Add model discovery integration tests without managed config

[andy-xu-db]

Summary

• Add Tests-table Cases 13–24, where managed config does not exist, stacked on Add model discovery integration tests with managed config #655.
• Cover 12 named tests and 20 configured/fresh executions across Claude Code and Codex.
• Add only the remaining unmanaged discovery inputs, fixtures, and Claude-owned cache inspection support.
• Keep hard observable assertions for provider- and location-scoped discovery; there are no skips, xfails, retries, or fallback expectations.

Cases

Cases	Scope	Executions
13–16	saved location and discovery-disabled native fallback	4
17–24	provider/location discovery and disabled fallback, configured + fresh	16

Each test function includes its Tests-table case number (test_case_13_... through test_case_24_...). Together with #655, the stack has 44 numbered discovery executions.

Managed workspace coverage inherited from #655

This branch is rebased on #655 head 76d42794a3061efe256f32dc3f5a6b5d7a644168. It therefore inherits the one-live-fetch-per-agent managed-config fixtures and their dedicated ca-central-1 MPS targets:

• Claude Code: main.default.ci_e2e_anthropic_mps
• Codex: main.default.ci_e2e_openai_mps

Both managed-config lanes pass on CI run 35284318193.

Live result

Run 35284318193 exercised Claude Code 2.1.268 and Codex 0.154.0 against the integration workspace.

• Managed Cases 1–12 pass for both agents.
• Unmanaged discovery-disabled Cases 15, 16, and 21–24 pass.
• Codex Cases 14, 18, and 20 currently fail their exact scoped-catalog assertions: Codex exposes its native catalog instead of the single saved/provider/location model.
• Claude Cases 13, 17, and 19 currently fail their exact scoped-picker assertions: Claude Code exposes its native picker and, where expected, does not create its gateway cache.

These failures are intentionally reported rather than weakened or skipped; they demonstrate the current client/catalog gaps the new tests cover. A separate Claude agent-launch shard also fails because system.ai.claude-opus-5 rejects Claude Code's advisor_20260301 tool type.

Stack

• Depends on Add model discovery integration tests with managed config #655 (andy/model-discovery-integration-tests), which contains managed-config Cases 1–12.
• Add model discovery integration tests with managed config #655 is based on codex/model-discovery-test-stack, rebased on current main.
• No dependency on Refresh Claude model cache before scoped launches #622 and no pre-launch Claude cache refresh.

Verification

• Head: 6fb6e7f9fb4ef6b25d15e771dee8d24eb806e6a6.
• Unit-test rerun 105415208862 passed; the original first-prompt socket failure did not recur.
• tests/test_integration_contract.py: 5 passed.
• This PR adds 20 numbered executions; cumulative discovery collection is 44; complete integration collection is 99.
• Ruff, formatting, and diff checks passed.
• No local Databricks profile was selected.