Fixes the receipt-normalization command in provider-dispatch instructions. The old command silently exited successfully without evaluating receipts. The corrected command invokes the shipped executable wrapper, returns a policy event for valid input, and rejects invalid input.
Regression tests execute the command extracted from the documentation and cover successful normalization, model-identity mismatch, and malformed input. Resolves #99.
Cursor baseline: 0.15.5, commit 12d587dfb20741cafc376c42c696c5f6e2a64487. Adaptations and exclusions. Previous tested release: v1.10.0.
Validation: PR #100 records 744 Bun tests, strict typechecking, static/plugin/maintenance/documentation checks, GitHub CI, independent review and real installed Codex and Claude Code checks. All 213 packaged files matched the candidate in both installations. Release checkpoint ea09860fa4e100f78447c34abc343d7e1e0eb69d has the same package tree as tested candidate efa07c0beefe37c9f9fd254b704a074ee3d593f7.
In each parent, a preserved authentic receipt produced a 970-byte complete JSON event; a wrong-model copy was rejected with exit 65 and empty stdout. These were receipt replays, not fresh external provider executions. Independent review confirmed the new regression tests fail with the old instruction. No Claude quota waiver was needed; Copilot was quota-limited and supplied no approval.