The 'Command line' block in the README lists --quote-lifetime <s>, but src/cli.ts only defines quote-ttl in its parseArgs options (and its own --help prints --quote-ttl <s>). Node's parseArgs runs strict, so the documented option aborts the process. This breaks the task's acceptance rule that every command in the README runs as written against the mock, and the README test added in this change does not catch it because it only executes the 'Try it by hand' block.

The assignment asked for a test that runs the README's commands and confirms the documented outputs match. The test extracts only the 'Try it by hand' block and asserts nothing about its output beyond what the block's own test "$STATUS" = 402 enforces; it then re-issues two requests itself and compares only the set of action names and the key set of one check response.

The documented quote response (201 {order:{id, status:"quoted", quote, ...}}), the 402 body fields, the 409 request_key_conflict and 422 invalid_input cases, and the other README shell blocks (node dist/cli.js conformance --flaky, conformance --url ..., the 'Command line' options) are not exercised. The suite therefore stays green while the README is wrong.

CHECKS in src/conformance.ts has 22 entries and the TAP plan line printed by every conformance run is 1..22. The two checks added since the number was written (same-bytes replay returns the first outcome; a schedule is priced per run) are not in the README's list either.

The test count in the Development section was not updated when test/readme.test.ts and earlier tests were added. The listed suite names also omit the README and crypto suites.

test/fixtures/live/capabilities.json, captured from GET https://api.imd.fun/requests/capabilities in the same change, lists the same seven actions (job.open, job.continue, launch.open, oracle.request, workflow.open, schedule.create, schedule.topup), the same policy versions, the same asset 0xd34a99bc0f67ae1bbd63c660e6d0b0dd03e263b7, the same amount 500000000000000000 and the same 600 s quoteTtlSeconds as the mock.

The only payment field that differs is payTo (live 0x4e0fa57b..., mock 0x8dc3ade1...), which the next bullet already covers. The README bullet tells readers the catalogue and prices are invented when the project's own evidence says they are not.

The poll response has a top-level status (the value a client must poll on, per src/client.ts poll()), plus payment and admission. The mock's own openapi summary says 'Order status: {status, order, payment, admission}'. Similarly, step 2 omits the top-level created flag of the quote response and step 3 omits the input field the 402 challenge carries.

These are omissions rather than wrong values, so no command fails, but a reader following the table polls order.status, which stays paid and never reaches admitted.

The check response no longer has a verdict; it has blockers. The bullet still uses the removed term. Separately, the live fixture test/fixtures/live/check-job-open.json shows that the exact body used in 'Try it by hand' ({"action":"job.open","input":{"objective":"Say hi."}}) is blocked on the live API with bad_path_count ('expected between 1 and 16 allowed paths') and plan: [], whereas the mock returns no blocker and a one-step plan.

The README says only that the shapes match; a sentence noting that the canned verdict differs from live for this input would stop readers from treating the mock's empty blockers as what the live evaluator would say.