[vicgupta]

Closes #260

Summary

Adds an experimental eval input to the GitHub Action so CI can run tinyagent --eval suites and gate on pass rate.

• New inputs: eval (suite YAML), eval-out, eval-min-pass-rate, plus cost-limit and time-limit (shared across the suite with eval).
• eval cannot be combined with task, task-file, review-base, review-commit, attempts, model or profile (the suite defines its matrix); eval-out/eval-min-pass-rate need eval.
• New outputs: eval-out, eval-report (report.json path) and eval-results (compact JSON: per-config pass rate and cost, below_threshold).
• Appends the eval report.md to the job summary (respecting the 1 MiB limit); job-summary and fail-on-error behave as for tasks. Artifact upload is left to the workflow.
• Docs: README, docs/eval-framework.md (CI example, cost limits, full-history checkout, fork-PR guidance), CHANGELOG.
• Tests: action_eval_test.go. The existing script-safety test is narrowed to catch the eval builtin in command position only, so flags like --eval and outputs like eval-out no longer trip it (with its own positive/negative test).

Checks: gofmt, go vet ./..., go test ./..., GOOS=linux go build ./... all pass.

Generated by a Cobra run (run_mvvnyk0j901m) that was cut off before it could push; finished and opened by hand.