chore(data): resolve five data-plane hygiene items from index audit (#178)

[sun-970]

Summary

Resolves all five items from #178:

1. embeddings_text uniqueness — migration 0026 adds UNIQUE (file_id, chunk_index) with defensive deduplication before the constraint. Integration test proves the constraint rejects duplicates.
2. memories unindexed FKs — partial indexes on source_file_id and created_by_user_id (WHERE IS NOT NULL) to avoid seq scans on ON DELETE SET NULL.
3. memory_relations cascade alignment — FKs now ON DELETE CASCADE to match memories.workspace_id, preventing orphan-edge failures on workspace delete.
4. N+1 in indexgeneration.List — batch-fetches generations for all builds in one query (WHERE build_id = ANY($1)) instead of per-build.
5. Keyset cursor for ListRelations — reuses encodeListCursor/decodeListCursor from memory/list.go with a relation-specific filter hash.

Changes

• server/internal/db/migrations/0026_data_plane_hygiene.sql — schema fixes with Down blocks
• server/internal/db/migrations_test.go — integration test for uniqueness constraint
• server/internal/indexgeneration/service.go + store.go — N+1 fix
• server/internal/memory/relation.go — cursor support
• server/internal/api/ — handler, interface, and test stub updates
• server/internal/memory/memory_integration_test.go — caller updates

Dependencies

This PR was previously stacked on #184. It has been rebased onto main and is now independent.

Test plan

• Migration 0026 applies cleanly on a populated database (not only empty schema)
• TestEmbeddingsTextUniqueConstraint passes with MEM_TEST_DB set
• EXPLAIN for ON DELETE SET NULL path uses the new partial index (item 2)
• ListRelations returns next_cursor when results exceed limit
• indexgeneration.List returns correct generations per build (no N+1)
• Existing integration tests pass

Closes #178

[PeterGuy326]

Gate decision

REQUEST CHANGES on current head 8f08cf6e57685fce5f21be8ac96ddbcfdbd8332f.

Blocking findings

1. The branch is stacked on #184: git log origin/main..origin/pr-185 has 89bb12b as its parent, and the diff therefore includes the live recall producer files. The PR body presents this as the five-item data-plane hygiene change but does not declare the dependency. Rebase/split it, or explicitly document and enforce the merge order so the same files are not merged twice.
2. The acceptance table leaves the migration, EXPLAIN, relation cursor, generation-list, and existing-integration checks unchecked; no checks are reported for this head. Please run and attach the exact-head CI/integration evidence.

Do not treat the unit-level changes as complete until the stacked scope and database acceptance are resolved.

[PeterGuy326]

Current-head CI approval readback (2026-09-10)

Head: 3104d35ceaf877e9ddf27ca570b87c475645e50c. The original pending-approval runs on this head have been approved and real check-runs now exist (18 at this snapshot). No old-head run was approved. Auto-merge remains disabled; this is not a code approval or merge-ready declaration.

• PR Policy / 34427150574 — attempt 2, completed / success
• ByteFolk Security Baseline / 34427138810 — attempt 2, completed / success
• PR Policy / 34427138751 — attempt 2, completed / success
• Validate Agent memory / 34427138816 — attempt 2, completed / failure
• CI / 34427138818 — attempt 2, completed / failure

Failures at this snapshot:

• Web
• Go
• PostgreSQL integration

[sun-970]

Superseded by @PeterGuy326's #195 which covers the same changes on current main. Closing per the '本轮先不合并' directive.