Stop a possessive from spending a head token-cap slot

[brad-richardson]

Implements the Worker half of docs/plans/2026-08-07-apostrophe-folding-locus.md. No rebuild required.

The defect

The apostrophe is a plain separator on both sides of the pipeline, so Len's Mill Store indexes as [len, s, mill, store] — four clauses against a HEAD_QUERY_TOKEN_CAP of three. The head refuses the query before any read.

Measured on the corpus production serves: 733,701 records are over the cap only because of that split. Live today, Queen's Medical Center returns empty while Queens Medical Center answers — from different records that happen to be spelled without one.

The fix

The AND lanes drop apostrophe-born one-character fragments. A fragment has no retrieval value to lose: its posting is saturated by construction, and merge_bounded_candidates already admits a record absent from a saturated posting when its own display tokens carry the token. Dropping it removes a cap cost, not evidence.

Three things it deliberately does not do

It does not touch the phrase lane. eN: keys are built from the record's full name, fragment included, so a filtered key names something the head does not contain — e2:domino pizza where the producer wrote e3:domino s pizza. places_construction_head_records now takes the unfiltered clauses and the filtered clauses separately, and a test pins exactly that reason.

It does not drop short tokens generally. 5,071,193 records carry a one-character token with no apostrophe anywhere — H&M, A&W, initials, single CJK characters — and are retrievable today only because those postings survive. A blanket rule would break a working class to fix a broken one.

It does not treat U+02BC as an apostrophe. I assumed it was one and the test caught me: Unicode classifies the modifier letter as a letter, so the index tokenizer never splits on it and there is no fragment to drop. The test now asserts that, so a future "add the other apostrophes" change has to face it first.

A query that is nothing but fragments (L'A) keeps its clauses — an empty AND retrieves nothing at all, which is strictly worse than what this replaces.

Verification

• Worker tests 249 passed (5 new), workspace green, fmt/clippy -D warnings clean, wasm32 checks clean.
• Python pipeline suite 1,810 passed, 2 skipped.
• The head cap contract test (test_worker_global_head_uses_one_three_token_cap) fired twice during development — once on a test local and once on a comment quoting the guard expression. The five production cap-guard sites are unchanged; the count is still 5.
• A queries_without_an_apostrophe_are_bit_identical_to_query_terms test pins that every case in both frozen sets tokenizes exactly as before.

Before merge

This is a retrieval-path change, so the standing rule applies: it ships behind a paired measurement on the gold and everyday sets — and now the proximity stratum too, via #261. It needs a production deploy with an immediate paired compare and rollback readiness, since there is no Worker preview environment.

🤖 Generated with Claude Code