Record vocabulary phase 2's agreed design in the RFC and decisions (#752)

AGENTS.md

@@ -306,7 +306,7 @@ Add a dedicated `copy.deepcopy()` round-trip test for it too (see `test_regexes_
 
 **A group-side reading of a family comma's given part that asks `tail_reading` disagrees with the given part's own walk, and the disagreement is accepted.** `tail_reading` reads the part as the comma-less name reads it, plus the bare salutations rules/H8 adds after a family comma, which a predictor of this part asks for with `salutations=True` (#835); assign's walk over the given part after a family comma reads by its own predicates -- a shape-only numeral VI to X stays a name word there (`Smith, John van der Berg VI` reads middle 'van der Berg VI'), V before a third comma part is read strictly (#144), and P6 attaches a period-marked title-particle (`St.`) the comma-less reading takes as a title. So a bound or a decline group computes with `tail_reading` acts where the walk then reads no title: #623's chain bound (`Smith, John Mc Prof. VI`) and #637's join decline (`Smith, abdul Prof. VI` → middle 'Prof. VI', `Smith, abdul St.` → family 'St. Smith'). Both were recorded rather than narrowed (Derek: garbage input; decisions/H5 and #P5, rules/P5's Accepted clause). Before writing another group-side predictor of the given part, sweep those three shapes and decide, rather than rediscovering the gap in review. And a change to the trailing peel ITSELF reaches the given part through those predictors without touching them: #718's words-to-spare count for a one-letter numeral, asked through `trailing_run_start`, made the chain take the V in `Smith, Freiherr von Berg V` (given 'von Berg V', unreported) while assign's walk, which never counts there (S4: "read without it"), read it as the suffix -- found by review on a grid of family-comma and empty-head shapes behind a title-particle; the predictor now passes `count_numeral=False`. When a change moves what the peel takes, put a family comma with a title-particle head in its grid.
 
-**A comparison that runs both sides in one tree reports 0 differences, which is what success looks like.** `tools/differential/compare.py` has been hardened against this — it generates its baseline worker into a temp dir, strips `PYTHONPATH` from the child, and aborts unless both the version AND the resolved path check out on each side; its README's three invocation traps are the analysis behind that design, worth reading before changing the harness. **The exposure is ad-hoc comparisons you write yourself**, where the same collapse has a cause the harness cannot disarm for you: **the shell's working directory persists between tool calls**, so a two-tree comparison written as two `cd`s silently runs both halves in whichever tree it landed in. Pin absolute paths, and assert `nameparser.__file__` on both sides the way `compare.py` does. The general rule outlives any particular trap: before believing a null result, prove the harness can report a difference — a clean run and a broken harness are the same output. The proof has to come from the input class the null result is about: #683 moved `tools/name_usage.py` to a POST and checked it with `shehu`, which is ASCII and so cannot show that the server decodes a Cyrillic body. A misdecoded query still runs and matches nothing, so the Cyrillic zeros it was meant to vouch for proved nothing until `Иван` (17,119 given) went through the same path. The converse trap is a mutation that moves nothing: it shows the code is unexercised by your grids, not that it is dead. #638 was filed calling group's chain re-ask dead on a grid with no title-particle head before a particle run and a dotted acronym; the next review found 400 moves (`Smith, Freiherr van Berg X.Y.Z.`) and the issue closed the same day. Before calling code dead, look for the input that reaches it. A CACHE of readings is the same trap one step removed: a wrong key copies one tree's readings in as another's, and the check then reports nothing. #859 first keyed `snapshot.py --local`'s kept baselines by `git rev-parse HEAD:nameparser`, which answers for the enclosing checkout, so a gitignored copy inside one (`build/lib`) took the checkout's key; two reviewers found it. Key a cached reading by what the import reads -- the files on disk, hashed again after the read -- never by what version control says should be there.
+**A comparison that runs both sides in one tree reports 0 differences, which is what success looks like.** `tools/differential/compare.py` has been hardened against this — it generates its baseline worker into a temp dir, strips `PYTHONPATH` from the child, and aborts unless both the version AND the resolved path check out on each side; its README's three invocation traps are the analysis behind that design, worth reading before changing the harness. **The exposure is ad-hoc comparisons you write yourself**, where the same collapse has a cause the harness cannot disarm for you: **the shell's working directory persists between tool calls**, so a two-tree comparison written as two `cd`s silently runs both halves in whichever tree it landed in. Pin absolute paths, and assert `nameparser.__file__` on both sides the way `compare.py` does. The general rule outlives any particular trap: before believing a null result, prove the harness can report a difference — a clean run and a broken harness are the same output. The proof has to come from the input class the null result is about: #683 moved `tools/name_usage.py` to a POST and checked it with `shehu`, which is ASCII and so cannot show that the server decodes a Cyrillic body. A misdecoded query still runs and matches nothing, so the Cyrillic zeros it was meant to vouch for proved nothing until `Иван` (17,119 given) went through the same path. The converse trap is a mutation that moves nothing: it shows the code is unexercised by your grids, not that it is dead. #638 was filed calling group's chain re-ask dead on a grid with no title-particle head before a particle run and a dotted acronym; the next review found 400 moves (`Smith, Freiherr van Berg X.Y.Z.`) and the issue closed the same day. Before calling code dead, look for the input that reaches it. A CACHE of readings is the same trap one step removed: a wrong key copies one tree's readings in as another's, and the check then reports nothing. #859 first keyed `snapshot.py --local`'s kept baselines by `git rev-parse HEAD:nameparser`, which answers for the enclosing checkout, so a gitignored copy inside one (`build/lib`) took the checkout's key; two reviewers found it. Key a cached reading by what the import reads -- the files on disk, hashed again after the read -- never by what version control says should be there. And a switch that changes no file defeats any file key: #752's spike set its options by environment variable, so after the first run every option copied that run's cached local readings until the variables joined the key (2026-10-24).
 
 **`SUFFIX_ACRONYMS` and `SUFFIX_WORDS` are DISJOINT by construction (since 2026-10-18; an import-time assert from 2026-09-08) — a word has one suffix kind, and its reading is an acronym or a word (`acronym`), so the vocabulary build refuses a word classified twice; the fix is to choose, not to relax the check** — the two branches normalize differently (the word test strips only edge periods, the acronym test strips all of them), so a word in both is matched by two rules and which one fired is unreadable from outside. `esq` was the single overlap and left `SUFFIX_ACRONYMS`, which is what made the assert possible; the whole algebra that had made the overlap safe, and the criterion that decided which set a word belongs to, are in the `suffix-acronym-collisions` entry of `docs/design/decisions/` if anyone wants them.