ci(e2e): remove --llm-api-key and Databricks credential setup

[TomeHirata]

Summary

All e2e tests now use the in-process mock LLM server — no real LLM credentials needed for the standard PR gate. Tests that require real credentials (prompt policy classifier) skip cleanly.

Changes

• Remove --llm-api-key, --profile default, --harness databricks from pytest invocation
• Remove "Set LLM credentials" and "Write gateway profile (~/.databrickscfg)" steps
• Remove OMNIGENT_TEST_MODEL_SPREAD / OMNIGENT_TEST_MODEL_POOL_GPT env vars (only needed for load-balancing real gateway calls)
• Update workflow description to reflect mock LLM usage

Test plan

• CI passes without credentials (mock LLM handles all tests)
• 2 prompt-policy tests show as SKIPPED (expected — they check DATABRICKS_TOKEN)

This pull request and its description were written by Isaac.

[omnigent-ci]

Polly AI Review

PR Review: ci(e2e): remove --llm-api-key and Databricks credential setup

---

1. Blocking Issues

None.

The removed pytest flags (--llm-api-key, --profile default, --harness databricks) and env vars (OMNIGENT_TEST_MODEL_SPREAD, OMNIGENT_TEST_MODEL_POOL_GPT) are cleanly excised with no dangling references in the remaining workflow. The heredoc credential write and $GITHUB_ENV exports are fully removed.

---

2. Security Analysis

This diff is a net security improvement:

• The old Set LLM credentials step injected LLM_API_KEY into $GITHUB_ENV, making the secret visible to every subsequent step in the job — a well-known GitHub Actions secret-leakage vector. Removing it is strictly correct.
• The Write gateway profile step wrote the token in plaintext to ~/.databrickscfg on the runner. Removing it eliminates a credential artifact that could be exfiltrated by a compromised action or a log-scraping attack on the runner's filesystem.
• DATABRICKS_BEARER=$LLM_API_KEY passthrough is gone — no longer a lateral-movement path if a sub-process were compromised.

No new attack surface is introduced. The mock-LLM path carries no credentials by definition.

---

3. Non-Blocking Suggestions

a. Scheduled run now uses mock LLM too.
The workflow fires on schedule (09:00 UTC daily). The credential setup is removed unconditionally, so nightly runs also exercise only the mock path. If nightly.yml already owns real-credential integration coverage this is fine, but it is worth a one-line comment confirming that intent, so a future reader doesn't add credentials back "to fix" the nightly.

b. Test plan checkboxes are unchecked.
Both items in the PR description are [ ]. Recommend confirming CI green and the expected 2 SKIPPED prompt-policy tests before merge, and checking these off (or noting them in the merge commit).

c. tests/_model_pools.py may be dead code.
The removed comment explicitly references tests/_model_pools.py as the consumer of OMNIGENT_TEST_MODEL_SPREAD. With that env var gone, the module is presumably unused in CI. Worth a follow-up to remove or document it so it doesn't confuse contributors.

---

4. Summary

A clean, well-scoped CI hygiene change. Removing real credentials from the PR gate is the right call once the mock-LLM path covers the full test surface, and the diff actually improves the security posture by eliminating secret-to-env and secret-to-disk patterns. The only risk worth watching is that the scheduled run has silently lost real-gateway coverage — confirming nightly.yml or another job still exercises the Databricks path would close that question entirely. No correctness or security blockers; safe to merge once the test plan is verified green.

---

_{Automated review by Polly · workflow run}