Problem

PR #14424 adds providers/ package with 8 provider profiles and a transport single-path (_build_kwargs_from_profile). The profiles are proven correct via 73 parity tests. However, wiring run_agent.py to activate profiles caused 50 test regressions because of a fundamental detection mismatch.

The mismatch

Legacy detection: URL pattern matching

_is_nvidia = "integrate.api.nvidia.com" in self._base_url_lower
_is_kimi = base_url_host_matches(self.base_url, "api.kimi.com") or ...
_is_qwen = self._is_qwen_portal()  # checks base_url

Profile resolution: Provider name matching

_profile = get_provider_profile(self.provider)  # "nvidia", "openrouter", etc.

A user with provider="custom" and base_url="https://integrate.api.nvidia.com/v1" hits the NVIDIA path in legacy but gets NO profile. Conversely, tests that set provider="openrouter" without a matching base_url now hit the profile path when they expected legacy.

Missing params in profile path

The legacy chat_completions _build_api_kwargs computes these agent-level params that the profile path doesn't:

1. anthropic_max_output — Claude max output for OpenRouter/Nous (calls _get_anthropic_max_output)
2. qwen_session_metadata — Random UUID per call
3. supports_reasoning — Model-family-gated reasoning safety check
4. github_reasoning_extra — GitHub Models reasoning config
5. _fixed_temperature_for_model — Temperature override from auxiliary_client
6. provider_preferences — OpenRouter provider routing preferences

Proposed approach

Wire profiles one provider at a time, simplest first:

1. NVIDIA (only default_max_tokens, zero missing params)
2. DeepSeek (minimal profile, zero extras)
3. Kimi (OMIT_TEMPERATURE + reasoning_effort — profile handles both)
4. Qwen (needs session_metadata generation in profile)
5. OpenRouter (needs anthropic_max_output + supports_reasoning + provider_preferences)
6. Nous (similar to OpenRouter)

Each wiring step: add the missing params to the profile, write an E2E test that calls _build_api_kwargs with real agent state, verify identical output, then enable.

Acceptance criteria

• Each provider profile produces identical build_kwargs output to the legacy flag path
• Tests that create agents with provider="X" work correctly on the profile path
• provider="custom" with provider-specific base_url still hits legacy path (profile returns None)