fix(bedrock): stop sending output_config.format on AnthropicBedrock aux routes; probe with maxTokens=16

[dskwe]

Summary

Two Bedrock fixes from #124923:

1. output_config.format sent to AnthropicBedrock (400 on every structured-output aux call). _AnthropicCompletionsAdapter.create translated response_format (top-level and extra_body) into Anthropic output_config.format even when the wrapped client is AnthropicBedrock. Bedrock's InvokeModel endpoint rejects that field — 400 {'message': 'output_config.format: Extra inputs are not permitted'} (ValidationException on InvokeModelWithResponseStream for global.anthropic.claude-opus-5-5, claude-opus-4-8, claude-sonnet-5 — endpoint, not model). The _REJECTED_ROUTES memo is per-process, so every new gateway/CLI process re-learns it by eating a 400 on each new session's title call. The adapter now detects AnthropicBedrock clients (class-name check — the anthropic SDK is an optional extra) and skips the translation up front; schema enforcement degrades to prompt compliance, the same fallback used for providers that don't support a format at all.

   I did not take the issue's alternative of unsupported_response_formats on the bedrock provider profile: that filter runs in _build_call_kwargs before the wire is known, and it would also strip response_format from the Mantle route (Bedrock-hosted OpenAI models, a plain OpenAI client with no evidence of rejection), losing structured output there for no reason. The per-client gate scopes the skip to the endpoint that actually rejects.

2. Context probe maxTokens=8 is below the OpenAI-on-Bedrock minimum. probe_bedrock_context_length sent inferenceConfig={"maxTokens": 8}; OpenAI-on-Bedrock models (openai.gpt-6-sol, openai.gpt-6-luna) reject max output < 16 with a ValidationException before the prompt-length check, so the probe never got the "prompt is too long" error it parses — nothing cached, the two probe tiers (~1.3M/2.2M tokens) re-ran in every new process, and the model resolved to BEDROCK_DEFAULT_CONTEXT_LENGTH (128K), triggering early compression. The probe now sends maxTokens=16.

Root cause

• agent/auxiliary_client.py: _AnthropicCompletionsAdapter.__init__ now records _is_bedrock_client = type(real_client).__name__ == "AnthropicBedrock"; create() routes response_format translation through a gate that logs-and-skips for Bedrock clients (both the top-level and extra_body forms).
• agent/bedrock_adapter.py: probe_bedrock_context_length sends maxTokens=16 with a comment explaining the OpenAI-on-Bedrock minimum.

How I tested it

• New regression tests:
  • tests/agent/test_auxiliary_client.py::TestAnthropicAuxiliaryReasoningTranslation::test_bedrock_client_never_receives_output_config_format — an AnthropicBedrock-named client never gets output_config.format on the wire.
  • ...::test_non_bedrock_client_still_gets_output_config_format — the skip is scoped to Bedrock; plain Messages clients keep the translation.
  • tests/agent/test_bedrock_adapter.py::TestBedrockContextProbe::test_probe_max_tokens_meets_openai_bedrock_minimum — the probe sends maxTokens >= 16 and parses the limit.
• Full local runs (venv pytest): tests/agent/test_bedrock_adapter.py + tests/agent/test_bedrock_context_cache.py → 135 passed; tests/agent/test_auxiliary_client.py + tests/agent/test_model_metadata.py → 331 passed.

Fixes #124923