Tsumugu — full Vietnamese translation campaign (paused 87.6%)

Verdict: Opus is translating the entire dictionary into everyday modern Vietnamese, overwriting the old-fashioned Grok bulk .vi. Characters are 100% done; overall 87.6%. Paused on the session usage cap; resumes in ~5h. Everything is committed on vi/data (not pushed) and staged for a clean resume — full pointer in tsumugu-ed/batches/vi/HANDOFF-vi-resume.md.

What we did

Authority is the ratified tsumugu/PRD-vietnamese-opus-everyday-translation.md (“full send”, D1–D5): Opus authors, everyday register, overwrite, ship as labeled beta. We flagged before starting that the PRD predated yesterday’s example-sentence run, which tripled the examples[].vi volume — Wedge confirmed “full send everything” anyway (~36k .vi strings: def / form / story / component glosses / examples across chars + words + patterns + chrome).

Built the VI lane from scratch and ran it as a workflow fleet:

• scripts/vi_checks.py — everyday-register gate (hard bans from native feedback: lâm sản→gỗ, cáo từ, thổ địa→đất đai, liễu giải→hiểu rõ, …; soft warnings → review).
• scripts/merge_vi.py — additive .vi merge (overwrites old, gate-bounces archaic words, tracks progress).
• scripts/vi_stage.py / vi_finish.py — split work into per-agent manifest files; reconcile→merge→commit per wave.
• batches/vi/vi-translate.workflow.js — agents each read their own manifest, translate every .vi field, self-check against the gate before returning. Worked exemplars 休/失眠 calibrate the register.

Register is landing well — natural 25-year-old Vietnamese (“lên ti vi”, “góa phụ”, “lướt điện thoại đến mất ngủ”, “đi khám bác sĩ”), 0 hard-ban hits, 0 merge bounces across all clean waves.

State (paused)

• chars 100% (25,609/25,609) · words 68% (2,269 entries left) · patterns 0% (515) · chrome not started.
• ~30+ scoped commits on vi/data, not pushed.
• 9 waves left: words ×6 + patterns ×2 + chrome ×1.

Lessons (throughput)

• 5 concurrent lanes is the safe ceiling. 6 lanes (96 agents) trips the transient rate limit (recoverable, wasteful — drops ~⅓ of agents, who mostly finish their file writes before the failed return, so coverage stays ~88% and we sweep gaps).
• The session usage cap is cumulative token spend on a timer reset — independent of lane count. That’s what we paused on, not concurrency.
• Per-agent manifest files (each agent reads its own g<gi>.json) beat shared-file index slicing, which failed earlier by re-authoring the same block.

Resume

See tsumugu-ed/batches/vi/HANDOFF-vi-resume.md: re-stage from RESUME-{words,patterns}-gaps.json, run ≤5 lanes, vi_finish + commit per wave, chrome last, then validate.py. Composer Lane B QA still owed (author≠checker); ships as labeled beta.

Related: 2026 06 19 example sentences bakeoff and opus recipe · tsumugu-ed vi/data.