Configured model output budgets
- Ordinary agent requests now use each model’s resolved
maxTokens from the first attempt, including models.json overrides, instead of silently starting at 32k.
- Length retries retain the same configured output limit rather than escalating through synthetic intermediate budgets.
- Exhausted responses trigger one compact-and-retry recovery only when provider evidence or token arithmetic indicates context pressure; genuine output-limit exhaustion still fails loudly.
- Numeric thinking budgets now fit within the configured total output maximum while preserving visible-output headroom.
- Provider payload handling, Codex override validation, retry events, TUI/dashboard messaging, tests, and documentation were updated consistently.
Implemented in PR 521, resolving issue 519.
