Release 0.133

v0.133.2 — 2026-07-15

Hotfix 2: migration 0119’s journal when predated 0118’s, and the migrator gates on when > max recorded created_at — so boxes that already ran 0118 skipped 0119 even with the journal entry present. Restamped to the +1-day ledger convention; the journal guard test now also enforces strictly increasing when values.

v0.133.1 — 2026-07-15

Hotfix: migration 0119 was missing its journal entry, so the migrate gate skipped it (“Already up to date”) while v0.133.0’s code queried the new content_chunks.search_tsv column. Journal entry added; a new guard test fails the suite whenever a migration .sql lacks a journal entry (or vice versa). Boxes that rolled v0.133.0 self-heal on this release — the migration SQL is idempotent.

v0.133.0 — 2026-07-15

Retrieval: hybrid passage search, spreadsheet profiles, corpus map. Born from a production recall audit. (1) search_chunks gains a keyword arm — weighted RRF over the new content_chunks tsvector (migration 0119) with a rescue floor, so exact rare tokens (error codes, field names, coined terms) are findable even when they embed poorly; the responder’s auto-context uses it too. (2) Spreadsheets index as one profile chunk per sheet (headers + sampled rows + honest coverage note) instead of thousands of embedded grid rows — they were 74% of one brain’s chunk table; full text still persists for file_read. Versioned exports (date/_version_NN families) get their older copies salience-down-ranked, newest self-heals. (3) Every responder turn now carries a cached corpus map — branch-grouped titles (+ page/table one-liners) on its own prompt-cache breakpoint, memory_config.corpus_map_limit to tune.