feat(webv2): reference trims as start + sample length with steppers; 200-frame default

[lstein]

Why

Reference frames cost denoise VRAM on every step (the packed sequence grows with reference length), so the useful sample of a reference video is a few seconds — but picking a small interval with independent start/end sliders meant fighting both ends.

What

• The video reference card's second control is now the sample length in frames (min 1). The start slider slides the whole window at constant length, stopping at the clip's end — a drag overshoot can't shrink your sample. The row's thumbnail still shows the resulting end frame (badged with its frame number, since the number field now shows the length).
• New video references default to a 200-frame window (~8 s at 24 fps, DEFAULT_REFERENCE_SAMPLE_FRAMES) instead of the whole clip.
• All trim fields gain increment/decrement steppers (both reference rows and the Initial Video's start/end) for single-frame nudging.
• The Ref2VA extend anchor keeps its seam contract: its end frame stays pinned to the Initial Video cutpoint — the start slider adjusts the lead-in, and the length control grows backward from the pinned end. (The adversarial review caught the first cut breaking this: a start-slider round trip could silently detach the anchor's window from the seam.)

Storage is unchanged — clips persist startFrame/endFrame, so the request contract, normalization, metadata recall and old generations are untouched; length is derived.

Testing

• The window math lives in core/settings as pure functions with 11 new unit tests (constant-length slide, wall stop, anchor pinning both directions, corrupt-trim self-healing, single-frame clips).
• pnpm lint, unit suite (7436), browser suite (1036), and the architecture gate all pass.
• Adversarial fresh-context review run; all three findings (anchor seam detachment, overshoot ratchet, corrupt-trim repair regression) fixed in the second commit.

🤖 Generated with Claude Code