feat: make internal AI optional and add capabilities endpoint

[jansitarski]

Description

Wardrowbe produces item tagging, outfit suggestions, and pairings through its internal LLM, with no way to turn that off or supply those results externally. This is PR 1 of 4 that lets the backend run with internal AI disabled and defer that work to an external agent (e.g. an MCP server such as saya6k/mcp-wardrowbe or jansitarski/wardrowbe-mcp) while every endpoint keeps working. The design is provider-agnostic. Everything is additive and defaults to internal AI on, so existing deployments are unaffected.

What this PR does

• Capability switches — AI_INTERNAL_ENABLED (master, default true) plus AI_VISION_ENABLED / AI_TEXT_ENABLED, which inherit the master when unset (a false master forces both off). Wired through .env.example, both compose files, and the k8s manifests so the switch actually reaches the containers.
• Guards — require_internal_ai(...) raises before any client is constructed; get_ai_service() and AIService.__init__ are backstops. With AI off the app boots cleanly with no AI_BASE_URL/AI_API_KEY/models set.
• Enforced at every consumer — recommendation/pairing services guard first (so deferral is unconditional); the tagging worker skips and leaves items ready (untagged); the worker skips AI init at startup; scheduled notifications skip cleanly; POST /outfits/suggest and POST /pairings/generate/{id} return a typed 503 ("deferred to an external agent") instead of a 500 or a hang.
• Introspection — GET /api/v1/capabilities reports the effective state (public, no user data); GET /health/ai degrades to disabled when off.

{ "ai": { "vision": false, "text": false },
  "features": { "external_tagging": true, "external_suggestions": true, "external_pairings": true },
  "version": "1.0.0" }

Roadmap — this is PR 1 of 4

Each PR is additive, independently reviewable, and defaults to current behavior:

1. Foundation (this PR) — capability switches, guards enforced at every consumer, GET /capabilities.
2. Item tagging — tagging_status/tagged_by/tagged_at state, an auto_tag ingest flag plus the enqueue-site guard (so vision-off uploads are never queued), a ?tagging_status=pending work queue, and write-back/retag with server-derived agent identity.
3. Deferred resources — replace the interim 503 with well-formed deferred responses, and add agent-authoring for suggestions and pairings (source=agent).
4. Compose — nullable outfit attribute columns (season/formality/palette/notes) and agent-authored outfit composition.

Related Issue

Related to #99 (Built-in MCP server support). This series exposes the API surface any external MCP needs; it does not add a built-in server.

Type of Change

• New feature (non-breaking change that adds functionality)

Checklist

• I have read the CONTRIBUTING guide
• My code follows the project's coding style
• I have added tests that prove my fix/feature works
• New and existing tests pass locally
• I have updated documentation as needed
• My changes don't introduce new warnings or errors

Testing

Full backend suite — 300 tests, all passing against Postgres + Redis. Coverage includes the flag-resolution truth table; the guards (per-capability, the vision-off/text-on isolation case, and the __init__ backstop); the tagging worker skip and startup skip; /capabilities and /health/ai on and off; the typed 503 for suggest/pairings; and fail-fast ordering (generate_recommendation/generate_pairings defer before any location/item validation).

The two background-removal tests are excluded — they require the optional rembg dependency and are unrelated to this change.

Test Environment

• Docker Compose
• Kubernetes
• Local development

Tests Performed

• Full backend suite (300 tests) green against Postgres + Redis.
• Deployed to a k3s homelab cluster and exercised both states:
  • Internal AI on (default): tagging, suggestions, and pairings behave exactly as before.
  • Internal AI off (AI_INTERNAL_ENABLED=false): no AI client is constructed; uploads land ready (untagged); POST /outfits/suggest and POST /pairings/generate/{id} return the deferred 503; GET /api/v1/capabilities reports ai.vision/ai.text as false.

Additional Notes

• The capabilities version is hardcoded "1.0.0" (matching the existing FastAPI app version), kept hardcoded to avoid pulling release-please version wiring into this PR.
• The features flags are static true: they describe the external-agent surface the series delivers; the dedicated source=agent authoring paths land in PRs 2–3.
• require_internal_ai("vision") is implemented and tested but currently enforced through the worker's direct check; it's the hook PR 2 uses at the enqueue site.