Add subtitle dictionaries, Polish number normalizer, ComfyUI lifecycle control

Text processing before TTS:

- dictionary.txt rewrites words and phrases (km/h, np., units). Rules compile
  into one alternation applied in a single pass, so replacements never chain
  and longest-literal wins. Case-sensitive, whole-word.
- A per-queue second dictionary, written by the GUI into the first video's job
  folder and merged with the shared one (--dictionary-extra).
- ALL-CAPS blocks become Title Case rather than sentence case: lowercasing
  destroyed proper nouns inside captions ("NASA - MARGO MADISON") and invited
  the normalizer to "correct" the name. Paired clips in output/clips/all-caps
  showed the TTS reads both casings the same.

normalizer/ — digits to spoken Polish, in its own venv because it needs
transformers>=5, which must not touch ComfyUI's pinned versions:

- Two backends. Plain integers go to num2words for the value and a general
  Polish LLM for the case; times, dates, decimals and identifiers go to a 0.6B
  fine-tune. Measured on 40 real subtitle lines, each is clearly better than
  the other on its half. 74% of numbered blocks take the first leg.
- Guards on every answer: runaway length, repetition loops, invalid bytes,
  leftover digits, a lost minus sign, a changed number, a lost word. A failure
  falls back to num2words, which is nominative but never wrong about the value.
- ComfyUI is asked to unload before normalizing; a normalizer that does not fit
  in VRAM is spilled to RAM silently at 5x the cost.

Backend lifecycle:

- ComfyUI is found by port through /proc, so a backend started by hand is
  managed too. SIGTERM escalates to SIGKILL, and a suspended process is killed
  outright since it never runs its handlers.
- A clip that kills the backend is retried three times, each with a different
  seed (the aborts are seed-dependent, not text-dependent), then skipped and
  reported so an overnight queue keeps moving.
- start_comfyui.sh execs python, so the script's PID is ComfyUI's.
- GUI gains ComfyUI and Queue dictionary buttons; quitting stops the backend.

--resume-from now reuses the job folder's tts_input.srt instead of reparsing
the subtitles, so block N still pairs with clip_NNNN.flac.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

readme.txt

@@ -30,7 +30,7 @@ Settings
 --------
 Defaults live in settings.ini next to the script:
   voice, speed, temperature, max_tokens, strip_leading_ms, trailing_fade_ms,
-  orig_volume, tts_volume, output_dir, comfy_url
+  orig_volume, tts_volume, output_dir, dictionary, numbers, comfy_url
 
 Precedence: command-line flag > settings.ini > built-in default.
 Use --settings FILE to select a different profile.
@@ -59,10 +59,12 @@ Both external SRT files and embedded subtitle tracks go through the same cleanin
 
 - HTML tags stripped:          <i>, </i>, <b>, <font color="…">, etc.
 - ASS override codes stripped: {\an8}, {\pos(x,y)}, {\c&H…&}, etc.
-- ALL-CAPS blocks lowercased and converted to sentence case
+- ALL-CAPS blocks converted to Title Case (keeps proper nouns readable)
 - Sentence fragments starting with a lowercase letter are capitalised
 - Missing terminal punctuation added (… or . depending on context)
+- Dictionary replacements applied (see below)
 - Dialogue lines (prefixed with -) are split into separate clips with proportional timing
+- Numbers spelled out as Polish words, in the right case (see below)
 
 With --embedded-subs the extracted subtitles are saved to the job folder as embedded_sub.srt
 so you can inspect them before or after the run.
@@ -72,6 +74,63 @@ The final MKV preserves all subtitle tracks from the source video — text-based
 stream copy with no conversion.
 
 
+Replacement Dictionary
+----------------------
+dictionary.txt next to the script rewrites words and phrases before they reach
+the TTS engine — abbreviations, units, symbols, names the model mispronounces.
+One rule per line:
+
+    km/h = kilometrów na godzinę
+    np. = na przykład
+    /(\d+)\s*%/ = \1 procent
+
+Matching is case-sensitive and whole-word only, so "np." leaves "znp." alone and
+"Joshua" leaves "joshua" alone. The replacement is inserted exactly as written.
+The first letter of each block is capitalised before these rules run, so add a
+capitalised variant for words that can start a subtitle.
+Longer patterns win over shorter ones regardless of the order they are written
+in, and replacement text is never rescanned, so rules cannot feed each other.
+A pattern wrapped in slashes is a regular expression, with \1 \2 … referring to
+its groups. Lines starting with # or ; are comments.
+
+The result is visible in the job folder's tts_input.srt — check there if a rule
+is not doing what you expect. --regen and --rebuild work from that file, so they
+do not re-apply the dictionary; edit tts_input.srt directly, or rerun.
+
+Use another list with --dictionary FILE (a relative name is looked for in the
+current directory first, then next to the script) or turn it off with
+--dictionary "". A missing file is not an error.
+
+A second list can be merged in with --dictionary-extra FILE. The GUI writes one
+per queue — the "Queue dictionary" button — into the first queued video's job
+folder, for rules that belong to one batch rather than to Polish in general.
+
+
+Number Normalization
+--------------------
+The TTS misreads long digit strings and sometimes turns them into non-words, so
+after the dictionary every block containing a number is rewritten with the
+numbers spelled out in Polish — in the correct grammatical case, which is why
+this is a small neural model and not a lookup table:
+
+    Wygrał 1247 złotych.     ->  Wygrał tysiąc dwieście czterdzieści siedem złotych.
+    Mówimy o 25 psach.       ->  Mówimy o dwudziestu pięciu psach.
+    Mam 3 koty.              ->  Mam 3 koty.        (single digits are read fine)
+
+It runs in its own environment (transformers>=5, which must not be installed
+next to ComfyUI's pinned versions). One-time install:
+
+    normalizer/setup.sh                          # venv + dependencies
+    HF_TOKEN=hf_... normalizer/setup.sh --model  # gated weights, accept terms first
+
+Terms: https://huggingface.co/Folx/qwen3-0.6b-pl-text-normalization
+Everything after that runs offline. See normalizer/README.md.
+
+Turn it off with numbers = no under [normalize] in settings.ini, or with
+--no-normalize-numbers for one run. If the environment or the model is missing
+the run continues with the digits untouched — it is never fatal.
+
+
 Output Files
 ------------
 Everything belonging to one video is kept together in a job folder named after
@@ -130,6 +189,13 @@ All Options
   --max-tokens N        Max audio tokens the model may generate per clip (default: 4096)
                         Raise this if long clips are cut off mid-word.
 
+  --dictionary FILE     Word/phrase replacements applied to the subtitles before
+                        TTS (default: dictionary.txt). Pass "" to disable.
+
+  --normalize-numbers   Spell digits out as Polish words before TTS (default: on)
+  --no-normalize-numbers
+                        Leave digits as digits. The TTS often misreads them.
+
   --orig-volume 0-1     Volume of the original audio in the Polish lektor mix (default: 0.60)
   --tts-volume 0-2      Volume of the TTS voiceover in the Polish lektor mix (default: 1.8)
                         Note: these affect Track 1 only. Track 0 is always the raw original.
@@ -151,7 +217,8 @@ All Options
 Recovery and Fixing Bad Clips
 ------------------------------
 
-Resume a crashed run from block N (reuses existing clips 1..N-1):
+Resume a crashed run from block N (reuses existing clips 1..N-1 and the job
+folder's tts_input.srt — subtitles are not reprocessed):
 
     python3 lektor_ai_comfy.py input/film.srt input/film.mkv --resume-from 47