e2e: fix the codex model exclusions, and stop truncating away the cause

[tt-le]

Changes

TestCodexLaunch::test_launch_codex_per_model has been red on main for at least five consecutive runs. The exclusion list had drifted in both directions.

Dropped gpt-5-6-luna / sol / terra. Their comment says the backing OpenAI endpoint returns ENDPOINT_NOT_FOUND in the CI region. That is no longer true — all three answer a /ai-gateway/codex/v1/responses request there now:

databricks-gpt-5-6-luna    OK -> gpt-5.6-luna
databricks-gpt-5-6-sol     OK -> gpt-5.6-sol
databricks-gpt-5-6-terra   OK -> gpt-5.6-terra
databricks-gpt-5-3-codex   INTERNAL_ERROR

So the list was excluding the model ucode actually defaults to (gpt-5.6-sol is the launcher's pick) while still testing one that cannot run.

Added gpt-5-3-codex, which is what fails today. It is discoverable, in the workspace catalog, and configured correctly by ucode — the gateway's upstream OpenAI project just can't serve that snapshot from the CI region:

ERROR: stream disconnected before completion: The requested model snapshot is not
available for your project's geography.

The gateway relays that as a bare INTERNAL_ERROR ("invalid response from an upstream server"), so the launch dies after codex-cli exhausts its five reconnects. No ucode change can fix it, which is what makes it an exclusion rather than a bug.

Net effect is more coverage, not less: 13 models tested before, 15 now.

Stopped truncating away the cause. The failure report used the first 200 characters of stdout/stderr; it now uses the last 1500. Direction matters more than length: codex-cli logs a non-fatal model-listing error first and the real cause last, so a prefix reliably reported the wrong problem. The geography failure above presented as a /v1/models routing error — the message died mid-word at Request path '/ai-gat — and cost two rounds of misdiagnosis before the tail was visible.

Testing

Verified against a workspace in the CI region:

• each of the three re-enabled models returns a response
• gpt-5-3-codex returns INTERNAL_ERROR
• tests/test_e2e.py::TestCodexLaunch::test_launch_codex_per_model passes across all 15 remaining models (73s)

Note for reviewers

The exclusion list is keyed to whatever workspace UCODE_TEST_WORKSPACE points at, and regional availability has now shifted twice. I verified against a workspace in the CI region, but that secret is masked, so I can't confirm it is the same one — if it isn't, the 5-6 removals could re-introduce failures.

Treating a gateway-side availability error as a skip rather than a failure would remove this maintenance burden entirely, since a model ucode configured correctly that the gateway can't serve isn't a ucode regression. That's a larger change than this fix, so it's left as a follow-up.

This pull request and its description were written by Isaac.