feat(minimax-h3): linear UI — generate-tab t2v/i2v and image output mode (PR 4/5)

[lstein]

Summary

PR 4 of the MiniMax H3 stack (on top of #140): the linear UI. H3 becomes usable from the Generate tab (t2v + first-frame i2v with a muxed stereo soundtrack) and, in image output mode, from the canvas as a txt2img base.

What's here

• Params state: minimaxH3DurationSeconds (5–14 s slider; the 14 s stop maps to the model's true ceiling of 345 frames = 14.375 s) and minimaxH3OutputMode ('video' | 'image') with safeParse-guarded reducers and selectors. Redux-persist _version 4 → 5 with a migrate step seeding both keys (fixtures follow the per-version key-intersection convention). ⚠ The unmerged flux2 branch also claims v5 — whichever merges second renumbers to v6.
• Graph builder (buildMiniMaxH3Graph.ts): one prompt node (guidance-distilled — no negative prompt, no CFG). Video mode is Generate-tab only: denoise.video_latents/audio_latents → minimax_h3_latents_to_video (+ video/audio VAEs), modes minimax_h3_t2v/minimax_h3_i2v. Frame counts snap to the 17n+5 grid within [124, 345]. Image mode runs the 5-frame minimum block through minimax_h3_latents_to_image + NSFW/watermark (mode minimax_h3_txt2img), canvas-capable.
• i2v coupling: the first enabled minimax_h3_reference_image feeds the SAME image + dimensions to BOTH the text encoder (vision context) and the frame-conditioning node — the backend denoise node enforces this pairing (PR 3).
• Ref image type minimax_h3_reference_image (model-less), registered across validators, conversion listeners, and default-config hooks; SUPPORTS_REF_IMAGES_BASE_MODELS gains 'minimax-h3'. The negative-prompt list deliberately does NOT.
• Dispatch/readiness: 'minimax-h3' cases in both enqueue hooks; canvas readiness blocks video mode with a "video runs on the Generate tab" reason; optimal dimension 768, grid 32.
• UI: output-mode toggle + duration slider in the Generation accordion (duration hides in image mode); the dead SD Scheduler and CFG Scale controls are hidden for H3. Metadata recall for both new params; "MiniMax H3" strings verbatim (license attribution).

Adversarial review fixes (already in)

• H3 added to BASES_WITHOUT_STANDARD_SCHEDULER and excluded from CFG Scale — both were rendering as dead controls for a guidance-distilled model.
• steps clamped to ≥ 2 in the builder (the H3 denoise node's scheduler contract; the shared Steps slider allows 1, which would have 422'd at enqueue).
• Duration snapping special-cases the 14 s stop to 345 frames, which was previously unreachable (nearest-grid rounding topped out at 328).

Known gaps (deliberate, review-noted)

• An enabled ref image is silently ignored in image output mode (panel stays visible) — follow-up UX decision.
• Dims not on the 32-grid can survive a model switch and fail at enqueue — pre-existing failure class shared with wan/cogview4, not introduced here.
• Duration metadata recall registers but can't fire yet: videos have no per-field recall viewer.

Testing

pnpm lint:tsc / eslint / prettier clean; pnpm test:no-watch 144 files, 1719 tests passed (incl. new v4→v5 migration fixtures); pnpm vite build succeeds.

Manual test gate: Generate tab → MiniMax H3 → 5 s 16:9 t2v with audible audio; ref image → i2v; output mode = Image lands a normal gallery image (also from canvas); metadata recall of duration/output mode.

🤖 Generated with Claude Code