results: AIOpsLab audit with its proposed fix, ITBench library audit, agent study, SREGym HEAD alert baseline, and k8s-ai-bench repeats and cross-task check

[TheDeadcoder]

Adds the results of the studies merged in #27 to #32, with dates and clock times scrubbed.

• AIOpsLab audit: runs/ (first attempts and repeats), runs-patched/ with the proposed fix, pilot/, and their
  ledgers.
• ITBench library audit: runs/, pilot/ and the ledger.
• Agent study: the Claude Code runs under runs/, the smaller model's runs under runs-haiku/, pilot/, and both
  ledgers.
• SREGym HEAD alert baseline: runs-head/ and ledger-head.json.
• k8s-ai-bench: the 15 repeated episodes and the cross-task check under interference/. The ledger now scores first
  attempts on their own. All 15 repeats got the same label as their first attempt.

It also widens the scrubber so it removes clock times and dates that follow an escaped newline, as in the agents'
JSON transcripts. No committed file changes under the new patterns.

Four agent episodes hit the plan's usage limit and were run again. The set-aside copies are not included.

Every study's tests pass, and every ledger's --check matches its runs.