GLOSSARY() recomputes the slot collapse once per witness

[phas02]

What it buys: one slot collapse per read.

What stands: no store relation is re-scanned per call —
Store::read_context is cached by version
(crates/glossary/src/store.rs:1570-1592) and Store::history per
relation snapshot (store.rs:1402-1434). The cost is in-memory
recompute: Store::slots (store.rs:898) runs once per witness with
a detector inside raw_read (crates/session/src/reads.rs:239) plus
once in collapsed_read (store.rs:1057), each pass re-filtering
the whole history; verdicts re-reads witnesses_all()
(reads.rs:229) though ctx.witnesses holds the rows;
read_context computes the pin and the version before its cache
check (store.rs:1571-1572); each detector run builds a fresh
SessionContext (reads.rs:299-300).

Done when: the slots are computed once per read and shared across
witnesses and the collapse; the witnesses come from the context; the
cache check precedes the pin.

[phas02]

First half landed (13da7ef): one raw-read pass per read serves every witness — each takes its aspect's rows from it — and the witnesses come from the ReadContext, so verdicts derives from the context alone (crates/session/src/reads.rs); the pass-through parameters fell off verdicts, settings, current_fact_values and current_query_slots.

What still stands: the collapse runs its own slots pass (collapsed_read, crates/glossary/src/store.rs) beside the verdicts pass — sharing one pass needs a read currency carrying rank and snapshot_id, which RawRow does not; and each detector run still plans over a fresh SessionContext. The pin/version-before-cache-check item is absorbed by the head cache (store_snapshots): both derive from the cached enumeration, so the pre-check cost is in-memory formatting only.

Done when (remaining): the slots are computed once per read and shared with the collapse.