[gedaiu]

Brings re-dev up to upstream dev at ab09b4a1: 5 commits, deeplethe#1349 (two worker slots stay free for jobs without model calls), deeplethe#1280 (a client reads a document back by external_id), deeplethe#1331 (containment recall starts from name facts), deeplethe#1326 (a GitHub repo source skips empty files, its config is editable) and deeplethe#1328 (a GitHub or Jira source needs its repo or project). The dev mirror was fast-forwarded to upstream first.

Conflicts.

• Containment recall reads the base's name facts through the predicate index instead of probing every entity's facts deeplethe/utopia#1331 vs our trigram containment (Containment matches a name's % and _ literally instead of as wildcards (#1320, ahead of upstream) #11, Containment recall reads indexes instead of scanning the base #12, 9002): kept ours. resolution.rs is unchanged by this merge (CONTAINMENT_SQL, contained_names, SET LOCAL jit = off, the Recall guard). The designs differ in where recall starts. Upstream starts from the base's known_as facts through facts_live_predicate_idx and joins back to entities, so an entity with no name fact (one created from a description, 0044) is never a candidate. Ours scans entities.canonical_name (trigram index, plus the reversed substring lookup) and unions the known_as facts for aliases, so description entities are candidates. Containment matches a name's % and _ literally instead of as wildcards deeplethe/utopia#1320's escaping tests and the alias pair test (Mr. Holmes / Sherlock Holmes through Holmes) pass. Containment recall reads the base's name facts through the predicate index instead of probing every entity's facts deeplethe/utopia#1331's a_described_entity_is_not_paired_by_containment does not pass against ours; I kept the test with the opposite expectation (a_described_entity_is_paired_by_containment_through_its_canonical_name) so the difference is pinned. Whether to adopt upstream's behaviour is a separate decision.
• A client reads a document back by the external_id it pushed it under deeplethe/utopia#1280: upstream's route, tests, pushed_key/is_push_kind and read_retained(.., serves) taken whole. Our older copy had include_missing and a missing_since field in the response; both are gone, as are the ADR 0052 and ingest.md sentences about them. A tombstoned identity is now always 404. Our token read on GET /documents/{id}/extractions (A token reads a document's extractions and exports its base deeplethe/utopia#1266) was kept. ADR 0014 keeps our revision line and not upstream's; web/src/docs/ingest.md keeps our "A name you push is matched by its spelling" bullet.
• Two worker slots stay free for jobs that make no model call, so a sync runs even while model work holds every other slot deeplethe/utopia#1349, A GitHub repository source skips empty files, and its repository, token and globs can be changed after it is created deeplethe/utopia#1326, A GitHub or Jira source cannot be saved without the repository or project its sync needs deeplethe/utopia#1328: no conflicts.

Behaviour to know about.

• embed_names is still priority 1 (job_priority from 0116 is unchanged and no migration came in). The new reserved lane (Two worker slots stay free for jobs that make no model call, so a sync runs even while model work holds every other slot deeplethe/utopia#1349) only claims priority 0 jobs, so embed_names waits behind extraction as before. Adding it to the list is still the 9003 decision.
• A client that relied on include_missing=true on the read-back route now gets 404.

Schema version 107. No upstream migrations came in; the migration count is still 107 and CURRENT_SCHEMA_VERSION is unchanged.

Checked on the merge: migrations ran twice, cargo fmt --all -- --check and cargo clippy --workspace --all-targets -- -D warnings are clean, UTOPIA_TEST_REQUIRE_DB=1 cargo test --workspace is green: 1675 passed, 0 failed, 6 ignored (server 708 passed, store integration suite 330, store unit tests 91). deeplethe#1326 changed web files, so the web checks ran: pnpm install --frozen-lockfile, tsc --noEmit clean, vitest run 315 tests in 46 files.

🤖 Generated with Claude Code