Check a real kimi run's wire.jsonl against the Kimi Harness's usage parser #1925

Finding

Spec review of #1923 (the Kimi Code Harness, #1888), finding P2: the Ticket says "Before the Harness is marked done, a real kimi run's wire.jsonl is checked to have the step.end usage shape above. If the shape has changed, the parser follows the real output and a test fixture records it." That check was not done.

Evidence

• tests/fixtures/normalized-harnesses.md: kimi-wire.jsonl is "reconstructed from Kimi Code 2.5.2's published source … not recorded from a live run".
• The shape was read from the 2.5.2 npm bundle: finishMachineStepProjection appends {"type":"step.end", …, "usage":{"inputOther","output","inputCacheRead","inputCacheCreation"}} as the event of a context.append_loop_event record; writeResumeHint prints session.resume_hint only after a completed turn; createWireMetadataRecord writes protocol_version: "1.5".
• The implementing session had no Fireworks or Kimi credentials, and running the downloaded CLI against a local stub provider was refused by its permission policy.

Why the call is not the implementer's

It needs a person with a Kimi sign-in or a Fireworks key to run, for example, kimi --output-format stream-json -m <alias> -p "Run echo hi, then say done." and compare ~/.kimi-code/sessions/*/<id>/agents/main/wire.jsonl (and the stdout lines) with tests/fixtures/kimi-wire.jsonl and tests/fixtures/normalized-kimi.jsonl. If they differ, update the parser in src/harness/kimi/wire.rs / stream.rs and record the real lines as the fixtures.