Probe cross-format provider behavior for reasoning, tools, and streaming

[Oaklight]

Context

Follow-up from #268 and part of the provider reasoning documentation effort in #267.

We need active probing across providers and API formats to decide shim policies and cross-format conversion behavior instead of relying on assumptions. The immediate trigger was Argo Anthropic rejecting prior thinking blocks with missing/empty/invalid signatures, which led to unsigned_reasoning_blocks: preserve for argo--anthropic.

Goal

Build and run a cross-format probing matrix covering:

• source API format → target shim/API format
• stream and non-stream paths
• reasoning/thinking blocks with:
  • no signature
  • empty signature
  • valid provider-generated signature where available
• tool calls and tool results
• provider metadata round-trip fields:
  • Anthropic thinking.signature
  • Google thoughtSignature
  • OpenAI-compatible reasoning_details / encrypted_content

Initial provider/shim scope

• argo--anthropic
• argo--openai_chat
• openrouter
• google
• minimax--openai_chat
• minimax--anthropic
• volcengine--openai_chat
• volcengine--openai_responses
• deepseek

Outputs

• Probe script(s) or reproducible examples under references/provider_model_params/raw_probes/ or another agreed location
• Matrix of observed behavior per provider/API format
• Shim config updates based on observed behavior
• Follow-up docs work in docs: Provider reasoning capability probing matrix and documentation #267 for provider reasoning compatibility

Related

• docs: Provider reasoning capability probing matrix and documentation #267 — provider reasoning probing/docs arching issue
• feat: configurable handling for unsigned reasoning blocks across providers #268 — unsigned reasoning block shim policy
• Normalize reasoning/thinking fields across OpenAI Chat-compatible providers via shim transforms #185 — reasoning/thinking normalization and metadata preservation
• Comprehensive per-model ARGO parameter probe for shim capability tables #220 — Argo model parameter probing
• PR Add shim policy for unsigned reasoning blocks #269 — initial unsigned reasoning policy implementation