feat(webv2): video conditioning media — keyframes, initial video, and trim

[lstein]

PR 5 of the Video generation panel stack — stacked on #127. All five conditioning modes now work from the panel.

What this adds

Initial Video Frames — first- and last-frame fields (gallery drag, file upload with account-scope-guarded async, clear), rendered per the capability matrix: both for I2V-A14B and MiniMax H3, first-only for TI2V-5B, neither for T2V-A14B.

Initial Video — gallery/upload intake building the source-clip record (frame count estimated from duration × fps — the records store no exact count), start/end trim sliders with a live frame preview (one long-lived muted <video> seeked via currentTime, the frame-scrubber trick — no per-drag-tick refetches), and a 2-frame minimum trim enforced in the setters, the defaults, and validation (the crossfade join consumes a 2-frame tail; a 1-frame trim would fail mid-encode).

Estimate-proof trim compilation — a trim end at or near the estimated ceiling compiles as a negative index, which the backend resolves against the clip's real frame count, so "keep to the end" can never land past it on a VFR upload whose estimate overshoots (the bundled templates' end_frame: -2 robustness, preserved). Mid-clip picks stay positive.

Mutual exclusion — as designed in the stack's kickoff: first frame ⊻ initial video (each setter clears the other; whichever is set disables the other's field with an explanation), while a last frame combines with either — FLF2V interpolation with a first frame, the extension's destination with a source video.

Dimension/FPS locking — conditioning media disables the aspect preset and orientation swap, and the derived-size readout names its source ("from the first frame"); Wan's FPS control disables in extend mode and the duration readout switches to the source clip's rate (the extension inherits it).

Housekeeping — deleted gallery images/videos reconcile out of the panel (patching only the cleared slots), the optimistic-rollback registry covers the three media slots, and a media value the current model can't consume (persisted from another model, or restored by a rollback) renders as a visible clear-stub instead of blocking invoke invisibly.

Smoke-test follow-ups (added during review)

• The frames section is now titled "Start & End Images" (was "Initial Video Frames").
• Trim bound thumbnails — the Initial Video section shows two small always-visible previews of the start and end frames of the kept range (same one-element seek trick), so both ends stay in view while either slider moves.
• Preview drag parity for videos — the preview widget's video frame gets a corner grip handle carrying the same gallery-item drag payload as gallery thumbnails, so a previewed video can be dragged into video-accepting inputs just like a previewed image. The <video controls> surface itself can't be the handle (drag activation would fight the native seek bar), and the grip is deliberately unfocusable: a focusable activator would let the KeyboardSensor start an invisible Enter/Space drag that Tab then drops on the closest-center droppable — a defect the follow-up's own adversarial review caught, fixed, and pinned with a browser-test assertion.

Verification

• Adversarial fresh-context review executed against the commit; all five confirmed findings fixed with regression tests (estimate-overshoot trim failure, sub-2-frame trims — including the default on ultra-short clips, invisible policy-hidden media blocking invoke, wrong duration readout in extend mode, whole-settings rewrite in the deletion reconciler). One race-gated finding is accepted and documented: a failed gallery deletion's rollback can restore a first frame alongside a later-set initial video; normalization resolves it first-frame-wins, and the clear-stub/setters recover.
• 117 video-feature tests; full webv2 suite 477 files / 6328 tests; tsc / oxlint / format clean; browser perf baseline regenerated.

🤖 Generated with Claude Code