Model defaults to only FP4, when only 1 curve

[Oseltamivir]

@claude fix

[Oseltamivir]

@claude fix

[github-actions]

Claude finished @Oseltamivir's task in 19m 23s —— View job • claude/issue-470-20260618-0436 • Create PR ➔

---

Fixing MiniMax M3 default precision

Reproduced on the live site: selecting MiniMax M3 from the model dropdown left precisions at the global [FP4] default. M3's day-zero rollout shipped 23 FP8 datapoints per sequence vs 1 FP4 row, so the chart was almost empty until users manually checked FP8.

Fix

• Added a per-model defaultPrecisions field on ModelConfig (packages/app/src/lib/data-mappings.ts) and seeded Model.MiniMax_M3 with ['fp4', 'fp8']. Other models inherit the existing [fp4] fallback.
• Wrapped the public setSelectedModel in GlobalFilterContext to re-seed precisions to the new model's defaults. URL i_prec=… and InferenceContext.applyPreset write to precision state immediately after the model setter, so they still win the React batch — URL hydration, preset application, and the unofficial-run auto-switch all behave as before.
• Updated initial selectedPrecisions state to honor the initial model's defaults.

Verification (Playwright, local dev server)

Path	Model	Precision
/inference → switch to MiniMax M3	MiniMax M3 428B	FP4 + FP8
Switch back to DSv4	DeepSeek V4 Pro 1.6T	FP4 (default)
/inference?g_model=MiniMax-M3&i_prec=fp8	MiniMax M3 428B	FP8 (URL wins)
/inference?preset=minimax-m3-launch	MiniMax M3 428B	FP4 + FP8

Chart renders 647 SVG elements, no "No data available" message.

• Find where MiniMax M3 default precision is configured
• Add defaultPrecisions to ModelConfig; set MiniMax M3 = ['fp4', 'fp8']
• Reset precisions on model change in GlobalFilterContext (URL i_prec still overrides)
• Add unit tests (getModelDefaultPrecisions)
• Typecheck ✅, lint ✅, fmt ✅, 2107 unit tests pass ✅
• Commit and push (cbb849e)

· branch: claude/issue-470-20260618-0436