[Benjamin-eecs]

Fixes #1157.

Problem. review.evidence._link_target checked an absolute symlink target after stripping the workspace prefix but before normalising it. A link the agent made as /root/analysis/scripts/../figures resolves inside the workspace, yet _relative_path rejected its .. part and the whole capture failed with EvidenceError: Unsafe evidence path: 'analysis/scripts/../figures', so a verified rollout was never reviewed. Relative targets were already normalised first.

Change

• src/benchflow/review/evidence.py: normalise the absolute target after relative_to(workspace) (posixpath.normpath). A result that is .. or starts with ../ raises "Evidence symlink escapes workspace", as an absolute target outside the workspace already does; anything else is checked and returned relative to the link's parent, as before.
• tests/test_review_evidence.py: a unit table for _link_target (in-workspace .. kept, . kept, plain absolute unchanged, two escapes refused), an end-to-end capture of an absolute in-workspace .. link through capture_workspace, prepare_workspace_upload and install_uploaded_workspace, and a negative control for a link that climbs out after normalising.
• CHANGELOG.md: a Fixed entry.

Validation

• Against main's code the new tests fail (6 of 8, all with Unsafe evidence path); with the fix all pass.
• tests/test_review_evidence.py, test_review_evidence_runtime.py, test_acceptance_evidence.py: 137 passed. Full suite with an empty HOME: 6599 passed, 4 failed; the same 4 fail on main's code here (no uv on the test PATH, no playwright browser, two host-specific fixtures). ruff check and ruff format --check clean.
• FrontierPhysics-Infra has run the same normalisation as a local patch (3111e65, bench patch 43) on 0.7.8 since 2026-10-10.

Not in this PR: release/0.8's kept_link (PR #1130) needs the same normalisation, otherwise it drops such a link as symlink_escape.