fix(copilot): use target_model not config default for api_mode resolution

[CharlesMcquade]

Problem

When selecting GPT-5.5 via Copilot as the primary model from the WebUI dropdown, every API call fails with:

HTTP 400: model gpt-5.5 is not accessible via the /chat/completions endpoint

GPT-5.x models on Copilot require the /responses endpoint (codex_responses API mode), but the request is sent to /chat/completions instead.

Root Cause

_copilot_runtime_api_mode() in runtime_provider.py reads model_cfg.get("default") — the persisted config.yaml default model (e.g. deepseek-ai/DeepSeek-V4-Flash) — instead of the actual target model (gpt-5.5) when deciding which API endpoint to use.

The WebUI correctly passes target_model to resolve_runtime_provider(), but _copilot_runtime_api_mode() does not accept or use it. Since the config default is not a GPT-5 model, _should_use_copilot_responses_api() returns False, and the API mode resolves to chat_completions.

The fallback path (try_activate_fallback in chat_completion_helpers.py) is unaffected because it uses _provider_model_requires_responses_api() with the correct model name — which is why GPT-5.5 works as a fallback but fails as a primary model selected from the WebUI.

Fix

Pass target_model through _copilot_runtime_api_mode() at both call sites (pool entry path at line 446 and non-pool path at line 2003) and prefer it over model_cfg.get("default") when determining the API mode.

Reproduction

1. Set config.yaml default model to a non-GPT-5 model (e.g. deepseek-ai/DeepSeek-V4-Flash on custom:wandb)
2. In the WebUI, select @copilot:gpt-5.5 from the model dropdown
3. Send a message — HTTP 400 from Copilot
4. Falls back through the chain (claude-sonnet-5 on copilot also fails if not in catalog, then to local MLX which may also fail)

Testing

• 218 copilot-related tests pass (1 pre-existing failure in test_copilot_acp_client unrelated to this change — HERMES_REAL_HOME env var issue)
• Simulated the fix: with target_model="gpt-5.5" and config default deepseek-ai/DeepSeek-V4-Flash, the API mode now correctly resolves to codex_responses

Note: This was authored using the GitHub Copilot commit email per the AGENTS.md fork conventions.

[alt-glitch]

This was generated by AI during triage.

Duplicate of #60294 -- both fix _copilot_runtime_api_mode in hermes_cli/runtime_provider.py by threading target_model through and preferring it over the stale model_cfg["default"] (so GPT-5.x on Copilot routes to /responses instead of 400-ing on /chat/completions). Same function, same mechanism, same file scope; #60294 is the earlier open PR (created ~5h before this). Related: #60086 (models.py/slot-level) and #32308 (broader superset). Tracking the fix in #60294.

[CharlesMcquade]

Closing as duplicate of #60294