Mark current Bifrost regressions as expected

[DavidBakerEffendi]

Description

Scheduled benchmark run 33842914977 reported 10 new Bifrost semantic failures. This change marks those cases as current expected failures through a runner overlay, so the scheduled benchmark reports them as expected_failure while the content-addressed legacy and preregistered evaluation case documents remain byte-for-byte unchanged.

Key Changes:

• Add a strict 10-case Bifrost expected-failure overlay with reasons taken from the run artifact.
• Add --expected-failures support and apply the overlay in the scheduled benchmark before scoring.
• Reject duplicate, missing, already classified, or expected-pass case entries, and test the checked-in overlay against all referenced cases.

Touch Points:

• benchmarks/expectations/bifrost-expected-failures.yaml
• src/runners/bifrost.rs
• src/main.rs
• .github/workflows/benchmark.yml
• benchmarks/README.md

Validation:

• cargo test --locked expected_failure_overlay (3 passed)
• cargo test --locked (256 passed; 3 existing local Git LFS archive-hash failures remain because the source archives are not materialized correctly in this worktree)
• Workflow and expectation YAML parsed successfully
• git diff --check