Fail fast on invalid benchmark output

[silverstein]

Summary

• validate the benchmark output path before reading the corpus or launching PDF Tools
• reject missing or unsafe output parents and occupied targets immediately
• repeat the no-overwrite checks during final atomic publication
• make the layout determinism test reuse one exact PDF byte source rather than compare time-varying generated files

Regression coverage

• proves an unavailable parent processes zero documents
• proves an occupied output processes zero documents and remains untouched
• reproduces and closes the CI-only fake-PDF hash flake under Node 20.19
• affected suites pass on Linux under Node 20.19 and 22, and on macOS under the project Node 22 runtime

Scope

The product change is limited to the offline olmOCR benchmark runner. The additional test-only correction makes an existing layout determinism assertion compare two runs over the same source bytes. Shipped PDF behavior and benchmark claim policy are unchanged.