feat([issue-4188]): mood board items gain persisted prompt-from-media analysis

[atomantic]

Summary

Phase 3 of #4188 (mood boards as a universe style-building tool): board items backed by gallery media can now be analyzed with the Prompt-from-media flow, and the result persists on the item.

• Analyze with AI on board items — gallery-backed items (an image:<file> media-key, a gallery-shaped imageUrl, or a video:<filename> media-key) gain a ScanEye action that opens PromptFromMediaModal seeded with the item's media. The analyzer is unchanged for its existing hosts; it gains an optional onResult callback (threaded through the modal along with kindDefault/children) so a host can persist a successful run.
• Persisted analysis on the item — a run auto-saves { prompt, negativePrompt, rationale, providerId, model, analyzedAt } onto the item via the item PATCH (moodBoardItemAnalysisSchema, media items only, null clears; analyzedAt stamped server-side when absent). Additive on the wire — whole-record LWW spreads unknown keys and older peers preserve them (same precedent as the pinterest sub-object), so no moodBoards gate bump. The card shows an accent-highlighted analyzed state; the modal shows the saved analysis with copy/remove.
• Gallery video resolves by filename — POST /api/media-jobs/prompt-from-media accepts sourceKind:'video' with a filename in place of the history videoId (board video items store the on-disk filename, not an id). The extraction id is stable per clip (pfm-vf-<stem>), so re-analysis overwrites its sampled frames instead of accumulating sets under data/video-thumbnails. The gallery flow's id path is unchanged.
• Analysis-source resolver (moodBoardItemAnalysisSource) mirrors the server's imageUrlToAppAsset contract: legacy bare gallery refs and query/hash-suffixed /data/images URLs normalize to the on-disk basename; external URLs, non-gallery paths, text items, and legacy video:<id> pins on image items are not analyzable (no button).

Analysis is per-item and user-triggered with an explicit vision provider/model picker — no batch auto-analysis (AI Provider Usage Policy).

A codex review pass surfaced two P2s, fixed in the third commit: the stable frame-extraction id (was a fresh UUID per run → unbounded thumbnail accumulation) and the legacy/query-string gallery-URL normalization (those items previously got no analyze action, or sent the query string as part of the filename).

Remaining phase of #4188 (board → universe style synthesis) is not in this PR — see the reconciliation comment on the issue.

Refs #4188

Test plan

• server/lib/moodBoardValidation.test.js — analysis PATCH cases: full/minimal accepted, null clears, rejects missing prompt / unknown keys / non-ISO analyzedAt.
• server/services/moodBoard/logic.test.js — normalized persistence (explicit nulls + stamped analyzedAt), client-provided analyzedAt kept, other patches leave it untouched, null clears, ignored on text items.
• server/services/mediaPromptFromMedia.test.js — filename-based video resolution with the stable pfm-vf-<stem> extraction id and no history lookup; extension-less filename rejected before any ffmpeg work.
• server/routes/mediaJobs.test.js — video source with neither videoId nor filename 400s.
• client/src/lib/moodBoardItemSrc.test.js — analysis-source resolution: video-by-filename with poster, image by media-key/app path (decoded), legacy bare refs, query/hash stripping before basename, and the null cases.
• client/src/components/media/PromptFromMedia.test.jsx — onResult fires with the payload; a filename-only clip sends no videoId.
• client/src/pages/MoodBoardDetail.test.jsx — analyze action only on gallery-backed items, persist PATCH payload and analyzed-state flip, saved-analysis view, remove via analysis: null.
• Touched suites green locally; biome check clean. Full-suite runs were degraded by extreme machine load (load average ~280) — every failing file also fails identically on main and none are touched by this change; CI is the authoritative gate.