fix(gateway): scope api_server secrets under multiplex profiles

[giggling-ginger]

Summary

• When gateway.multiplex_profiles is on, credential reads go through get_secret and fail closed unless a profile secret scope is installed.
• Messaging platforms already wrap turns in _profile_runtime_scope; api_server (/v1/chat/completions SSE and /v1/runs) did not, so agent setup raised UnscopedSecretError on OPENROUTER_BASE_URL (and any other scoped credential).
• Install the same profile secret scope on api_server agent entry (create + run), inside the executor thread where ContextVars actually apply. No-op when multiplexing is off.

Fixes #61276

Test plan

• scripts/run_tests.sh tests/gateway/test_api_server_multiplex_secret_scope.py
• Existing test_run_agent_binds_api_session_context_for_tool_env still passes
• Optional live check: gateway.multiplex_profiles: true + api_server streaming chat completion with OpenRouter

[alt-glitch]

This was generated by AI during triage.

Fix PR for #61276. Distinct call site from the same multiplex secret-scope family — #52399 (API_SERVER_* enable/key in config.py), #52307 (_apply_env_overrides) — this installs the profile secret scope on the api_server agent entry. Related, not duplicate.

[teknium1]

Thanks for tracing the API-server executor boundary. The premise is confirmed on current main: gateway/platforms/api_server.py:4052-4096 runs _create_agent() and run_conversation() in a bare executor, while multiplex mode makes unscoped get_secret() calls fail closed (agent/secret_scope.py:149-157). The PR scopes that path and the independent /v1/runs construction/executor path.

Suggested changes

• Add a /v1/runs regression covering scope visibility during both agent construction and run_conversation; the new test currently covers only _run_agent, whereas /v1/runs has separate regions at proposed gateway/platforms/api_server.py:4330 and :4379.

Automated hermes-sweeper review.

[teknium1]

Please add a /v1/runs regression that verifies a scoped secret is available during agent construction and during the executor-run conversation. The new test currently covers _run_agent only; this path has an independent construction/executor split.

[giggling-ginger]

Addressed the review feedback in commit 1952ff37c by adding a /v1/runs regression test that verifies profile secret-scope visibility during both agent construction and run_conversation in the executor thread.

[teknium1]

Merged via PR #65700 — rebuilt on current main with your authorship preserved in git log (rebase merge, commit credited to you). Your diagnosis of #61276 was correct and your PR was ahead of the codebase: the /p/<profile>/ routing rewrite (7aa21e3) landed after you opened this and superseded the mechanical diff, but left exactly the gap you identified — plain requests on the default listener still ran unscoped. The rebuild puts your fix at the new seam: _profile_scope(None) enters the default profile's runtime scope when multiplex is active, covering all three wrapped call sites at once. Your regression tests were ported to the new seam rather than discarded. Closes #61276. Thanks!

Note: #60726 (background-task UnscopedSecretError) is the same bug class at a different call site (_run_background_task in gateway/run.py) and still needs its own fix using the same pattern.