Fix Pixelbot post-merge validation issues

[pierrebrunelle]

The merged 3.0 upgrade could create a fresh catalog, but schema reconciliation and service startup exposed follow-up failures: nested transcript sentence views stalled schema application, JSON schemas changed order after PostgreSQL JSONB persistence, Studio still referenced removed views, and CI action refs were invalid.

This change:

• consolidates video/audio transcript text and embedding indexes onto their iterator views;
• canonicalizes provider JSON schemas and tool declarations so a second schema diff is clean;
• updates Studio and the architecture/database UI to the deployed 19-model Gemini pipeline;
• returns 404 for unknown API paths and fixes the bundled SPA title;
• removes the vendored .agents/skills copy and skills-lock.json; contributors are directed to the canonical pixeltable/pixeltable-skill repository;
• pins checkout, setup-node, and setup-uv to verified immutable release SHAs;
• reads the managed service log from its documented local path because pxt service logs rejects local services in Pixeltable 0.7.7.

Validation:

• Ruff format/check and mypy pass.
• Backend pytest: 12 passed.
• Frontend: 4 tests passed; lint and production build pass; initial chunk 271.64 kB.
• Wheel build and packaged SPA inspection pass.
• Fresh isolated Pixeltable 0.7.7 catalog: dry run reports 19 creates, apply succeeds, and second diff reports 19 up to date with zero unsupported/destructive changes.
• Real managed service starts on port 8000; health, SPA, catalog, and audio/video transcription routes return 200; removed Reve/database mutation and unknown API routes return 404; service stop leaves no managed services.

Provider calls remain mocked or uninvoked. No paid calls, Cloud deployment, or old agents catalog mutation was performed.