Auto-transcribe uploads, one-button generate all, accept-all tags

[davior]

Summary

Addresses a batch of AI-enrichment feedback:

• Video/audio now auto-transcribe on upload. ingest_upload chains a transcribe job the same way transcription already chains into embedding — guarded on a Deepgram key actually being configured, so an unconfigured library gets no spurious activity rows.
• One "Generate all" button. New generate_all job kind runs summarize → describe → autotag serially over one asset (backend/app/enrichment/generate_all.py), with its own endpoint (POST /api/assets/{id}/generate-all). Order is summarize-then-describe (not the UI order) because describe's prompt already knows to skip repeating an existing summary — running summarize first is what gives that check something to check against. Autotag runs last so it can read both.
• AI triggers moved beside their field labels. Description, Summary and Tags each get a small icon-only button next to their label instead of a full-width "Describe/Summarise with AI" button below the field. EnrichmentButton gained an iconOnly mode; TagInput gained a labelAdornment slot. "Generate all" sits at the top of the panel.
• "Accept all" for suggestions. SuggestionPanel no longer owns its own trigger (that moved to the icon button beside Tags) and now offers one click to accept every pending tag/title suggestion, watching both autotag and generate_all jobs for the reload signal.
• Description/summary limit raised 2000 → 6000 characters, and each prompt now states its own limit explicitly, so a reply that used to be silently clipped mid-sentence either fits or the model stops itself in time.

Test plan

• cd backend && pytest -q — 754 passed, same 6 pre-existing failures (missing ffmpeg on PATH / an unrelated httpx typing issue) confirmed via git stash baseline, no new failures.
• cd frontend && npm test && npm run build — 256 passed, typecheck + build clean.
• npm run lint / prettier --check clean on changed files.
• Manual check: ran both dev servers, signed in with a self-minted JWT, uploaded an image, and exercised every new control in a real browser (icon buttons next to Description/Summary/Tags, "Generate all", and "Accept all" against seeded suggestions — title and tags applied correctly in one click).

🤖 Generated with Claude Code

https://claude.ai/code/session_019RTuL5VnMWCM6j7hDiw46X

---

Generated by Claude Code