feat(webv2): video generation parameter recall

[lstein]

PR 8 of the Video generation panel stack — stacked on #131. Gallery videos gain Recall Metadata.

Backend

Nothing needed — verified rather than assumed: VideosInterface.save already persists the graphs' core_metadata through the WithMetadata mixin fallback, and GET /api/v1/videos/i/{name}/metadata serves it (the August video-metadata groundwork plus PR 3's graph stamping pay off here).

Frontend

A sibling of the image-recall machinery (videoRecall.ts / executeVideoRecall.ts), with the video-specific semantics:

• Verbs: Recall All / Remix Video / Use Prompt / Use Seed on the gallery video context menu, capability-gated by what each video's metadata actually contains (gated on the nine video generation_mode stamps — image metadata never recalls into the Video panel).
• Model by key, family transition first — recalling a Wan run while the panel holds H3 snaps frames/fps/resolution before the recorded values land; an uninstalled model is skipped and everything validates against the current one.
• Mode is part of the recall, but never recalled directly: conditioning media is restored (re-hydrated against the gallery for dimensions/probe data; deleted media drops) and the mode falls out. Held media the recorded run didn't have is cleared — media selects the graph family, so "Recall All" on a t2v video must not compile an i2v graph. In extend mode the recorded first_frame_image (the frame the graph extracted from the source clip) is deliberately not recalled.
• LoRAs recalled by key — the first recall implementation to handle LoRAs at all — reproducing the recorded set exactly (including empty / no-longer-installed → cleared), with the accelerator toggle restored when the recalled set is precisely what the toggle would install at its step count.
• Reuse-primary CFG semantics survive: a missing guidance_scale_low_noise key restores null (reuse), not a stale explicit value.
• Size recall is exact-match only — recorded W×H inverts to an aspect+resolution preset only when a preset pair reproduces it precisely; a near-miss would claim a preset the run never used.
• Prompts route through the shared project draft (where the compiler reads them); an empty recorded negative — indistinguishable from disabled — doesn't flip the toggle on.

Verification

• 22 mapping tests covering round-trip fidelity, capabilities, media/mode rules, accelerator restoration, and the review regressions
• Adversarial fresh-context review executed compile→recall→recompile round-trips against the real graph compiler; all four confirmed findings fixed (its own broken browser-test assertion, lost null-low-noise semantics, uncleaned absent media changing the generation mode, and the stale-snapshot window — now snapshotting after the metadata fetch and skipping the values write for prompt-only recalls). The remaining lost-update window during media hydration is shared with — and no wider than — the image recall path.
• Full suite 478 files / 6359 tests; browser context-menu suite; tsc / oxlint / format / architecture:check clean

🤖 Generated with Claude Code