Tsumugu: Step-4 char wave passes review (2026-06-13)

Orchestrator review of Opus’s Step-4 authoring wave — the downstream of the morning seed-gate unblock (seed-gate entry). Verdict: ship it. The work is sound, the headline claims hold against the actual repo, and the adjudications are correct rather than merely plausible. One real correction to the agent’s proposed next step; two items to confirm.

What was verified, not taken on report

Re-ran the real validator against the content store (tsumugu/personal/dictionary): OK — 2716 entries valid, ids unique, filenames canonical, exit 0 (1859 chars + 857 words). The earlier-in-the-day green was 2425; the wave added the difference. Both commits are real and scoped — 0c56dc7 (the 56 chars) is HEAD~1 at 58 files (56 chars + LOG.md + one sources file, no stray -A), a804ef2 (the expand_chars.py indent=2 fix) is HEAD. Validation only failed in my sandbox first because of a stale jsonschema + missing pypinyin — same toolchain artifact the seed-gate entry already diagnosed, not a corpus fault.

The three conflict resolutions are paleographically correct, checked against the real etymologies, not rubber-stamped:

• 彌 = 弓 + 爾聲 (grounding note states “not 水” — correct)
• 謎 = 言 + 迷聲 (not 米 — correct)
• 賜 = 貝 + 易聲 (尚聲 reading explicitly rejected — correct)

QA earned its keep: shipped 晶 carries no “Not X but Y” and no hedges; 會 / 蔓 state phonetics honestly with no invention. The flagged fixes landed in the committed files.

The 噅 (U+5645) skip is the right call. Its crib is bodied for 嗯 — header reads ”# 嗯 en”, 說文無此字, composition UNRESOLVED. Nothing in it grounds U+5645. Skipping beats fabricating, consistent with zero-uncertainty-on-student-surfaces.

The correction: crib_diff’s comparator is mis-scoped, not just noisy

The agent’s conclusion is right — green=0 is a comparator artifact, not a broken corpus — but its diagnosis is incomplete, and that changes the fix. Read-only repro of the compare logic (no mutation of the working-tree merge): CONFLICT=1479, green=9. Per-section breakdown is the tell:

• Composition: 1462 / 1488 conflict. Series: 1473 / 1488. ~98% cannot be real disagreement — that is grok and cursor formatting those sections differently. norm() hardening is justified here.
• 說文: 1230 / 1488 (~83%), but the conflicts are genuine divergence, not phrasing. Samples: 三 quotes two different 說文 editions; 丈 is grok-none vs cursor-present (coverage gap); 丘 is short-vs-long of the same entry. This is exactly where the cross-check earns its seat — it is how the 敵 仇也/彊也 misquote got caught.

Decision: do not approve a blanket norm() hardening. Make it section-aware — aggressive structural normalization on Composition/Series, conservative on 說文 (normalize only absence and punctuation, keep content-strict). A blanket fix turns green back on by blinding the one section that catches real misquotes. (Agent flagged the hardening mid-flight and asked; this is the scoped answer.)

Open items

• python3.14 failure-mode note unverified. The agent says it wrote the toolchain failure + brew reinstall python@3.14 fix “to memory”; it is not in Projects/memory/ or the auto-memory index. Likely lives only in its own Claude Code context → at risk of being lost. Needs a durable home (location TBD with Wedge).
• 苛 — re-examined, shipped correctly; earlier flag withdrawn. First pass called it a contested 說文 split needing a two-telling. It is not. Received 說文 is unambiguous (苛，小艸也, small weeds); cursor quoted it right, grok quoted 病也 — a loan-sense (苛 ~ 疴, ailment) mislabeled as the 說文 gloss, same failure class as the 敵 catch. The agent adjudicated to small-weed correctly. Two-tellings is for contested origins; this origin isn’t contested, so meta.notes (resolved misquote) is the right home and openQuestions would falsely signal doubt. No change. Optional only: 苛 does carry an archaic illness sense (苛癢) — left out as the weaker mnemonic for a B2 entry; correct encoding call.

Working-tree note

The uncommitted RUN-SHEET / RUNBOOK / char-queue.json + regenerated cribs are the word lane’s, disjoint from the char commit, left for Wedge’s push — confirmed. Nothing should re-run crib_diff.py before that push; it would clobber the staged merge.