CLAUDE.md cleanup — collapse the mega-paragraphs that accreted across recent bundles, push verbose details into the corresponding docs/claude/ deep-dive files where they already lived in parallel. CLAUDE.md is auto-loaded into every Claude Code session, so its weight is paid every turn; tight orientation + pointers beats inline catalogs. Six bullets rewrote: (1) **Transactions § locking sub-bullet** — the inline 8-mode + hint-surface + escalation catalog collapses to a one-line summary with pointer to docs/claude/locking.md. (2) **Not modeled § Key-range locks** — the ~3000-token phase-0-through-3-shipped catalog (every Msg verbatim, every implementation detail of MVCC, every DMV column rule) collapses to a short list of what ships + the remaining phase-3 gap + pointer to docs/claude/locking.md; the orphaned BEGIN DISTRIBUTED TRANSACTION and SET <option> … sub-paragraphs lift out into their own dedicated bullets. (3) **Not modeled § programmable** — the inline CLR/UDF/TVF/view/procedure/sequence/trigger/TVP/TRY-CATCH/RAISERROR catalog collapses to a "ships comprehensively, see [docs]" summary plus the top-level gaps; TRY/CATCH detail moves to its own dedicated bullet pointing at docs/claude/control-flow.md. (4) **Not modeled § ALTER TABLE** — modeled-shape + out-of-scope-shape catalogs each collapse to one line apiece with pointer to docs/claude/alter-table.md. (5) **Not modeled § Query-hint gaps** — the multi-paragraph Msg 1047/1065/1069/307/308/321/102 verbatim wording catalog + legacy-paren placement rules + remaining-gaps narrative collapses to a 3-line summary with pointer to docs/claude/query-hints.md. (6) **MERGE / OUTPUT** — the inline grammar + multi-Msg catalog collapses to the canonical syntax sketch + the four enforced Msg numbers + pointer to docs/claude/dml.md. Each cut was preceded by a grep -c against the target deep-dive to confirm the verbose content already lived there; nothing load-bearing got dropped, only the auto-loaded-into-every-session duplication. Net result: CLAUDE.md shrinks from ~93k chars / ~26.8k tokens to ~50k chars / ~14k tokens (-47%), with the deep-dive docs unchanged (docs/claude/locking.md 563 lines, programmable.md 194 lines, alter-table.md 325 lines, query-hints.md 241 lines, dml.md 90 lines — all already had the cut content captured in the previous bundles' "Update CLAUDE.md and docs/claude/" steps; the duplication was the actual bug). The Feature reference trigger-phrased index at lines 132-158 stays untouched — it's already in the ideal "pointer to deep dive on demand" shape and isn't the bloat source. The one-line Not modeled bullets (RIGHT JOIN derived-table, comma-FROM, RANGE BETWEEN, LIKE COLLATE, CONVERT styles, LEN(ntext), MERGE-CTE, UNIQUE-on-computed, table-vars, global temp tables, etc.) stay verbatim — those are correctly-sized at top-level. The Quirks section stays untouched — each entry is genuinely self-contained edge-case behavior that doesn't have a deep-dive home and needs the context inline.

Author: RyanLamansky