test(evals): confirm the eval gate triggers, and show output when it fails

[walkerhughes]

Two purposes, one file changed.

1. Confirm the trigger

This PR touches only tastytrade-mcp/evals/validate_local.sh. If the tastytrade-mcp workflow runs and its Validate eval task verifiers step executes, the paths filter is confirmed to fire on eval-only changes. If the harbor-mcp jobs stay absent, per-subdirectory scoping is confirmed in the same run.

For reference, the step already ran green in #7, which touched tastytrade-mcp/README.md:

7. Validate eval task verifiers: success

This isolates it to an eval file specifically.

2. Make the gate useful when it fails

validate_local.sh discarded both runs with >/dev/null 2>&1, so a failure printed only a task name and two reward values:

FAIL  net-liq-value  (oracle=0 negative=0)

That is enough to know something broke and nothing about what. Diagnosing meant reproducing locally. The oracle, verifier, and negative-control output now go to files inside the existing $work tempdir and print, indented and tail-limited, only on failure. Passing runs are byte-identical.

Verified by breaking one oracle deliberately:

FAIL  net-liq-value  (oracle=0 negative=0)
      --- oracle ---
      simulated oracle failure: endpoint unreachable
      --- verify ---
      verifier error: [Errno 2] No such file or directory: '.../app/answer.json'
      reward=0
      --- negative ---
      verifier error: 'net_liquidating_value'
      reward=0

11 passed, 1 failed

The cascade is visible: the oracle exited non-zero, so no answer.json existed for the verifier to read. Restored, 12 passed, 0 failed.

The negative control was already sound and is untouched: it still requires reward 1 from the oracle and reward 0 from an empty answer, so a verifier that blindly awards 1 fails.