Add docs/design/vocabulary-model-rfc.md, where the vocabulary model is going

docs/design/decisions/vocabulary-model.md

@@ -13,7 +13,7 @@ The vocabulary was a set of base sets plus marker subsets (`particles` and `part
   8. A patronymic field belongs in 3.0.
   9. Gender is recorded only where a word gives a definitive clue.
   10. Title-phrase modifiers count as titles only next to a head title (parked, not filed).
-  11. Name readings are hand-declared, and only for words some language file claims as vocabulary, so nobody maintains surname lists.
+  11. Name readings are hand-declared, and only for words some language file claims as vocabulary, so nobody maintains surname lists. (Widened 2026-10-19, below: or refuses.)
   12. Bearer counts are structured provenance (source + count), never used to compute a reading: two data sets deciding one reading would make the decision unclear. A new census flags changed names for review.
   13. `Dean Smith` reads given `Dean`, as today.
   14. Order and shape are separate: a language supplies only a default order; the shapes (where titles, suffixes, particles and commas sit) belong to each order and live in the core.
@@ -37,5 +37,7 @@ The vocabulary was a set of base sets plus marker subsets (`particles` and `part
   - **The comments moved with their words.** The suffix words' per-word facts became fields (`note`, `issue`, `same_as`, `expands`, `excluded`); the class-level parts of the vetting notes were checked against the decisions entries that already hold them, and what had no home went to decisions/W2's entry of this date. In the mechanically converted categories the block comments stand as TOML comments above the entry they stood above, in source order, and inline comments became reading `note`s; classifying those categories is the later work that turns them into fields.
   - **Measurement (2026-10-18, Python 3.11.17).** `Lexicon.default()` and all thirteen raw constants equal the pre-change tree's, compared against values pickled from master at 6803a261; `tools/differential/snapshot.py` reports "matches the snapshot". Negative control: dropping `may_be_name` from `ma` moves 120 snapshot entries and makes `Lexicon.default()` unequal, and editing the TOML without rebuilding fails the staleness test with the refresh command. `import nameparser` measured 48.5–48.8 ms median against master's 49.4 ms (25 fresh processes each, two runs), so the generated module costs nothing visible; the spike measured parsing the TOML at import at about +18 ms on a 48 ms import, and the TOML grows with descriptive data the compiled module never carries.
   - **Declined, with evidence.** PARSING THE TOML AT IMPORT: +18 ms net on every import (spike, 2026-10-18), and it could not be made lazy on 2.x without making the config modules and the shim's `CONSTANTS` lazy too. DERIVING THE MARKERS NOW: see the flags bullet. A `peelable` BOOLEAN: `attached` is the general property from the start, a table so its other side and separators can be added. A `language` KEY in the file: a second source of truth for the file name.
-  - **Phases 2–4 are rejected by the compiler, not accepted as inert data**: `addresses = "full"`/`"any"`, `attached` on a side other than after, and the names later phases will use (`default`, `bare_only`, `leading_only`, `trailing_only`, `attaches`, `separators`) are build errors naming their phase, so nothing reads as working that does not. Locale-pack vocabulary and the Korean surname list stay in Python: surnames are a segmentation list, outside the per-word model, and packs move in phase 3.
+  - **Phases 2–4 are rejected by the compiler, not accepted as inert data**: `addresses = "full"`/`"any"`, `attached` on a side other than after, and the names later phases will use (`default`, `prefer`, `strict`, `bare_only`, `leading_only`, `trailing_only`, `attaches`, `separators`) are build errors naming their phase, so nothing reads as working that does not. Locale-pack vocabulary and the Korean surname list stay in Python: surnames are a segmentation list, outside the per-word model, and packs move in phase 3.
   - **Recompute the composition** rather than quoting it: `uv run python tools/vocab/query.py --kind generation --count` (readings, one per file a word is in), `--kind suffix`, `--excluded generation`, `--issue 661`.
+
+- **2026-10-19 — decision 11 widened to words a file refuses (Derek).** A name reading may also sit on a word whose vocabulary reading some file refuses: the reading is the evidence the refusal rests on, as `fils`'s `family` reading in fr.toml backs its excluded `generation`. The build already allowed this (its soft warning fires on a name reading only where the word has neither a vocabulary reading nor an exclusion), so the decision now matches the shipped data; #758's review found the mismatch.