Fix Braintrust redaction false positives

[rgarcia]

Summary

The first full ClawBench run graded all 129 trials but published nothing to Braintrust: the publisher's safety check rejected the payload. The redactor had three false positives, all from code inherited from kernel-mcp-server:

• Labels followed by a line break. In a page snapshot, "Email address:" followed by a new line was read as an address field whose value was the next lines.
• Method calls. Agent code like const cookie=page.getByRole(...) was read as a cookie= credential.
• Quoted labels with spaces. In 'Work email':'[REDACTED]', the checker took the label's closing quote for an opening one, then skipped [REDACTED] as if it were a snapshot attribute.

In each case the redactor and the safety check disagreed, and one disagreement aborted the whole experiment.

Changes:

• Fix all three cases in both the redactor and the checker.
• Before the final check, the publisher now replaces any string that fails the check with [WITHHELD] and logs a count. An unexpected value costs that value, not the experiment. The safety check still runs on the final payload.

Validation

• bun test passes (21 tests, including regressions for each case and for withholding), along with tsc and prettier.
• Building events for every local Harbor job directory now succeeds with nothing withheld, including a previous 129-trial run (9,894 events). Before this change, each of them failed.

[rgarcia]

Validated in CI: run 38440542932 published 129 rows to Braintrust with nothing withheld. In the next run, 4 values were withheld and the experiment still published.