feat(minimax-h3): dropdowns for canvas size and frame count

[lstein]

Summary

Two MiniMax H3 node-parameter changes, both turning constrained numeric fields into dropdowns so an invalid value cannot be typed in the first place.

1. Ideal Dimensions: resolution presets

A target_resolution popup on minimax_h3_ideal_dimensions (1.0.0 → 1.1.0):

• 768 highres (default) — the released canvas policy, unchanged: short edge 768, soft area cap 768x1344.
• 768 lowres — pins the long edge to 768 instead, so the canvas carries roughly half the pixels for non-square ratios (a 2:3 keyframe becomes 512x768 rather than 768x1152; 16:9 becomes 768x448). Useful for fast preview renders; it sits below the model's training resolution, so expect some fidelity cost.

Both presets share the 1:4–4:1 aspect validation (extracted into _validated_aspect_ratio; the highres path is line-for-line unchanged) and round each axis to the 32-pixel grid every H3 node requires.

2. Denoise: frame count as a grid-aligned dropdown

num_frames on minimax_h3_denoise (1.2.0 → 1.3.0) was a free integer that had to land on the video VAE's 17n+5 grid; a mistyped value only surfaced as a validation error once the graph was already running. It is now a choice list of every legal count from 90 (3.75 s) through 345 (14.38 s), defaulting to 124, plus the single 5-frame block the still-image output mode uses. Labels show the duration ("124 frames - 5.17 s").

The accepted floor for video moves from 124 frames (the released 5 s training window) to 90, so short test renders are possible; validate_num_frames enforces the same range for hand-authored graphs. Choice values are strings — the only enum shape both workflow editors render as a dropdown — and the node converts to an int at invoke time. An import-time assert keeps the Literal in sync with the computed grid.

Compatibility. A workflow saved before 1.3.0 stores num_frames as a number, which the new enum template rejects. A node migration (alongside the existing migrateImageCollectionInputValues) converts a stored count that is still on the grid and falls back to the default otherwise; it is wired into both paths that rebuild a node from its template — updateNode for saved workflows and graphToWorkflow for workflows recalled from an older session's graph metadata. Deliberate tradeoff: an integer edge into num_frames (e.g. an iterate sweeping clip lengths) is no longer type-compatible — that is inherent to making the field a choice list.

Review

Both commits went through fresh-context adversarial review.

The presets diff came back clean: ~340k integer aspect pairs brute-forced against five invariants (32-grid, long edge pinned at exactly 768, area within the cap and never above the highres result, no orientation inversion) with zero violations; the highres refactor verified byte-identical including NaN/inf and boundary-ratio exception paths.

The dropdown review found one real defect and several nits, all fixed in the follow-up commit:

• A pre-1.3.0 saved workflow rendered the frame-count row with no widget at all — silently, because the editor's recursive merge kept the stored number and InputFieldRenderer returns null when the instance fails the enum schema, leaving the user no way to fix it before the enqueue 422. Fixed by the migration described above (the codebase's documented rule is that a merge which would produce an invalid node needs either a major bump or a migration hook).
• Both bundled workflows still carried integer-field-config form settings for the field (inert, but stale) — dropped.
• The canvas workflow panel rendered enum options as raw values, ignoring ui_choice_labels — now shows the labels (generic fix, newly visible for this node).
• The 5-frame label was inconsistent with the others — now "5 frames - still image only".

Verified-and-held attacks worth recording: the validator's accept/reject verdict changed for exactly two inputs over −50…500 (90 and 107, both intended); every value the linear UI's duration slider can produce is in the enum; pydantic does not coerce 124 → "124" for a str Literal even in lax mode (so nothing silently passes); and both generated artifacts are byte-identical to what CI's recipes produce.

Testing

• Backend: full suite 4304 passed. New: per-choice acceptance and grid-membership tests, label formatting, lowres canvas math (exact values, the never-larger-than-highres property, envelope extremes), and default-preserves-old-behavior.
• Frontend: pnpm test:no-watch 1728 passed (incl. three new node-migration tests), lint:tsc and eslint clean.
• ruff@0.11.2 format --check and ruff check clean; schema.ts and openapi.json regenerated per the CI recipes.

Stacked on #77. Note that #77 also picked up three CI-hygiene fixes that were failing the whole stack: a prettier pass on openapi.json, ruff-0.11.2 formatting of three vendored files, and an unused export that failed knip. Separately, the python-tests failures across #62–#77 were traced to a test that allocated ~24 GB (real int8 zeros at the 32B text encoder's shapes) and OOM-killed the runner; that fix landed at minimax-h3/07-quantized-text-encoder and was rebased through the stack.

🤖 Generated with Claude Code