feat: intent-driven dual-model architecture (text2music + lego)

[ChuxiJ]

Summary

• Add Text2MusicTaskParams, GenerationIntent, ModelCategory types and 'mix' track type
• Extend modelStore with ensureModelForIntent() — auto-selects the right model family (text2music vs lego) and ensures LM is loaded when needed
• Add generateText2Music() pipeline: creates mix track/clip → submits text2music task → polls → downloads audio → optionally auto-splits into stems
• Add IntentSelector component: Full Song / Single Track toggle with model status indicator
• Add FullSongForm: dedicated form for text2music (prompt, lyrics, BPM/key/duration, split-to-stems, advanced params)

Backend requirements

For end-to-end operation, the backend needs:

1. category field on ModelEntry in /v1/model_inventory ('text2music' or 'lego')
2. text2music task_type accepted by /release_task

Test plan

• 36 new unit tests pass (inferModelCategory, intentToCategory, intentNeedsLm, modelStore category getters, ensureModelForIntent)
• All 2679 existing tests pass (0 regressions)
• npx tsc --noEmit — 0 type errors
• npm run build — succeeds
• Visual verification: IntentSelector toggles correctly, FullSongForm renders all fields, no console errors

Closes #884

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR introduces an intent-driven generation flow that distinguishes between text2music (full-song) and lego (single-track) model families, adds a full-song text2music generation pipeline, and updates the generation UI to let users choose intent while the app auto-selects/initializes the required model(s).

Changes:

• Added new API/types for intent-driven generation (GenerationIntent, ModelCategory, Text2MusicTaskParams) and introduced the 'mix' track type.
• Extended modelStore with category-aware helpers and ensureModelForIntent() to auto-select the correct model family and (attempt to) ensure LM readiness.
• Added generateText2Music() pipeline and new UI components (IntentSelector, FullSongForm) integrated into GenerationSidePanel.

Reviewed changes

Copilot reviewed 13 out of 13 changed files in this pull request and generated 9 comments.

Show a summary per file

File	Description
tests/unit/generationSidePanel.test.tsx	Updates assertions to match new generation panel tab markup and default intent behavior.
src/types/project.ts	Adds 'mix' to TrackType for full-song output track typing.
src/types/api.ts	Introduces Text2MusicTaskParams, GenerationIntent, and ModelCategory plus ModelEntry.category.
src/store/modelStore.ts	Adds intent/category model selection (ensureModelForIntent, category getters, overrides).
src/store/tests/modelStoreIntents.test.ts	Adds unit tests for intent mapping and category-aware model resolution behavior.
src/services/generationPipeline.ts	Adds generateText2Music() full-song pipeline including optional stem splitting.
src/services/aceStepApi.ts	Adds inferModelCategory() and expands task submission types to include text2music.
src/services/tests/inferModelCategory.test.ts	Adds tests for category inference behavior and precedence rules.
src/constants/tracks.ts	Adds catalog entry for new 'mix' track type.
src/constants/trackHeight.ts	Adds default lane height for 'mix' in auto preset.
src/components/generation/IntentSelector.tsx	New intent toggle UI with model status indicator.
src/components/generation/GenerationSidePanel.tsx	Integrates intent selection and routes to FullSongForm for full-song flow.
src/components/generation/FullSongForm.tsx	Adds a dedicated full-song (text2music) form and trigger to run the new pipeline.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Advanced numeric inputs (inferenceSteps/guidanceScale/shift) use Number(e.target.value) and can also become NaN if cleared. Consider guarding against NaN and ensuring only finite values are sent (or restoring the previous valid value on blur).

[Copilot]

The comment says cover/repaint should "just check LM", but this branch returns immediately and does not perform any LM-related checks. Either update the comment to match the behavior, or implement the intended LM check here so future intents don’t silently skip required initialization.

Suggested change

	// Cover/repaint work with either model — no switching needed
	// Cover/repaint work with either model — no switching needed
	// But we still need to ensure LM is initialized if required.
	if (needsLm) {
	const refreshedState = get();
	const lmLoaded = refreshedState.availableLmModels.some((m) => m.is_loaded);
	if (!lmLoaded && refreshedState.availableLmModels.length > 0) {
	const defaultLm = refreshedState.activeLmModelId ?? refreshedState.availableLmModels[0].name;
	await refreshedState.switchLmModel(defaultLm);
	}
	}

[Copilot]

This submits a full-length silence WAV even though the comment notes text2music doesn't need source audio. For longer durations this can add significant upload/downsample time and risk request-size limits. If the backend supports it, consider skipping the src_audio upload entirely for task_type='text2music' (e.g., adjust releaseTask/releaseLegoTask to omit the blob for text2music).

[Copilot]

This wrapper now accepts Text2MusicTaskParams in addition to lego/cover/repaint, so the name releaseLegoTask is no longer accurate and can be confusing for future callers. Consider renaming to something task-agnostic (or introducing a new wrapper) to better reflect its responsibilities.

[Copilot]

Using Number(e.target.value) on a will produce NaN when the user clears the field, and that NaN will be sent to generateText2Music() (and ultimately the backend). Consider treating "" as null/undefined and validating/clamping before enabling submission.

[Copilot]

ensureModelForIntent() awaits switchModel(), but switchModel() swallows init errors (it only sets modelLoadingState='error'). That means this function can continue and run generation with the wrong model loaded. Consider making switchModel/switchLmModel rethrow, or have ensureModelForIntent verify after refresh that an expected model is loaded and throw if not.

[Copilot]

When needsLm is true (full-song), this silently does nothing if availableLmModels is empty, which will likely lead to a backend failure later. It also relies on switchLmModel(), which currently swallows init errors. Consider throwing a clear error when no LM models are available, and/or validating that an LM is actually loaded after attempting init.

Suggested change

	const lmLoaded = refreshedState.availableLmModels.some((m) => m.is_loaded);
	if (!lmLoaded && refreshedState.availableLmModels.length > 0) {
	const defaultLm = refreshedState.activeLmModelId ?? refreshedState.availableLmModels[0].name;
	await refreshedState.switchLmModel(defaultLm);
	if (refreshedState.availableLmModels.length === 0) {
	throw new Error('No LM models available. Check backend LM model inventory.');
	}
	const lmLoaded = refreshedState.availableLmModels.some((m) => m.is_loaded);
	if (!lmLoaded) {
	const defaultLm =
	refreshedState.activeLmModelId ?? refreshedState.availableLmModels[0].name;
	await refreshedState.switchLmModel(defaultLm);
	// Re-validate that an LM is actually loaded after attempting init
	const postSwitchState = get();
	const lmLoadedAfterSwitch = postSwitchState.availableLmModels.some((m) => m.is_loaded);
	if (!lmLoadedAfterSwitch) {
	throw new Error(
	`Failed to initialize LM model "${defaultLm}". Please verify LM backend configuration.`,
	);
	}

[Copilot]

generateText2Music() calls api.releaseLegoTask() with Text2MusicTaskParams. Since this API wrapper is now used for multiple task types, the name (and related log/error strings inside aceStepApi) is misleading. Consider adding a task-agnostic wrapper (e.g. releaseTask/releaseGenerationTask) or renaming releaseLegoTask to reflect broader usage.

[Copilot]

Same NaN risk applies to durationSeconds: clearing the input yields NaN, but the UI still renders ${durationSeconds}s and the request will submit NaN as audio_duration. Consider parsing safely ("" → default/null) and preventing submission unless duration is a finite number within [MIN_DURATION, MAX_DURATION].

Suggested change

	value={durationSeconds}
	onChange={(e) => setDurationSeconds(Number(e.target.value))}
	min={MIN_DURATION}
	max={MAX_DURATION}
	step={1}
	className="mt-1 w-full rounded border border-[#444] bg-[#2a2a2a] px-2 py-1 text-sm focus:border-indigo-500 focus:outline-none"
	disabled={isDisabled}
	/>
	<span className="mt-0.5 block text-[10px] text-zinc-600">{durationSeconds}s</span>
	value={Number.isFinite(durationSeconds) ? durationSeconds : ''}
	onChange={(e) => {
	const rawValue = e.target.value;
	if (rawValue === '') {
	// Treat empty input as reset to a safe default duration
	setDurationSeconds(MIN_DURATION);
	return;
	}
	const parsed = Number(rawValue);
	if (Number.isNaN(parsed)) {
	// Do not update state with an invalid number
	return;
	}
	const clamped = Math.min(Math.max(parsed, MIN_DURATION), MAX_DURATION);
	setDurationSeconds(clamped);
	}}
	min={MIN_DURATION}
	max={MAX_DURATION}
	step={1}
	className="mt-1 w-full rounded border border-[#444] bg-[#2a2a2a] px-2 py-1 text-sm focus:border-indigo-500 focus:outline-none"
	disabled={isDisabled}
	/>
	<span className="mt-0.5 block text-[10px] text-zinc-600">
	{Number.isFinite(durationSeconds) ? `${durationSeconds}s` : ''}
	</span>