[M6] Improve AI provider quality, cost, and deployment strategy

[weiuou]

Goal

Use M5 evaluation results to improve quality, cost, and deployment strategy for real image-to-3D generation.

Scope

• Decide whether to continue with Meshy API, move to local TripoSR, evaluate Hunyuan3D/TRELLIS further, or keep multiple providers.
• Plan GPU deployment or API-cost controls only after M5 data is available.
• Add model-specific paperability improvements such as mesh cleanup, simplification policy, prompt/view hints, or mask conditioning.
• Consider caching, quotas, retries, and provider-specific observability.

Acceptance Criteria

• M5 evaluation evidence is summarized before implementation choices are made.
• A provider quality/cost decision record exists.
• Next implementation slices are split by selected provider and paperability bottleneck.