[JacobStephens2]

Summary

Trial Session evidence stays readable when unrelated Trial metadata or usage spans are corrupt, including completed legacy Experiments.

CLI Session command / Console Trial Session page
  Experiment Session evidence reader
    validate saved Experiment metadata
    discover, identify and confine Session evidence
    read selected UTF-8 contents
Full Experiment view
  same evidence selection over captured attempts
  existing progress, accounting and Arm usage

Session IDs, converted-log preference, JSON and rendering stay stable. Retained filesystem paths remain private. Navigation, guide and site descriptions now explain the shared owner and independent reads.

Closes #1302

Evidence

• Before: cargo test --test console selected_session_evidence_survives_unrelated_corruption_and_legacy_verdicts failed on malformed unrelated usage spans (expected value at line 1 column 1).
  After: the same regression passes for malformed spans, results, pending history, active markers and attempt indexes, with running and completed legacy Experiments.
• cargo test --bin thirdshift evals::experiment::: 37 passed, including shared evidence reads and retained capture/accounting regressions.
• cargo test --test console experiments::: 4 passed, including exact Session JSON, raw/converted rendering, escaping, confinement and unavailable responses.
• cargo fmt --check and npm run docs:check --prefix site: passed.
• cargo clippy --all-targets -- -D warnings: passed.
• /usr/bin/time -p scripts/affected-tests main: passed. All 50 site browser tests and the full Rust suite passed (cargo test fallback); 2,001.64 seconds total, including compilation and site checks. The Rust unit suite had 1,441 passed and 3 ignored tests; all integration suites passed.

Seams under test: the Experiment-owned Session evidence read with Native and Observed filesystem adapters; the captured full Experiment view; the CLI Session command paired with the Console HTTP page.

Review coverage

Independent thirdshift-code-review against main: Standards 0 findings; Spec 0 findings. Each reviewer read all 16 changed files, including the generated documentation hunks. Unread changed files: none. Generated-output exclusions: none. Both reports, including their final files-read lists, are retained with the Session evidence.

Unaddressed findings

Standards

None.

Spec

None.

Security

None.

Security review outcome

No unaddressed introduced findings.

Merge Danger

Door: two-way

Retained-reader refactor with no persisted-format or Experiment-execution change.

Blast Radius: Evals

CLI and Console Trial Session evidence reads and full Experiment Session listings share the selection implementation.

Built with codex · gpt-6.1-sol · xhigh