fix(recall): inherit observation entities through source_memory_ids

[youchi1]

Summary

Recall returns entities: null for every observation result, even when include_entities=true is set, while GET /v1/default/banks/{bank_id}/memories/{id} returns the same observation with its inherited entities populated. The asymmetry strips entity context (e.g. URLs, named projects) from observation-only recall and silently drops those entities from the top-level entities aggregate map.

Related: #1374 fixes the same conceptual gap on the document detail view (per-document graph + Observations tab were empty because the observation→source edge wasn't traversed). This PR is the recall-projection sibling: observation results lose their entities for the same underlying reason — code paths that need to traverse the observation→source edge weren't doing it.

Root cause

get_memory_unit (hindsight_api/engine/memory_engine.py) already handles observations explicitly: if a unit has no rows in unit_entities, it inherits entities from its source memories. The recall path's entity fetch only queried unit_entities directly; observations have no direct rows there, so every observation in the recall result lost its entities.

The inheritance edge is dialect-shaped: PG stores it on memory_units.source_memory_ids, Oracle keeps it in the observation_sources junction (same setup #1374 navigates with _observations_via_source_match_sql).

Fix

Two commits:

1. fix(recall): inherit observation entities through source_memory_ids — mirror get_memory_unit's fallback inside recall's include_entities block: for observation result IDs with no direct unit_entities rows, look up source_memory_ids, fetch entities for the union of source IDs, project back. Downstream code derives the per-result entities field and the top-level aggregate map from the same map, so inheritance flows through both automatically.

2. refactor(recall): consolidate observation entity inheritance in one SQL helper — replace the procedural fallback with _entity_rows_for_units_sql, a private engine helper that returns one dialect-correct UNION SELECT covering both direct and inherited entity rows. Mirrors the shape of _observations_via_source_match_sql from fix(api): include observations in per-document graph and counts #1374. Also fixes silent Oracle-parity gap (the procedural patch reached for source_memory_ids directly, which is PG-only). get_memory_unit is refactored to use the same helper, so the two call sites that hand-rolled the same inheritance now share one source of truth.

Reproduction (before fix)

# Same observation via memory get returns entities
curl -sS "$API/v1/default/banks/$BANK/memories/$OBS_ID" | jq '.entities'
# => ["HeadClaw","Reddit","waitlist users"]

# But via recall they are stripped, even with include.entities forced on
curl -sS -X POST "$API/v1/default/banks/$BANK/memories/recall" \
  -H 'Content-Type: application/json' \
  -d '{"query":"...","types":["observation"],"include":{"entities":{"max_tokens":1000}}}' \
  | jq '.results[0].entities, .entities'
# => null
# => null

A mixed-type recall (["observation","world","experience"]) populates the top-level entities map but only via world/experience contributions; observation results still come back with entities: null.

Test plan

• New regression test (tests/test_recall_observation_entities.py) seeds:
  • A world fact with two entities linked via unit_entities,
  • One observation referencing that fact via source_memory_ids and zero direct entity links,
  • One observation with its own direct unit_entities link,
    and asserts that:
  • both observations come back from recall_async with the expected entities (inherited and direct),
  • the top-level entities aggregate map is populated with all of them,
  • get_memory_unit on the same observations also surfaces inherited and direct entities (guards against drift between the two call sites that share the helper).
• Test fails on main with AssertionError: Observation entities must inherit through source_memory_ids; got [] and passes after the fix.
• tests/test_observations.py and tests/test_graph_filtering.py continue to pass.
• ruff check and ruff format clean.