translate: Finish translation

2026-06-18 09:16:10 +08:00
parent b849f7429f
commit 8c1feb0a31
6 changed files with 162 additions and 180 deletions
@@ -212,6 +212,7 @@ Do not run extraction pipelines until scope is clear.
 - **Verify patches with `cat`** — `read_file` dedup makes it unreliable
 - **Re-read before fixing** — user may have made interim edits
 - **Em-dash drift (translation mode only)**: AGENTS.md mandates `—` (Unicode em-dash) → `---` (three hyphens) in English djot for the translation workflow. When proofreading an existing DOCX manuscript, do NOT flag em-dashes — the manuscript's English is authoritative and this is a formatting concern for the translation workflow. If you're in translation mode and the target has Unicode em-dashes, run a single find/replace `—` → `---`. Long files typically have 30–50 such instances.
+  - **Before raising this**: confirm the user wants mechanical cleanup. The user may deliberately leave `—` in the .dj because it's their working source — the conversion to `---` is something they do themselves at the docx-export step, or they may want comments/findings only. Default: ask before applying a 100+ instance find/replace. If unsure, mention it in `translation-findings.dj` with a note that it's mechanical, but do NOT auto-apply.
 - **Batch terminology lookups** — when checking many terms against the terms DB, run them in one `execute_code` script that loops over a query list and calls `search.py` via `subprocess.run`. One terminal call per term floods the context with repetitive output.
 - **Proofread ≠ translation review** — when the user says "校对" or "proofread" and the input is a DOCX manuscript with existing English, you are in proofread mode. Do NOT flag translation quality, terminology, or djot formatting. Do NOT apply patches to bilingual.dj unless asked. Write `edit-suggestions.dj` with manuscript-level issues only. If the user later asks for translation review of the same article, write findings to a separate `translation-findings.dj`.