Tsumugu: the word campaign, executed (2026-06-13, night)

Continues the campaign setup. The seed landed and four agents ran it down in a single wave.

What ran

• Seed: 3,529 normalized TOCFL word gaps prepended band-ordered into backbone-queue (562 of the 4,091 were already queued/authored from the frequency corpus — deduped, no collisions). TOCFL-first ordering, frequency tail intact behind it.
• Four author-only agents, one per disjoint slice (~1,023 words each), via the parallel-wave model: write entries only, no git / queue / LOG / cross-slice touches. Central bookkeeping (validate, commit) deferred to one pass.

Result — TOCFL words effectively complete

5,960 / 6,006 distinct multi-char TOCFL words authored = 99.2%. A1, A2, B1, B2 are 100%; 46 words remain, all C1. Validation green across 9,481 entries (2,526 chars + 6,955 words), ids unique, filenames canonical — four agents wrote ~4,047 entries in parallel with zero id or filename collisions. The author-only isolation model held at scale; per the wave-1 cost analysis it’s also the cheaper path (no per-slice QA tax).

Variant features shipped and verified

• Cross-links are bidirectional (台灣↔臺灣, 水分↔水份) via related[].
• 140 scholar-notes proposed on orthographic twins (注/註 water-vs-speech radical; 表/錶 clockwork specialization; 畫/劃 added knife; 僱/雇 the seasonal-hire loan). Quality high and accurate — Wedge approved all 140. The mechanism rode an existing scholarNote field the schema and style card already defined (“a variant gloss… author proposes, Wedge approves”) — no new machinery.

Edge cases / what was checked

• The 46 skips are one dependency. Every skipped C1 word is blocked on one of the 55 C1 characters — cribbed and (now) cross-checked, but not yet authored. So the whole remaining gap, 46 words + 55 chars, collapses to a single action: author the 55, then sweep the 46.
• Scholar-note side-placement (undecided): 68 twin pairs carry the note on only one side (甚麼 has it, 什麼 doesn’t). Cosmetic — cross-link still connects them. Open: mirror to both, standardize onto the orthodox form, or leave.
• QA false positives cleared: the register scan’s hits were hedge-meaning headwords (也許, 說不定, 定義); the “5,327 thin” were legitimate story:null transparent compounds. Zero real thin entries, zero real register violations.

Operational notes (recurring friction)

• The researched→queued gate bit a third time. The 55 C1 chars landed cross-checked but stuck at researched; nothing auto-promotes a merged-crib char to queued, so the author loop ignores them until a manual flip. Worth fixing at the crib-merge step (crib_diff could promote to queued once both lanes agree) so the gate stops recurring.
• .git is mount-blocked in the Cowork sandbox — commits/pushes for this repo (and the KB) run host-side only. The 4,047-entry milestone could not be committed from here; the char loop clears the stale index.lock host-side at startup and commits via git add -A. (Generalizes the earlier KB push-constraint note.)

Status / next

• Char loop running in another Claude Code window (KIND=chars) — authors the 55 C1 chars and commits the word-campaign milestone in its first batch.
• Then: sweep the 46 C1 words (re-run the word agents on their SKIPS.txt, or a one-shot sweep) → TOCFL A–C complete, characters and words.
• Then: reconcile_queues.py --apply to sync the drifted queue (agents authored without status updates — the deferred reconciler).
• Still parked: the 289-entry single-source audit (Composer cross-check); the scholar-note side-placement decision.