Archive v0.7: scheduled catalog observer

[WangEn]

Goal

Turn the v0.6 observeFirstPartyIdentity ingestion primitive into a periodic first-party Catalog Observer that preserves vendor snapshots and continuously feeds the identity-drift engine.

Collection model

Migration 0009_catalog_observer.sql adds:

• catalog_observer_sources: source registry, cadence, parser, credential reference and due time;
• catalog_collection_runs: durable attempt/status/diagnostic history;
• catalog_source_snapshots: append-only raw response snapshots linked to source_records.

Due work is claimed with FOR UPDATE SKIP LOCKED and next_run_at is advanced before HTTP I/O, so multiple runner processes can coordinate without holding database locks across network requests.

Observer

PgCatalogObserver:

1. lazily registers current OpenAI / DeepSeek first-party model-list and docs sources;
2. collects due sources on their own 6h / 24h cadence;
3. hashes and stores every successful first-party response before identity parsing;
4. parses OpenAI-compatible model lists;
5. maps only remote model IDs that match an existing sourced current first-party binding;
6. reuses the fetched source_record when calling v0.6 observeFirstPartyIdentity;
7. records unknown remote IDs and missing known IDs as evidence without auto-creating, retiring, or remapping canonical Models.

Docs are snapshot-only in v0.7: they create source/snapshot evidence but do not make parser-derived identity claims yet.

Runner scheduling

Queue-mode runner enables the observer by default and launches collection beside Run processing. The runner polls PostgreSQL every 60s for due sources; source rows own the longer collection cadence.

Optional runtime controls:

• MODELAPSE_CATALOG_OBSERVER_ENABLED
• MODELAPSE_CATALOG_OBSERVER_POLL_MS
• MODELAPSE_CATALOG_OBSERVER_TIMEOUT_MS
• MODELAPSE_CATALOG_OBSERVER_MAX_RESPONSE_BYTES

A one-shot catalog-admin collect-first-party-catalog command uses the same implementation for operator collection / diagnosis.

v0.6 invariants retained

• every retrieval gets a source-backed immutable snapshot;
• alias history remains append-only;
• execution-binding history remains immutable except one-way close;
• an unchanged model list can append provenance without creating false binding drift;
• lack of an explicit remote snapshot preserves the current known Snapshot;
• unknown remote IDs never become canonical identity merely because they appeared in a list.

Coverage

Adds parser unit coverage and PostgreSQL integration coverage showing two scheduled model-list snapshots with changing provider snapshot IDs produce source-backed alias + execution-binding drift through the existing v0.6 engine while an unmapped remote ID remains evidence only.

Deployment and operator semantics are documented in docs/catalog-observer.md and the Coolify runner guide.