fix(agent): surface HTTP errors from the unload_models tool

[kevin9327]

What

The unload_models chat tool discarded the response from POST /model/unload-all and unconditionally returned the success string, so a failed unload (any 4xx/5xx) still told the assistant — and the user — that VRAM had been freed.

Why it triggers

execute_tool (api/routers/agent.py) wraps every tool call in a shared except httpx.HTTPStatusError handler that turns an HTTP error status into "API error <status>: ...". But httpx does not raise on an error status unless you call raise_for_status(). Every other POST-based tool — list_models, decimate_mesh, smooth_mesh, get_generation_status — captures the response and calls raise_for_status(), and the MCP server's modly_unload_models does too. unload_models was the only one that fired the request and threw the response away, so a 500 from the unload endpoint was reported to the model as a success.

elif name == "unload_models":
    await client.post(f"{MODLY_API}/model/unload-all")   # response ignored
    return "All 3D generation models have been unloaded from VRAM.", None

Fix

Capture the response and call raise_for_status(), matching the sibling tools and the MCP server. The existing HTTPStatusError handler then surfaces the real failure.

Verification

• Added api/tests/test_agent_router.py (unittest + httpx.MockTransport). test_http_error_is_surfaced fails before this change (it gets the success string back for a mocked 500) and passes after; test_success_still_reports_unloaded guards the happy path.
• python -m unittest discover -s tests in api/ (venv with fastapi + python-multipart + httpx): all pass, 0 failures.