P106: the dictionary-group index + the 84000 harvest

[arvicco]

The dictionary-group index + the 84000 harvest

Dictionary shelves now group by lect, every dictionary-shaped surface answers from the group, and the 84000 shelf's two long-named follow-ups — the back-matter glossary term bank and the Derge title crosswalk — are built. Two hardening riders close real incident classes in the lect facet's lifecycle and the batch-link replay contract.

The dictionary-group census (dictionary_stats, migration 036)

One precompiled row per dictionary: shipped language code, the lect node it resolves to through the nabu-lects ladder (per-source overrides included; NULL = the registry defines no such node — censused as unresolved, never silently invented), live entry count. Derived wholesale in both rebuild flavors and rake builders:refresh, per source after any dictionary-bearing sync.

Surfaces re-read the group:

• nabu language <code> gains its dictionary-group block — every shelf at or under the code's resolved node (prefix semantics: lat collects the la-med shelf's lat:med), entries descending, render-capped with an honest truncation line;
• define's header and --lang validation (and the MCP define lang enum) read the census instead of live-scanning the shelf table, with the byte-identical fallback for underived catalogs;
• the site census gains dictionary_lects / dictionary_unresolved_codes (54 / 61 live), cited from the languages page through Liquid only, guard-tested.

Recorded v1 limit: a dictionary carries ONE language code; mixed dictionaries group under their primary. A companion registry PR (arvicco/nabu-lects#17) covers the ancient-language tail of the unresolved census — after it merges and reaches canonical, resolution rises 175 → 326 of 464 on the same census.

The 84000 glossary shelf (second dictionary lane)

Every publication's back-matter <div type="glossary"> entity records — Tibetan script, Wylie, Sanskrit, English rendering, definition — load as ONE shared xct-headed shelf (e84000-glossary) through the secondary dictionary lane. Discovery delegates to the primary adapter's cone and duplicate-pair winner rule, filtered to gloss-bearing files, so the lane can never disagree with the passage shelf about editions. Tibetan headwords strip the trailing shad before folding so EWTS-typed queries hit (define gser 'dab → garuḍa), while key_raw keeps the verbatim upstream term; both real upstream gloss shapes parse (bare <term> + inline definition, and translationMain + <note type="definition">). License: the entries inherit 84000's stated CC BY-NC-ND 3.0 grant (class nc), credited as the source row states.

The 84000 → Derge title crosswalk

Every 84000 header — the 3,926 placeholder stubs (which carry the titles of every untranslated text) and the published translations — names its Toh text in English and Wylie; the Derge shelves mint those documents title-less. The new pass fills documents.title by EXACT Toh key (a part key titles the part document; no base-container inference), published headers winning over placeholders. Runs in both rebuild flavors, builders:refresh, and after e84000/derge syncs.

Riders

• Lect-facet lifecycle hardening: LectFacets.rebuild!'s delete + rewrite + census now run in one transaction (a killed rebuild can no longer strand a zero-row axis); builders:refresh covers the lect facet, the dictionary census and the title crosswalk; nabu health gains three loud invariants — stats claiming staged documents beside an empty facet, per-key census drift, and override-ruled sources holding live documents with zero staged rows.
• Place scopes in the replay contract: place mine / place link record producer + scope + gazetteer in config/link_scopes.yml, and the links-stage replay (extracted to LinkScopeReplay, dispatch and ordering test-pinned) re-runs both — place-link after place-mine, with a loud failure when the places registry is absent.
• Registrations: the new table classified in the derivability manifest; the two new corpus stages labeled in the rebuild profile (their time reaches the report and the durable ETA record).

The live round (run, verified)

bin/nabu sync e84000 --parse-only
# 388 documents =skipped (content unchanged) · dict +56,794 entries into e84000-glossary
# 124,223 passages reindexed · crosswalk refs re-derived (488 edges)
# derge titles: 4,451 of 4,560 documents titled (the 109 remainder have no
# exact-key 84000 header — the honest gap until a bare-base header exists)

dictionary_stats derived live: 464 rows, 54 lect nodes, 61 unresolved codes. Suite 9,971 runs / 207,527 assertions / 0 failures; rubocop 1,265 files clean; site census restamped from the live catalog at this gate.