Tsumugu: characters done, words next (2026-06-13, evening)

The seed-gate unblock (see earlier today) let authoring rip. By this snapshot the corpus stood at ~2,526 characters + ~2,910 words and the character campaign had effectively closed.

Milestone: the TOCFL character set is essentially complete

A1, A2, B1, B2 are at 100% character coverage; C1 is down to the 55 Grok just cribbed (the resume worklist worked — clean restart, all 55 landed, 2 phonetics flagged unresolved). Author those and the whole A1–C2 character backbone is done.

C2 finding: C2 contributes no new characters — and in fact the 華語八千詞 / 8000-word list this corpus runs on has no C2 tier at all, at either the character or word level. “What about C2” has a clean answer: it isn’t in the source data. True C2 (精通級) coverage would require an external word list and a separate campaign — explicitly out of scope for now.

The bottleneck flipped to words — and the word lane was pointed at the wrong source

While characters were chased to completion, TOCFL word coverage quietly lagged to ~32%. The reason is structural: the word lane drains backbone-queue, a frequency-ranked corpus, not the TOCFL list. So it authors frequent words, which only partially overlap TOCFL — completion by accident, not design. The raw gap looked like ~4,154, but the list encodes variants and options inline (你/妳, 台灣/臺灣, 小孩(子), 名字(˙ㄗ)); normalizing those (slash→primary, strip parentheticals) cleared ~206 false misses. True gap: 3,948 — A1 170, A2 191, B1 425, B2 1,121, C1 2,041.

Decision: open a TOCFL word completion campaign

Mirror the character campaign on the word axis. Confirmed with Wedge:

• Seed all 3,948 normalized gaps, band-ordered A1→C1.
• 3–4 parallel agents, staged into disjoint slices via the parallel-wave driver (the simple loop races git/queue on concurrent commits).
• TOCFL-first drain priority — close the TOCFL gaps before the frequency-corpus tail.

Key enablers and guards:

• Words need no crib lane — they decompose into the now-complete characters, so seeded words author immediately; Grok/Cursor sit idle for this campaign.
• Reconciler deferred to post-campaign (Wedge’s call) — so every agent must run KIND=words to stop the stale char-queue count from pulling them into char runs.
• PRD written: batches/parallel-wave/WORD-CAMPAIGN-PRD.md (problem, success criteria, scope, constraints, plan, open questions).

Variant policy (resolved) + staging done

The slash-variant question resolved toward maximum coverage: every written form is its own entry — distinct words, reduplication pairs, orthographic twins (台灣/臺灣, 什麼/甚麼), and erhua forms all kept, none collapsed. Forms cross-link via related[]; orthographic twins additionally carry a scholarNote explaining the split (why 甚 vs 什, 份 vs 分, 佔 vs 占) — which turned out to be a field the schema and style card already define for exactly this (“a variant gloss… author proposes, Wedge approves”). Erhua kept because simplified Mandarin is the next phase after the Traditional dictionary completes.

stage_words.py ran → batches/parallel-wave/word-batch-2026-06-13/slices/: 4,091 gaps (A1 211, A2 202, B1 434, B2 1,164, C1 2,080), 4 disjoint band-ordered slices of ~1,023, each word annotated with its variant siblings + a twin/distinct hint. RUN-PROMPT.md gained the variant-handling rules; the PRD’s open questions are now resolved decisions.

Outstanding

• Seed: prepend the 4,091 into backbone-queue band-ordered — the one remaining destructive step, awaiting explicit go (backup first).
• Then run 3–4 agents on the slices via the parallel-wave driver, KIND=words.
• Still parked: the 289-entry single-source audit (Composer); the 55 C1 chars + their cross-check; reconcile_queues.py --apply (Wedge runs post-campaign).