fix(core-ts,mcp-server): RFC-0045 step 7a, held rows for old-shape values and whole-list re-translation - #221

RFC-0045 rollout step 7a: the two latent issues to fix before step 7b enables reference and repeater fields. Both types stay refused on every production write path; the shared suites enable them with SiteGraphStoreOptions.fieldTypes.

Closes #215, closes #218.

#215: old-shape values keep their entry_refs rows (core-ts)

• normalizeWrite → onUnconvertible used to return early when convertingPath(def) was undefined. That covers every top-level conversion from references with no text ↔ reference path: reference (multiple) ↔ repeater, reference single ↔ multiple, and a repeater converting away.
• So a stored old-shape value that did not convert was not held, and entryRefStatements dropped its rows on any unrelated write.
• Such failures are now recorded at the old type's field paths: each id-shaped string at <field> from a reference, and each item's id-shaped sub-value at <field>.<sub> from a repeater. heldRefs and entryRefStatements keep them, and every stored failing id is held, as before for conversions from references.
• The rebuild's reshapedRows and the writer share one helper, references/old-shape.ts (readsOldShape, oldShapePairs), so the rows are the same whichever touched the node last.
• Test: src/testing-field-operations-old-shape.ts, registered in the shared suite, so it runs on SQLite, Miniflare D1 and DO SQLite. It fails on all three without the fix.
  • Reference → repeater with a malformed stored list ([id, 5], injected with json_set as an import could hold it):
    1. a title edit keeps the rows, and deleting the target is entry_referenced;
    2. a reimport (the rebuild) gives the same rows and the same refusal;
    3. a write after the rebuild gives the same again.
  • Repeater → reference with an item of another shape: the same rows after a write, a rebuild and a write.
  • Target deletes there are field_operation_running for the whole operation, because a conversion into references holds them (R12). So the row checks carry this case.

#218: re-translating a repeater the translation already holds (mcp-server, core-ts)

1. When the existing translation (the one save_translation writes) holds its own list for a localized repeater, get_translation_source returns the source's whole list for that key under fields, in source order and with every sub-value, instead of item paths. A new wholeLists output names those keys. The logic is in core-ts translationSourceValues (new export); MCP's translationSourceFields takes the existing translation.

2. Updated text:

   • the translation_shape_mismatch hint in field-hints.ts (send the whole list, as get_translation_source returns it, and no item paths for it);
   • TRANSLATED_FIELDS;
   • get_translation_source's fields and wholeLists descriptions;
   • save_translation's description.

   The README tool table is unchanged after node scripts/mcp-tools-doc.ts: it truncates before the edited part, and --check passes.

3. Docs:

   • The RFC's "Amended in 5" bullet gets a short Amended in 7a qualifier: the length check applies only when item paths come in the same call, and a whole list sent alone is the translation's own and stored as sent.
   • A new "Amended in 7a" section records both fixes.
   • The new changeset .changeset/rfc45-step7a-fixes.md (core-ts and mcp-server, patch) qualifies the step-5 wording. The step-5 changeset is untouched.
   • Code comments in fields/translation.ts (baseList) and translation-fields.ts say the same.

4. Tests:

   • mcp-server/test/references.test.ts, "re-translates a stale repeater: list_stale_translations → get_translation_source → save_translation":
     1. the first save goes by item paths;
     2. the source is edited and lists the translation as stale;
     3. get_translation_source returns wholeLists: ["agenda"] and the source's list;
     4. item paths are refused with the new hint;
     5. the handed-over list, translated, saves, and nothing is stale afterwards.
   • core-ts fields-translation.test.ts: unit cases for translationSourceValues, and for a whole list sent alone being stored as sent.

Local checks

• Single files: core-ts test/store.test.ts:
  • better-sqlite3: all RFC-0045 suites, 111 passed;
  • D1 and DO SQLite, each: §7 suites 51 passed, step-2 integrity 49 passed. They ran one slice at a time, because a full D1 pass here exhausts local ports (EADDRNOTAVAIL).
• core-ts fields-translation.test.ts: 11 passed. mcp-server references.test.ts: 17 passed.
• pnpm size:check, typecheck, biome and mcp-tools-doc --check: clean.
• Full pnpm check: result in the session report.

🤖 Generated with Claude Code