Freeze high-quality model configuration

[Mister-JP]

What changed

• Centralized and froze one high-quality generation profile for every live model.
• Increased research, visual curation, repair, and redraw reasoning and token budgets.
• Synchronized public preset metadata with runtime limits.
• Increased spend headroom so expensive models can enter the shared standard profile.
• Released cost holds for rejected provider requests and stopped retrying definitive model-access failures.
• Preserved lead-image previews when retrying and included the current celestial animation refinements.

Why

Production behavior diverged by model even though prompts were shared. Model-priced admission holds, retries on definitive provider failures, and stale catalog limits made Luna and other expensive models behave differently from Nano.

Impact

Every selectable model now receives the same request configuration except for the model identifier, with higher reasoning and output headroom. Rejected model-access requests no longer consume repeated uncertain holds.

Validation

• Production build passed
• Full test suite passed
• Focused provider and research tests passed
• ESLint completed with zero errors
• Git diff check passed