Tsumugu-ed: composer batch + backlog (2026-06-17)

Verdict: The deferred Composer-native maintenance pass is done — validation green, zhuyin cross-check clean. Next value is either learner-facing (example-sentence pass) or encoding-facing (phonetic-family browse, handwriting lane). Grok Build plan mode remains unsafe on CJK filenames until the shell UTF-8 panic is fixed; use exec-mode Python for tsumugu-ed mechanical work.

What ran (2026-06-17)

Direct session execution via scripts/run_composer_batch.py and new maintenance scripts (no subagents — prior subagent launches cancelled; plan mode crashed Grok on kui1-虧.json byte-boundary panic).

Task	Result
Char audit (虧, 含)	Drift fixed
related trim	24 words → all ≤2
Erhua zhuyin	gen_helpers.py patched; 18 *r5* entries finalized
Level normalize	4,771 entries → TOCFL n (est.)
Def prefix normalize	639 adj/adv → adjective/adverb
Phonetic report	batches/maintenance-2026-06-14/PHONETIC-STATUS.md (59 sound / 85 no-sound)
Register lint	scripts/register_lint.py — 13 hits, mostly false positives
Validate	OK — 9663 entries + 515 patterns

Scripts added: run_composer_batch.py, normalize_levels.py, normalize_def_prefix.py, phonetic_unresolved_report.py, register_lint.py.

Grok shell constraint

grok-shell plan mode panics on UTF-8 filenames when slicing at byte index inside CJK (kui1-虧.json). Workaround: exec/agent mode; drive edits through python3 scripts/... with ASCII-safe globs. Already noted in production-line memory.

Lint (term)

Lint = automated checker that scans text or code for rule violations and reports them without editing. Named after lint rollers — the fuzzy fluff caught when clothes are inspected before wear. In software: first widespread use was lint (1978, Stephen C. Johnson, Bell Labs) for C code — caught suspicious constructs a compiler would allow. Modern linters (ESLint, register_lint.py) follow the same pattern: project rules, scan, flag. Our register_lint.py enforces STYLE-CARD §Banned register; needs tightening to exclude gloss definitions (e.g. 也許 → “possibly”) and structural lookalike lines in related.why.

---

Backlog — what else to do

Recorded from Composer review session 2026-06-17. Ordered by lane.

Corpus quality (mechanical, high signal)

1. Phonetic UNRESOLVED adjudication — 85 chars on worklist have no sound role (PHONETIC-STATUS.md). Confirm genuine 會意 vs unresearched; mark resolved in entries.
2. Tighten + wire register lint — refine patterns; add to audit_gates.py for CI; exclude definition glosses and lookalike related.why contrasts.
3. Extended drift audit — optional read-only finder pass on char corpus (449-char audit found 6 fixes; pattern may surface more).
4. Commit + push — ~5,400 entry touches from normalize passes plus surgical fixes; scoped commit before more authoring.

Product / learner value (bigger lifts)

5. Example-sentence pass — ~99% of words still lack examples; largest gap vs site promise (“Examples inside the level” still ◌ in progress).
6. Phonetic-family browse — composition rows already name series (方 → 放房訪防); needs index + page template.
7. Handwriting lane — Outlier course JSON still minimal; enrich from DevTools capture; crib batch for 手寫. Complements 2026-06-17 YGSF integration journal.
8. Site copy fix — home/method says stories in “every entry”; 77% of words correctly have story: null (transparent compounds). Distinguish char vs word in copy.

Growth (reach, not polish)

9. 579 frequency-tail queued words — beyond TOCFL; parallel-wave when reach past exam list is wanted.
10. C2 coverage — needs external 精通級 word/char list first (華語八千詞 tops at C1).
11. 成語 idiom segment — flagged on site as next on the loom; not in corpus yet.

Tooling

12. Avoid Grok plan mode on tsumugu-ed until CJK path panic fixed.

Suggested pick-one next

• Learner impact: example-sentence pass (TOCFL 1–3 high-freq words first).
• Encoding impact: phonetic-family browse (mostly mechanical).