OpenAI conformance: error-payload classes and sampling-field coverage

[Zigfreidish]

Child of #1384.

Consolidate Conformance Matrix rows for all four error classes (unsupported field, invalid schema, backend unavailable, context overflow) so each returns a typed payload naming the incompatible field/phase, and confirm seed/frequency_penalty passthrough rows. Hardening of existing behavior, no new runtime feature.

Done when

• Matrix rows assert typed error payloads for the four error classes.
• Matrix rows assert sampling passthrough for seed and frequency_penalty.

[Akane-CN]

Watch note: request bounds and reasoning-continuation parity

• Design principle: policy-forced sampling should only pin token-selection knobs; request length bounds and reasoning-continuation presentation must remain request-scoped and parity-tested across streaming and non-streaming paths.
• Melix-specific goal: keep max_tokens as an output cap even when recommended/forced sampling is active, reject negative reasoning budgets with typed OpenAI-compatible errors, and ensure raw completion streams do not emit synthetic reasoning-open markers that the non-streaming continuation path would hide.
• Measurable target: conformance matrix rows cover forced-sampling + request max_tokens, invalid thinking_budget < 0, and streaming/non-streaming raw completion parity for prompts ending in a reasoning-open token; all rows must pass against mock and real local backends.
• Next executable slice: extend the OpenAI conformance harness with these three rows first, then adjust effective-sampling merge logic and stream chunk normalization only where the rows prove Melix currently diverges.