fix(qwen): make thinking mode explicit

[jtstothard]

Summary

Qwen-compatible endpoints can default to server-side auto reasoning. For short utility or coding requests, hidden reasoning can consume the entire completion budget and leave empty visible output.

This PR makes Qwen thinking explicit at the provider boundary:

• absent, malformed, false, or non-boolean reasoning config → enable_thinking: false
• explicit {"enabled": true} → enable_thinking: true plus a bounded thinking_budget: 32768
• existing Qwen session metadata remains top-level metadata

Why this is provider-specific

Qwen uses enable_thinking in the request body. This is separate from the custom-provider / llama.cpp/vLLM work in #12427 and #64153, which uses chat_template_kwargs.enable_thinking for a different provider path. This PR does not attempt to redesign generic extra_body configuration or model-specific field mapping.

Reasoning contract

Hermes' reasoning parser produces {"enabled": true, "effort": ...} for explicit reasoning levels and {"enabled": false} for none. Only the genuine boolean enabled=True is authoritative here; effort is intentionally non-authoritative so explicit enablement is not accidentally suppressed by effort="none" or malformed effort values.

Unset reasoning is deliberately sent as enable_thinking=false rather than relying on a server-side auto default. This prevents hidden reasoning from silently consuming utility-request output budgets. The enabled 32K thinking budget is below the profile's 64K completion default, preserving visible-output headroom.

Validation

• python3 -m pytest tests/providers/test_provider_profiles.py -q — 16 passed
• python3 -m pytest tests/providers/test_profile_wiring.py tests/providers/test_transport_parity.py -q — 22 passed
• Independent adversarial review — APPROVE
• Review covered malformed/absent/false/true/effort edge cases, actual ChatCompletionsTransport forwarding, metadata preservation, diff scope, and secret safety

Validation exercised the Qwen-compatible llama.cpp/Lemonade transport path. No live Qwen Portal credential or production service is required or included.

Related upstream work

• fix(agent): also send chat_template_kwargs.enable_thinking=false for llama.cpp/vLLM #12427 / fix: emit chat_template_kwargs for llama.cpp/vLLM thinking control #64153 — custom-provider llama.cpp/vLLM thinking opt-out
• Custom provider thinking toggle doesn't map to model-specific field names #63195 — configurable model-specific thinking field names
• feat(config): support model.extra_body passthrough for CLI and gateway agent requests #7209 — general model extra_body passthrough
• Feature request: Configurable extra_body per provider (enable_thinking=false for DashScope) #8160 — DashScope thinking budget issue

Scope

Only the Qwen provider profile and its focused tests are changed. No routing, Hindsight, LiteLLM, Lemonade, credentials, or deployment changes are included.

[teknium1]

Thanks for the focused provider-boundary fix. Current main still has the relevant gap: QwenProfile.build_api_kwargs_extras() returns {} at plugins/model-providers/qwen-oauth/__init__.py:85-96, while the active transport forwards profile additions into extra_body at agent/transports/chat_completions.py:580-628.

Problems

• The new tests call the profile hook directly. They do not verify the outgoing transport payload; the existing Qwen transport coverage only asserts image-resolution and metadata behavior in tests/providers/test_transport_parity.py:124-148.

Suggested changes

• Add transport-level assertions for disabled and enabled Qwen reasoning. Verify enable_thinking, conditional thinking_budget, vl_high_resolution_images, and top-level metadata together in ChatCompletionsTransport.build_kwargs().

Automated hermes-sweeper review.