fix: use configured model name in ai_enricher instead of hardcoded llama3.2 ...

- Remove hardcoded 'model_version': 'llama3.2' from system prompt
- Remove hardcoded 'llama3.2-mock' from mock response
- Inject actual settings.ollama_text_model into enrichment result programmatically
- Ensures DB records reflect the real model used (gemma4:e2b-it-q4_K_M)

Co-Authored-By: Claude <noreply@anthropic.com>

fix: code review critical and high issues ...

- tenacity+structlog: replace before_sleep_log with structlog-compatible lambda
  to prevent TypeError on retry
- NormalizedProperty: filter AI response dict by allowed dataclass fields before
  unpacking to avoid TypeError on unknown keys
- property_pipeline: remove duplicate update_status(failed) from _stage_normalize
- security: add URL validator (SSRF protection) for ImageDownloader and archive-check
- ai prompts: replace raw <user_data> tags with JSON-serialized payload
  to mitigate prompt injection
- queue_worker: wrap _process_one in try/except so DB errors don't kill the loop
- image processing: parallelize with asyncio.gather + Semaphore(3)
- ai services: unify OllamaFatalError handling — all propagate instead of swallow
- router_properties: catch only pydantic.ValidationError/ValueError in ingest,
  let infrastructure errors return 500

Co-Authored-By: Claude <noreply@anthropic.com>

feat: classify AI errors — ValidationError→invalid, fatal/network→failed ...

Item 6 from REVIEW_FOLLOWUP:
- ai_normalizer: pydantic.ValidationError returns is_real_estate=False (invalid)
  OllamaFatalError / unexpected exceptions now raise → pipeline status=failed
- ai_enricher: ValidationError returns None (skip enrichment), fatal returns None,
  unexpected raise → pipeline status=failed
- pipeline._stage_enrich: propagate OllamaRetryableError so pipeline can mark failed

Co-Authored-By: Claude <noreply@anthropic.com>

feat: implement review items 8-14 ...

- Soft-delete/archive for listings: archived_at column + archive-check endpoint
- Rate limiting on /ingest: slowapi with 60/minute per source_slug
- Prometheus metrics: /metrics endpoint + custom counters/histograms
- Graceful shutdown: track active jobs in app.state, wait up to 30s
- Prompt injection protection: wrap user data in <user_data> XML tags
- Image download size limit: 50MB max with httpx streaming
- Raw data cleanup: admin endpoint to delete completed raw data older than N days

Co-Authored-By: Claude <noreply@anthropic.com>

feat: implement review items 1-7 ...

- Decompose PropertyPipeline into 8 explicit stages with PipelineContext
- Add tenacity retry (3 attempts, exponential backoff) to OllamaClient and ImageDownloader
- Add simple in-memory circuit breaker for Ollama calls
- Resize images to 1024px before base64 encoding for vision model
- Add /health endpoint (DB, Ollama, disk checks)
- Add DB performance indexes + alembic migration
- Classify AI errors: OllamaRetryableError vs OllamaFatalError
- Add strict PayloadSchema validation for ingest endpoint

Co-Authored-By: Claude <noreply@anthropic.com>

feat: core pipeline + FastAPI API (Phases 0-6) ...

Implemented the full VMK Data Collector foundation and processing pipeline:

- Config, logging, exception hierarchy
- DB models (listings, raw data, images, snapshots, enrichments, custom fields)
- Alembic async migrations
- Repositories with upsert/snapshot support
- Domain entities and Pydantic schemas
- Ollama AI client with mock mode
- AI normalizer, image analyzer, enricher
- Image downloader with SHA-256 dedup
- PropertyPipeline: raw -> AI validate -> upsert/snapshot -> images -> enrich
- FastAPI app with /api/v1/ingest endpoint

Co-Authored-By: Claude <noreply@anthropic.com>