fix(matching): keep abbreviation titles out of bare-year keys (#1426)

A leading 2-5 character uppercase token before a dash, pipe or colon was
always read as a provider tag, so a film whose NAME is such a token lost it and
normalized down to its release year alone. "AKA - 2023" became the key "2023",
where it collided with BDE, BRO, OUT, WIL and IF — and they were offered to
each other as alternative sources in VOD multi-source.

"IT - 65 (2023)" (the Italian copy of the film "65") is structurally identical,
so only the token's meaning can separate them. The leading token is now tested
against a vocabulary — TRAILING_TAG_VOCABULARY plus a prefix-only list derived
from the real catalog — but only when the strip would leave no real word
behind. A compound is read by its head, so the open-ended "4K-<lang>" family
keeps working while "INU-OH" and "PC-4L" are recognized as film names. An
unknown token keeps its title: a refused strip costs one unmatched copy, a
wrong one corrupts that film's identity in VOD multi-source, the TMDB Similar
rail, DB_MATCH_TITLES and pin keys.

"No real word" is decided by running the rest of the pipeline on the stripped
form and looking at what comes out, never by re-implementing what later stages
remove. Quality tags, trailing and underscore tags, double-dash suffixes and
season markers each otherwise smuggle the strip through, and a stage added
later is covered for free.

Validated over the live catalog, movies and series: 83 keys fixed, 0 corrupted
across 1,616,111 titles. Deriving the vocabulary from movies alone missed AMZ,
D+ and P+ and broke the Paramount+/Disney+ copies of the numeric series 1923,
1883, 24 and 9-1-1.

docs/architecture/vod-multi-source.md

@@ -216,6 +216,50 @@ reader is free to be permissive because the gate in front of its riskier
 forms — and the fact that a row only appears once it HAS matched — keeps a
 bad guess cosmetic.
 
+The one case where the shape genuinely cannot decide is a strip that would
+leave **no real word behind**. "IT - 65 (2023)" is the Italian copy of the film
+"65"; "AKA - 2023" is the film "AKA" and its year. Same shape, opposite
+readings, so the leading token is tested against a vocabulary instead —
+`TRAILING_TAG_VOCABULARY` plus a prefix-only list (`NF`, `EX`, `NRC`, `AMZ`,
+`D+`, `P+`, `OSN`, `VO`, …), with a compound read by its HEAD so the
+open-ended `4K-<lang>` family works and "INU-OH"/"PC-4L" — real film names —
+do not. Unknown token, wordless remainder: the title is kept whole. That
+direction is chosen deliberately, because a refused strip costs one unmatched
+copy while a wrong one produced a bare-year key; on the live catalog
+AKA/BDE/BRO/OUT/WIL/IF all collapsed onto `"2023"` and were offered to each
+other as alternative sources. Gating the case where a word DOES survive was
+rejected for the mirror reason: it would strand every genuine tag the
+vocabulary has not heard of.
+
+"No real word" is decided by running the REST OF THE PIPELINE on the stripped
+form and looking at what comes out — never by re-implementing what the later
+stages remove. That is the load-bearing part of the design: every stage drops
+something, so a guard that predicts them is a list to keep in sync, and each
+omission is a silent collapse:
+
+| title               | dropped by         | would key as |
+| ------------------- | ------------------ | ------------ |
+| `\|TA\| RRR - HEVC` | quality tag        | `""`         |
+| `CAT - Multi ENG`   | trailing tag       | `""`         |
+| `IF - 2024_sub`     | underscore tag     | `2024`       |
+| `AKA --xyz`         | double-dash suffix | `""`         |
+| `CAT - 2022 S01`    | season marker      | `2022`       |
+
+All five fall out of one question asked of the real output, and a stage added
+later is covered for free. The season check deliberately uses
+`SEASON_SUFFIX_PATTERN` directly rather than `stripSeason`, whose
+"never return empty" fallback would report a lone season marker as a surviving
+word. A tag word sitting next to a real one is still part of the title
+("EN - Sub Zero" → `sub zero`).
+
+The vocabulary is evidence, not intuition. Each entry prefixes hundreds to
+thousands of ordinary lettered titles in the real catalog; nothing is added
+because it "looks like a streaming service" (MAX and HULU would qualify, and
+"MAX - 2015" is a film). Deriving it from movies alone missed `AMZ`, `D+` and
+`P+`, which broke Paramount+/Disney+ copies of the numeric series 1923, 1883,
+24 and 9-1-1 — so validate any change over movies AND series: 83 keys fixed,
+0 corrupted across 1,616,111 titles.
+
 The category path exists because many panels tag the CATEGORY ("EN | Netflix",
 "DE | Apple TV") and leave stream titles bare. Discovery returns every visible
 category name a stream sits in, and the two query tiers get there