Make ACP model-selection integration test resilient to removed models #1439

Problem

The coder-acp-copilot integration test hard-codes TEST_MODEL=claude-opus-4.6. The current ACP session still advertises model-selection capability, but that model is no longer present in models.availableModels.

selectModel() logs that the requested model is unavailable, attempts session/set_model, catches the rejection, and returns undefined. The test then assumes undefined means model-selection capability was not advertised and requires this different warning:

does not advertise model selection capability

The actual warning is:

session/set_model failed for "claude-opus-4.6"

As a result, the coding prompt succeeds but the model-selection assertion fails.

Expected behavior

The integration test should remain meaningful without depending on a server-controlled model identifier that may be removed independently of the pinned CLI version.

Prefer selecting a non-default model from the session's advertised available-model list and asserting that selection succeeds. A short-term constant update may unblock the test, but dynamic selection is the durable fix.

Related work

• fix(coder-acp-copilot): honor --model flag in ACP mode #699 added ACP model selection and the integration coverage.
• fix(coder-acp-copilot): make model selection test resilient to server-side changes #905 handled the earlier case where the server stopped advertising model-selection capability entirely. The current failure is different: capability is advertised, but the requested model has been removed.
• chore: extract shared ACP integration test utils into packages/test-utils (#650) #1437 currently retains the hard-coded claude-opus-4.6 value while refactoring the shared ACP test utilities, so this fix should be coordinated with or applied after that change.