eval-harness v0.5.0

v0.5.0

• Adds typed capability, regression, and product evaluations; portable manifests; strict grading; promotion readiness; history, JUnit, and SARIF reporting; and the composite action.
• Documented 0.x breaking change: default shell checks now allow only constrained jq/printf/wc -l pipelines. Migrate other checks to typed graders or use unsafe_shell: true / EVAL_ALLOW_UNSAFE_SHELL=1 only with trusted case content; this is not an OS sandbox.
• Adds the Janus 0.1.10 offline benchmark. The matched file-output case did not invoke Janus's candidate executor, while checks: [] was accepted as PASS with zero checks; native adapter recommendation remains no-go until both behaviors are fixed.
• Benchmark used deterministic stubs. No live model quality, cost, latency, or reliability measurements were made.