feat(webv2): video panel entry points and polish

[lstein]

PR 7 — the top of the Video generation panel stack (on #129).

What this adds

• Launchpad: a "Generate a video" intent tile — /app?new=true&intent=video opens a compose-preset draft with the invoke route set to video.
• Gallery context menus: Send to Video on images seeds the first frame; Extend in Video on videos seeds the initial video with its default trim. Both open the widget, respect the first-frame ⊻ initial-video exclusion, and route the invoke source (same semantics as the existing Send to Upscale).
• Widget-header reset: restores every model-governed setting (frames, fps, target resolution, steps, CFG, accelerator state) to the selected model's defaults — prompts and conditioning media untouched.
• Docs: the video-generation page now leads with the panel before the workflow templates.

Review findings fixed (fresh-context adversarial pass)

• Video normalization now heals partial records field-by-field, upscale-style. Previously it returned null for anything missing the core shape — so seeding a never-opened widget from the context menu produced values the mount-time reconciler would wipe, silently discarding the payload on the feature's most likely first-contact path. Healing also repairs any other partial-write path (rollback restores included); readiness still blocks correctly through the model-required check.
• The header reset gates on a loaded model catalog — defaults computed from an empty catalog would have claimed the accelerator pair was "not a default" and stripped it as the reset.
• The reset's at-defaults check compares LoRAs as a set (order is insertion history, not a setting), so adding a concept LoRA no longer lights the button spuriously.

Full suite 477 files / 6329 tests (+ 756 browser tests in the review run); tsc / oxlint / format / i18n-key coverage clean.

🤖 Generated with Claude Code