feat(github-copilot(GHE)): route models through advertised responses endpoint

[Proxicon]

Fixes #1706

Summary

GitHub Copilot currently routes every model through /chat/completions, including models whose authenticated /models metadata advertises only /responses.

This change adds provider-local endpoint capability discovery and selects the correct API:

• /responses → ResponsesClient
• /chat/completions only → existing ChatClient
• both → prefer /responses
• neither → fail before inference

No model-name or version allowlists are used.

Scope

Changes are limited to the GitHub Copilot provider and focused tests.

This PR does not change shared model contracts, configuration, daemon startup, generic routing, retry/failover behavior, authentication, token exchange, or other providers.

Validation

Live validation against GitHub Enterprise Copilot passed:

• gpt-5.5
• gpt-5.4-mini
• gpt-5.3-codex
• mai-code-1-flash-picker

For each model:

• GHE token exchange returned HTTP 200
• tenant /models discovery returned HTTP 200
• inference completed successfully
• unsupported_api_for_model was absent
• fallback was not invoked

Existing gpt-5.4 compatibility remains intact.

Build validation

• Full solution build: 30 succeeded, 0 failed
• Linux x64 CLI and daemon publish: passed
• Container build and deployment: passed
• Published daemon SHA256 matched the daemon packaged in the container

Test evidence:

token exchange → HTTP 200
/models → HTTP 200
LLM streaming call completed
no failover

Live GHE validation evidence

Test environment:

• GitHub Enterprise OAuth device authentication
• Tenant-specific Copilot API host returned through endpoints.api
• Fresh Linux x64 self-contained Netclaw daemon from commit 206fad03
• Deployed container image digest: sha256:32c8484f1629fe4d5353cafa8b7d4f72e31ff1cc1d58b8c061196fed0ceca0f3
• Published daemon and image-contained daemon SHA256: 14848BA5753D10AB6010C5D526199370D1BC27D2052DD3724A2EA0E93426323B

Before the fix

Main model	Advertised endpoint	Previous result
gpt-5.5	/responses	HTTP 400 unsupported_api_for_model; incorrectly sent to /chat/completions; fallback invoked
gpt-5.4-mini	/responses	HTTP 400 unsupported_api_for_model; incorrectly sent to /chat/completions
gpt-5.3-codex	/responses	HTTP 400 unsupported_api_for_model; incorrectly sent to /chat/completions
mai-code-1-flash-picker	Responses-capable model returned by authenticated discovery	HTTP 400 unsupported_api_for_model; incorrectly sent to /chat/completions; fallback invoked

Representative failure:

model "gpt-5.5" is not accessible via the /chat/completions endpoint

After the fix

Main model	Token exchange	Authenticated /models	Inference	unsupported_api_for_model	Fallback invoked	Result
gpt-5.5	HTTP 200	HTTP 200	Completed; input 10,543 / output 19	No	No	PASS
gpt-5.4-mini	HTTP 200	HTTP 200	Completed; input 10,551 / output 86	No	No	PASS
gpt-5.3-codex	HTTP 200	HTTP 200	Completed; input 10,546 / output 23	No	No	PASS
mai-code-1-flash-picker	HTTP 200	HTTP 200	Completed; input 10,553 / output 104	No	No	PASS

Sanitized log evidence

gpt-5.5:

Copilot token exchange: HTTP 200
GET https://copilot-api.<tenant>.ghe.com/models: HTTP 200
LLM streaming call completed in 2127ms (input: 10543, output: 19)
No unsupported_api_for_model error
No "failing over to next candidate" event

gpt-5.4-mini:

Copilot token exchange: HTTP 200
GET https://copilot-api.<tenant>.ghe.com/models: HTTP 200
LLM streaming call completed in 2285ms (input: 10551, output: 86)
No unsupported_api_for_model error
No "failing over to next candidate" event

gpt-5.3-codex:

Copilot token exchange: HTTP 200
GET https://copilot-api.<tenant>.ghe.com/models: HTTP 200
LLM streaming call completed in 3035ms (input: 10546, output: 23)
No unsupported_api_for_model error
No "failing over to next candidate" event

mai-code-1-flash-picker:

Copilot token exchange: HTTP 200
GET https://copilot-api.<tenant>.ghe.com/models: HTTP 200
LLM streaming call completed in 2884ms (input: 10553, output: 104)
No unsupported_api_for_model error
No "failing over to next candidate" event

Result

The live GHE tests demonstrate that endpoint-capability discovery is active and that previously failing Responses-only models now complete inference without falling back to the known /chat/completions-compatible GPT-5.4 model.

[Aaronontheweb]

LGTM