fix(evals): coalesce null output for Braintrust reporter and guard log() throws

[iroiro147]

What

A no-turn eval (e.g. t.target.dispatchSchedule(...) + t.check over DB state) legitimately produces result.output === null per the eval API's own derivation. Braintrust's validateAndSanitizeExperimentLogFullArgs rejects null/undefined output ("output must be specified"), and the throw escaped onEvalComplete, killing the entire eve eval run — remaining evals never executed and no artifacts were written, even though the crashing eval itself passed all its gates.

Two fixes in BraintrustReporter.onEvalComplete:

1. output: result.result.output ?? "" — coalesce null to an empty string so the Braintrust SDK accepts it.
2. try/catch around experiment.log() — any reporter throw is logged to stderr but does not abort the run. The remaining evals still execute and artifacts are written.

Regression tests

Added to braintrust.test.ts:

• coalesces null output to empty string for no-turn evals — drives an eval result with output: null, asserts log receives output: "".
• survives a log() throw without aborting the run — mocks log to throw "output must be specified", asserts no rethrow, then asserts a second call still reaches log (proving the next eval is not dropped).

pnpm exec vitest run --config vitest.unit.config.ts src/evals/runner/reporters/braintrust.test.ts — 7/7 tests green (5 existing + 2 new).
pnpm exec tsc -p tsconfig.json --noEmit — clean.

Linked issue

Refs #1405

[vercel]

@iroiro147 is attempting to deploy a commit to the Vercel Team on Vercel.

A member of the Team first needs to authorize it.