feat(fingerprints): wire generic ingestion into the fingerprint pipeline

[shiba4life]

Summary

Every record written through IngestionService now passes through the text extractor, so emails and phone numbers in normal user-data ingestion (Notes, Contacts, smart-folder, file-upload, Apple imports) become Fingerprint / Mention / Edge records — and contact-shaped sources auto-form tentative Personas. Before this change, only the dedicated /api/fingerprints/* endpoints fed the identity graph; the rest of ingestion was disconnected from it.

The change lands as five phases in one PR:

• Phase 0 — auto_propose::maybe_spawn_persona_sweep helper centralizes the post-ingest sweep spawn across ingest_text_signals_batch, ingest_photo_faces_batch, and import_contacts (the last of which previously never triggered the sweep — bug fix).
• Phase 1 — SignalBinding enum on the text extractor. Strong emits StrongMatch (0.95) edges; CoOccurrence (default) preserves existing 0.3 behavior. Shared-inbox demotion (info@, support@, …) still applies in both modes — a contact card listing info@acme.com lands at 0.475, below the 0.85 sweep floor by design.
• Phase 2 — schema_policy::binding_for_schema allowlists Contacts / AddressBook / AppleContacts / CalendarEvent / AppleCalendarEvent / EmailHeader for Strong binding. Everything else stays on CoOccurrence.
• Phase 2.5 — SYSTEM_DESCRIPTIVE_NAMES mirrors the 12 fingerprint subsystem schemas locally; a unit test pins it to the upstream PHASE_1_DESCRIPTIVE_NAMES so it cannot drift.
• Phase 3 — ingestion::fingerprint_hook::run_after_batch runs the extractor over every non-system mutation after mutate_batch succeeds, then spawns the persona sweep if anything yielded. Wired into both execute_mutations_with_tracking and the per-item decomposition.rs path.

Invariants

• Best-effort: per-record extraction failures log at warn and never propagate. The user-data write already succeeded before the hook runs.
• Recursion-safe: writes on the fingerprint subsystem's own schemas are skipped via is_system_descriptive_schema, so a Mention insert does not trigger another extraction pass.
• Idempotent: ingest_text_signals keys its Mention + ExtractionStatus on (source_schema, source_key, "text_regex"), so re-ingesting the same record converges to the same row.

Out of scope

Photo face extraction in the generic ingest path (would need ONNX inline), name extraction (P2-6 NER), per-schema identity_assertion config field, and Apple-import side-call cleanup at apple_import.rs:904 (now duplicate work via the hook — follow-up PR).

Test plan

• cargo clippy --workspace --all-targets -- -D warnings (clean)
• cargo test -p fold_db_node --lib (859 passed, 0 failed)
• cargo test --workspace --all-targets -- --skip cli_does_not_auto_spawn_daemon_when_port_unset (all green; the skipped test is a pre-existing flake on main)
• New unit tests cover: SignalBinding::Strong emits StrongMatch at 0.95; CoOccurrence keeps 0.3; shared-inbox demotion in both modes; binding_for_schema allowlist + case-insensitivity; SYSTEM_DESCRIPTIVE_NAMES ↔ PHASE_1_DESCRIPTIVE_NAMES parity; collect_string_fields walks nested objects/arrays and skips non-strings; source_key_for uses KeyValue Display + falls back to uuid.
• Manual end-to-end demo: paste sample_data/contacts/address_book.json into the Ingestion tab → wait ~30s → tentative personas appear in People → Personas. (Will run before merge.)

🤖 Generated with Claude Code