fix(tests): isolate credential pool routing tests from real ~/.claude/.credentials.json

[JonthanaHanh]

Summary

TestFailureAttribution._make_pool() calls load_pool("anthropic") which triggers auto-discovery of real Claude Code credentials from ~/.claude/.credentials.json on the developer's machine. This changes the fixture from the intended single-entry pool to a two-entry pool and makes test_unmatched_key_does_not_retry_only_pool_entry fail with assert True is False.

Root Cause

The _make_pool helper only overrides HERMES_HOME via monkeypatch, but load_pool("anthropic") still runs the auto-discovery path at agent/credential_pool.py:2245 which reads ~/.claude/.credentials.json from the real home directory. When a developer has usable Claude Code OAuth credentials, the auto-discovery inserts a second claude_code entry into the pool, turning a single-entry test into a multi-entry test.

Fix

Monkeypatch read_claude_code_credentials and read_hermes_oauth_credentials to return None before calling load_pool(). This matches the isolation pattern already used in tests/agent/test_anthropic_adapter.py (lines 235, 256, 278, 298, 318, 359, 367, 403).

Test Plan

• pytest tests/agent/test_credential_pool_routing.py -q -> 15 passed
• Verified fix isolates from real ~/.claude/.credentials.json

Fixes #75230

[teknium1]

Closing with credit: this exact host-credential leak in the _make_pool fixture was fixed on main via PR #82500 (merged in c28114a). Your PR was submitted first (Jul 31) and correctly identified the load_pool auto-seeding problem — thank you! #82500 was merged as the wider variant since it also blanks the Anthropic env vars, stubs is_provider_explicitly_configured, asserts exact pool contents, and removes a stale sys.path override in a sibling test file. Both contributions acknowledged.