dropped redundant `samefile()`, removed the `verify()` half-size heuristic (duration-tolerance is the sole truncation canary — a truncation test was added), identify() reuse. - repair: prompt-injection XML-wrap of the fansub reference; whole-term glossary-cap (no mid-name truncation). - generate: per-show `lastrun.json`; **CUDA error gating fixed** — gates on `isinstance(RuntimeError)` instead of substring-matching `"cuda"`, so a non-GPU error no longer permanently poisons an episode (it clears `.fail` and logs `crash.json` for retry). - `glossary_verify.adjudicate()` parallelized (`VERIFY_WORKERS`, default 4); `ordering.read_start()` default fixed; `anime_library.sh --dry-run`; COMMON/BLOCKLIST moved to `data/` with inline fallback; cross-file `os.chown`-failure logging; Authorization sections filled in the older specs; reflow readability. **Phase D — Signs/songs + font audit** (dub_signs_merge.py, mux.py) - Diagnostic logging: style-name collisions, `WrapStyle` differences, resolution mismatch (WARN-ONLY — no track drop / no coordinate transform, deferred to V3), and a forced `ScaledBorderAndShadow: yes`. - `mux.verify()` font-attachment audit — compares source vs muxed font counts (returns `font-count-mismatch`), strictly additive (never a false `"ok"`). ## Notable decisions - **`ordering.read_start()` (C4)** — the literal task said "no priority file → return 0", but that would have silently dropped the existing `SEASON_START` env override. Kept `SEASON_START` working and made the "watch-order disabled" log fire only when the result truly resolves to 0. (User-adjudicated.) - **`verify()` half-size removal (C16)** — confirmed the duration-tolerance check runs unconditionally on the sole success path before removing the size proxy; added a truncated-output regression test. - **D2 font audit** — the task's literal code targeted the wrong `mkvmerge -J` field (`tracks` w/ `type=="attachments"`); corrected to the real top-level `attachments` array (else it'd be a permanent no-op). - **Fixed a pre-existing V1 bug**: `Dockerfile.builder` never `COPY`'d `common.py` (added in V1) → the container would `ImportError` at startup. Now copies `common.py`, `data/`, and `shell/`. **A rebuild of the builder image is required on deploy regardless of this PR.** ## ⚠️ Deploy decision required — WHISPER_AUDIO_FILTER The A8 default applies a `highpass=f=80,compand=...` filter to **every new transcription** going forward (already-stamped episodes are untouched). This is the spec's intent but is **unverified on real audio**. **Please make a go/no-go before rebuilding.** To opt out with zero code change, set `WHISPER_AUDIO_FILTER=""`. ## Deploy notes - **Rebuild `Dockerfile.builder`** (closes the V1 `common.py` ImportError; ships `data/`+`shell/`+`jellyfish`). - **New env vars** (all default to prior behavior except WHISPER_AUDIO_FILTER) need DockHand per-stack wiring only if you want to change defaults: `REPAIR_BACKEND`, `REPAIR_LLAMACPP_URL`, `REPAIR_MODEL_SECONDARY`, `REPAIR_TIMEOUT_CONNECT/READ`, `VERIFY_WORKERS`, `WHISPER_AUDIO_FILTER`, `SEASON_PRIORITY_FILE`, `HARDLINK_ROOTS`. - Hardware-gated items **PENDING manual verification** (not tested here): A9 large-v3-turbo bench, the live llama.cpp backend (192.168.1.232:8080), and a real `container_run.sh` end-to-end sweep. ## Test plan - `pytest -q` → **193 passed** (baseline 122 at V2 start; +71 across the 6 units). Run via `rtk proxy python -m pytest tests/` (this shell's RTK hook otherwise mangles pytest output). - `ruff check .` → **0 errors whole-tree** (V1 had left 2 in `plex_refresh.py`; B11 cleared them). - LLM/HTTP paths tested with mocked transport (routing + parsing); `verify()` font/truncation paths and the CUDA-gating retry-vs-poison branches have dedicated tests. ## Known minor follow-ups (non-blocking) `mux.partners()`/`HL_ROOTS`/`DELETE_BROKEN` are dead code on the current hardlink-safe path (polished but unused); `llm_llamacpp` ignores its model arg, so two-pass under `REPAIR_BACKEND=llamacpp` re-runs the same model (wasted compute, not incorrect). Both documented for a later pass. 🤖 Generated with [Claude Code](https://claude.com/claude-code) 
Renames the Ollama call to llm_ollama(prompt, model=None) (unchanged request
shape/response parsing), adds llm_llamacpp(prompt, model) for llama.cpp's
/completion schema, and makes llm(prompt, model=None) dispatch between them
on REPAIR_BACKEND (default ollama, byte-for-byte the old behavior). New
REPAIR_LLAMACPP_URL env var. Live llama.cpp integration not reachable from
this environment; covered here with mocked urllib.request.urlopen only.
Adds REPAIR_TIMEOUT_CONNECT (default 10) and REPAIR_TIMEOUT_READ (default
120). urllib.request.urlopen only exposes one timeout for connect+every
read, so the transport moves to a small http.client-based _post_json():
the connect timeout is set on the connection, the read timeout on the
socket right after connect(). llm_ollama/llm_llamacpp now go through it.

process() times each LLM call with time.monotonic() and adds a latency_ms
column to dubtitles.repair.csv.
Adds REPAIR_MODEL_SECONDARY (default REPAIR_MODEL, no-op when equal). After
the first LLM pass produces an accepted repair, _needs_secondary_check()
decides whether to re-send the same prompt to the secondary model: length
ratio < 0.6 or > 1.5, or the output contains a glossary name not present in
the original line. Per the corrected spec, the name-appeared condition is
expected to fire on ~every successful name repair (re-verify all name
changes with the stronger model), not just a rare-case fraction.
process() now writes <stem>.dubtitles.repair-summary.json (via out_for(),
same as the srt/csv) with targets, repaired, skipped_no_ref, mean/p95
latency (nearest-rank _p95(), no numpy), model, model_secondary and a
repaired_lines detail list. skipped_no_ref counts targets with no fansub
anchor (previously silently dropped).
_phonetic_match(token, names) fires after the guarded-fuzzy tier when a
non-English token's Metaphone code equals a glossary name's code (e.g.
"spondum" -> "Spandam", SPNTM==SPNTM, which the fuzzy cutoff rejects at
~0.71 similarity). Import is wrapped in try/except ImportError so a
missing jellyfish degrades to the existing 3-tier behavior.

Reuses the fuzzy tier's _one_indel() exclusion: a one-char insert/delete
mishear (e.g. "spandm"/"Spandam") also Metaphones identically but is the
same risky-edit case the fuzzy tier already defers to the LLM repair
stage -- tier 4 must not bypass that guard via a different matching path
(caught by the existing test_correct_guarded_fuzzy_refuses_one_char_indel).
jellyfish>=1.0 backs the A4 phonetic-match tier in glossary.py. Pure
Python (~50KB), no C deps -- added to project dependencies and to
Dockerfile.builder's pip install line alongside pysubs2.
_card_word_probs(card, words) joins a card's [start, end] window against
the full per-episode word list (collected via getattr(w, "probability",
1.0) in the whisper-adaptation loop) by time overlap, since reflow's Card
doesn't retain which whisper words it was built from. Computed on each
kept card's PRE-collapse boundaries so a later hallucination.collapse_runs()
merge (which keeps run[0]'s text verbatim) also keeps the word_probs that
actually correspond to that text, rather than the widened merged window.

Optional/backward-compatible: only added to a conf.json row when non-empty,
same pattern as the existing "flag" field.
has_low_prob_word(c) returns True if any value in c.get("word_probs", [])
(V2 A6) is below LOW_WORD_PROB (0.25) -- catches a single badly-mis-heard
word hiding inside a card whose avg_logprob otherwise looks fine. Added
as an OR condition in is_target() alongside the existing avg_logprob
check. Missing/empty word_probs (conf.json predating A6) -> False,
backward-compatible with the existing avg_logprob/name_suspect gate.
A8: extract_wav() appends -af "$WHISPER_AUDIO_FILTER" to the ffmpeg
command. Default is a highpass(80Hz) + compand filter (spec's Data
contracts value), tuned to clean up noisy/quiet dub audio before
transcription. Empty string disables it, reproducing the exact pre-A8
ffmpeg command (backward-compat opt-out).

BEHAVIORAL NOTE: since extract_wav() previously applied no filter at all,
this default changes audio extraction for every NEW transcription from
here on (already-generated .srt/.conf.json sidecars are untouched --
process()'s idempotency stamps mean only not-yet-transcribed episodes see
it). This is the spec's stated intent for the accuracy phase, not an
accident; flagged here per the task brief instead of silently shipped.

A9: WHISPER_MODEL was already env-configurable (no code change needed).
large-v3-turbo has NOT been bench-tested in this dev environment (no GPU
available here) -- recorded as PENDING manual test on the real GTX 1060
6GB server in both the module docstring and an inline comment. Default
stays large-v3; no turbo benchmark result is claimed or fabricated.
anime_library.sh/all_seasons.sh reload the Whisper model per show; merge_watcher.sh
spins up a fresh container every interval just to run merge_pass.sh. container_run.sh
does both more efficiently (resident model, in-process merge loop). No functional
change — comment-only, scripts retained for reference.
The old Dockerfile only builds the signs+dub merge step; Dockerfile.builder is the
full transcribe+repair+merge+mux pipeline and is the maintained path. README quick
start now builds/runs the builder image instead of dub-signs-merge.
mine/verify already used `|| echo` fallthroughs; the generate.py call did not, so a
crash there would previously just fall through to rc=$? and the stall-detection
logic. Under set -e that same crash would abort the whole container before rc=$?
ever ran. Switch it to `&& rc=0 || rc=$?` (verified with a standalone repro) so the
real exit code is still captured and crash-resume keeps working, while any other,
un-tolerated failure now fails loud instead of silently limping on.
The image should already have ffmpeg/mkvmerge/pysubs2 baked in (Dockerfile.builder);
silently apt-get-installing them on every pass masked a misbuilt image and added
unnecessary network calls to a hot path. Now exits 1 with a clear FATAL message
naming the missing tool.
Reproduces the 9 dir names currently hardcoded in common.py::EXTRA_DIRS exactly.
Feeds both common.load_extras() (Python, next commit) and shell/lib.sh's
extras_grep_pattern() so generate.py, mine_glossary.py, merge_pass.sh, and
post_show.sh all read the same list instead of 4 independent copies.
load_extras(path="data/extras.txt") reads the consolidated data file into a
lowercased set; falls back to the pre-consolidation hardcoded set on OSError so
dev checkouts / older images without the data file still work. common.EXTRA_DIRS
now comes from load_extras() at import time -- verified equal to the old hardcoded
set (same 9 entries).
Reads data/extras.txt and emits a grep -iE alternation pattern. Verified
byte-for-byte identical to the pre-consolidation inline regex in merge_pass.sh/
post_show.sh: '(Behind The Scenes|Deleted Scenes|Featurettes|Interviews|Scenes|
Shorts|Trailers|Other|Extras)'. Missing/unreadable file degrades to '()' (an
always-non-matching pattern); callers (B9) supply their own inline fallback.
Both previously imported the hardcoded common.EXTRA_DIRS constant directly; now both
call load_extras() themselves (still via common.py, still falling back to the same
hardcoded set) so all Python consumers agree on data/extras.txt as the source. Split
into a plain `from common import load_extras` + separate `EXTRA_DIRS = load_extras()`
assignment rather than a semicolon-joined one-liner -- the latter tripped ruff's I001
isort check with no auto-fixable diff. Verified: mine_glossary.EXTRA_DIRS ==
common.EXTRA_DIRS == generate.EXTRA_DIRS (module import deferred here since
generate.py needs faster_whisper, which the test suite stubs); 156 tests still pass.
Both scripts had their own copy of the inline title-cased alternation regex. Now both
source shell/lib.sh and call extras_grep_pattern() against data/extras.txt, falling
back to the literal old regex if sourcing fails or the data file is missing/unreadable
(extras_grep_pattern signals that case by returning 1 with no stdout, so the `||
echo fallback` idiom captures cleanly instead of concatenating garbage).

merge_pass.sh uses its existing $APP variable (APP_DIR, default /scripts, set to /app
by container_run.sh) rather than the tasks.md snippet's hardcoded /app, since the
script is also still reachable via the deprecated merge_watcher.sh's /scripts mount.
post_show.sh has no such variable and already hardcodes /scripts everywhere else, so
its new lines match that (its only caller, anime_library.sh, is deprecated as of B1
and always mounts to /scripts, never /app).

Verified equivalent to the old inline regex with a standalone repro (fake dir tree
with a Behind The Scenes + an Extras subfolder): old regex, new success path, and
new fallback path all produce the identical filtered file list.
Sidecars/stamps/logs the pipeline writes into the media tree (or, during local
testing, next to the repo) shouldn't ever land in git status as untracked.
Verified: touching one file per pattern produces no untracked entries.
Bare os.environ["PLEX_URL"]/["PLEX_TOKEN"] raised an opaque KeyError if unset.
Now os.environ.get(...) + sys.exit("PLEX_URL not set") / "PLEX_TOKEN not set".
Also splits the combined `import os, sys, urllib.parse, urllib.request` line into
one-per-line (already alphabetical, so no reordering needed) -- these were the
last 2 ruff errors in the tree; `ruff check .` is now 0 errors whole-tree.
common.py was never added to Dockerfile.builder's COPY list when V1 introduced it --
every module doing `from common import ...` (generate.py, mine_glossary.py, mux.py,
repair.py, dub_signs_merge.py) would ImportError at container start. Found while
verifying the B7/B9 EXTRA_DIRS consolidation needs data/extras.txt and shell/lib.sh
in the image too (referenced via $APP/data/extras.txt and $APP/shell/lib.sh, where
$APP defaults to /app under container_run.sh). Added all three; used separate COPY
<dir>/ /app/<dir>/ instructions for data/ and shell/ so their contents land in a
subdirectory rather than being flattened into /app/ (Docker's COPY semantics for a
directory source copy contents, not the directory itself).
Missed staging this in the B9 commit. Without this, a missing/unreadable
data/extras.txt would make extras_grep_pattern print the literal string "()" and
still exit 0 -- callers' `PATTERN=$(extras_grep_pattern ... || echo fallback)`
would then use "()" instead of falling back, and grep -ivE "()" matches (and thus
excludes) every line, since an empty alternation matches the empty string
anywhere. Verified: missing-file case now returns 1 with no stdout, so the `||`
fallback triggers cleanly.
Read HARDLINK_ROOTS once into a local instead of calling
os.environ.get() twice for the same var.
Matching (st_ino, st_dev) already establishes hardlink identity;
the extra samefile() stat-and-compare was dead weight.
A full HL_ROOTS tree walk per file is expensive; module-level
_partners_cache memoizes the result for the process lifetime
(one mux sweep), since hardlink partners don't change mid-sweep.
The size gate false-positived on compact muxes where mkvmerge
shrinks the CUES or drops a large embedded .ass. Verified first
that verify() has exactly one success path and the duration-tolerance
check already runs unconditionally on it -- that's the real
truncation canary, so no extra guard is needed once the size
proxy is gone. Adds a regression test asserting a truncated
(short-duration) output still fails verify().
The reference text comes from an untrusted third-party fansub file.
Wrapping it in <official_subtitle_reference> tags makes the model
read it as quoted data rather than instructions -- a prompt-injection
guard. Adds test_build_prompt_wraps_reference_in_xml_tag.
A raw [:1000] slice on the joined string could cut a name in half
mid-word, feeding the model a garbled "canonical spelling" fragment.
Now accumulates whole terms until the next would exceed 1000 chars.
After all episodes in a --root/explicit-file run finish, persist elapsed_s,
episode/card/drop/collapse/flag totals plus model + glossary revision info
to GLOSS_DIR/<show>.lastrun.json for ops visibility across runs.
verify() previously ran one blocking Ollama HTTP call per pending term,
serially. Add VERIFY_WORKERS (env, default 4) and run adjudicate() calls
concurrently via ThreadPoolExecutor.map, which preserves per-term result
pairing/ordering regardless of completion order.
Default path is None; resolves only from SEASON_PRIORITY_FILE (drops the
hardcoded /config/season_priority.txt silent fallback). No path configured
now logs "watch-order disabled" instead of silently probing a default file.
A malformed "Show:NN" value logs a warning instead of silently returning 0.
SEASON_START env fallback and file-beats-env precedence are unchanged.
Walks the show list and classifies each video by sidecar/stamp presence
(no .done/.ass/.srt/.fail -> generate; .srt no .ass -> repair; .ass no
.done -> mux), prints "would generate N, repair M, mux K", exits 0. No
containers launched. Verified via bash -n/sh -n plus a manual fixture
walk (mixed episode states, empty show dir, missing show, default LIST
resolution) -- real container sweep not runnable in this dev environment.
mine_glossary.COMMON -> data/common_proper_noun_deny.txt (one word per line);
hallucination.BLOCKLIST -> data/hallucination_blocklist.txt (one regex
alternative per line, joined with "|" and compiled with re.I). Both load at
module import and fall back to the exact original inline list/pattern if the
data file is missing/unreadable, verified byte-identical in tests.
generate.py::process(), repair.py::process(), and dub_signs_merge.py::
process_one() each had a bare "except OSError: pass" after os.chown --
replaced with log(f"chown failed for {p}: {e}") in all three so a
permission/ownership problem shows up in the run log instead of
disappearing. mux.py's os.chown is already covered by process()'s outer
except-and-log handler -- no bare-pass instance to fix there.
Adds "Who can execute" / "Behavior without permission" to a1-reflow-timing,
b1-hallucination-gate, c1-glossary-precision, d1-mux-fonts, and
glossary-wiki-verify. Notes one pre-existing gap in glossary_verify.py
(fetch_titles()'s os.makedirs() isn't exception-guarded) as out of scope
for this docs-only task.
Replace the (max_len, text) tuple packed into `fallback` with two named
variables, best_max_len and fallback_text -- readability only, behavior
identical (test_reflow.py unchanged and green).
The exception handler around transcription used to match "cuda"/"out of
memory"/etc. as a case-insensitive substring of str(e) -- a plain
ValueError or ZeroDivisionError that happens to mention "cuda" in its
message/stacktrace would falsely poison the episode and exit(3), even
though it has nothing to do with the GPU context.

Replace with isinstance(e, RuntimeError) -- what faster-whisper/
ctranslate2 actually raise for real GPU errors (OOM, device ordinal,
cuBLAS). RuntimeError still poisons (.fail kept, exit 3, fresh-context
restart). Any other exception type now clears the .fail marker so the
episode retries next sweep, and persists a small JSON retry log
(<stem>.dubtitles.crash.json: path/exc_type/msg/time) for later triage.

Two new tests prove both branches: a RuntimeError keeps .fail and
exits 3; a ValueError mentioning "cuda" in its message (the exact case
the old substring gate would have mishandled) clears .fail and lets the
show continue.
read_start() honors SEASON_START as a fallback when no SEASON_PRIORITY_FILE is
configured (unchanged, correct), but logged "watch-order disabled" even when a
non-zero SEASON_START meant reordering would actually happen. Now the disabled
log only fires when the resolved start is genuinely 0; a non-zero SEASON_START
gets an accurate "using SEASON_START=<n> (no priority file)" log instead. Also
corrects the docstrings to describe SEASON_START as a legitimate global
fallback rather than a hidden deviation.
build()'s style-merge loop already kept the first track's definition
when two tracks reuse a style name; it now also logs when the two
definitions disagree on fontname/fontsize, so a silent mismatch is
visible in run logs instead of just quietly winning.
Each source .ass track's WrapStyle is now compared against the base
track's value; a mismatch logs a warning naming both values. The base
track's WrapStyle still wins — this is diagnostic only, no output
events or styles change.
Sets base.info["ScaledBorderAndShadow"] = "yes" once the base canvas
is established, overriding whatever the source track had, so
outlines/shadows render consistently across Plex, mpv and VLC.
Collects (PlayResX, PlayResY) from every source track and logs a loud
warning if they differ, since positioned signs are only correct
relative to the resolution they were authored at. WARN ONLY -- no
coordinate transform and no track is skipped or dropped; the actual
resolution normalization is deferred to V3 (spec "Decisions taken").
mkvmerge -J reports embedded fonts as a top-level "attachments" array
(sibling of "tracks"), not as track entries -- confirmed against
mkvmerge's own JSON identification schema and mux.py's existing
tracks-by-type usage, which never sees an "attachments" type. The
task's originally-specified tracks-filter approach would have been a
permanent no-op against real mkvmerge output.

verify() now compares identify(orig)["attachments"] vs the muxed
output's, after the existing av/dubtitles-track/duration checks, and
returns "font-count-mismatch" on a count mismatch -- strictly adding a
new non-ok path, never changing an existing "ok" result. .get(...,
[]) treats a fontless file (key absent) as 0 so equal-zero still
verifies "ok". Bonus: logs (doesn't fail) when an attachment's
content_type is the generic application/octet-stream.
spec.md's D5 acceptance said 'skips the mismatched track', contradicting
tasks.md/the Decisions table (warn-only, transform deferred to V3). Skipping a
track would drop its signs entirely — worse than a warning. Aligned to warn-only.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
fix(shell): POSIX '.' instead of bash 'source' for shell/lib.sh
Some checks failed
CI / test (push) Failing after 1m23s
Details
CI / test (pull_request) Failing after 1m20s
Details
tests / test (push) Failing after 1m30s
Details
tests / test (pull_request) Failing after 1m26s
Details
eb76b64e11 
merge_pass.sh/post_show.sh are #!/bin/sh (dash in the subgen container), where
'source' is undefined — the B7/B9 EXTRA_DIRS consolidation silently fell through
to the inline fallback in-container. '.' is POSIX and works in dash and bash, so
the shared data/extras.txt is actually used.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Sign in to join this conversation.
Reviewers
No reviewers
Labels
Clear labels
No items
No labels
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
xenarathon/DubTitlerr!2
No description provided.