Campaign adapter: make Claude timing totals deterministic

[jeffhuber]

Problem

During the six-provider v1.2.0 qualification campaign in #853, the maintained Claude adapter executed successfully twice but both generated result files failed closed-schema validation because the sum of step elapsed_seconds values differed from the top-level elapsed_seconds beyond tolerance.

A bounded recovery invocation that explicitly required integer step times and an exact top-level sum produced a valid Claude-authored result. Authentication, CLI invocation, package installation, and the release candidate were not the failure.

Ask

• Make the maintained Claude campaign adapter produce timing arithmetic that satisfies the closed code_mower.adoptionResult.v1 contract reliably.
• Preserve provider-authored evidence and fail-closed validation; do not silently rewrite or fabricate a provider result.
• Do not add an automatic paid retry.
• Add a focused regression test for inconsistent timing totals and the corrected prompt or adapter behavior.
• Add a documentation line only if operator behavior changes.

Privacy

Keep the existing metadata-only boundary. Do not store or upload prompts, transcripts, source, diffs, raw stdout/stderr, auth output, local paths, or secrets.

Campaign evidence: #853. Both required providers ultimately passed; the two rejected Claude attempts remain in local campaign history.

[jeffhuber]

Shipped in v1.2.1.

• PR Make Claude campaign timing deterministic #855 added deterministic Claude timing instructions while preserving fail-closed provider-authored evidence and avoiding paid retries.
• The maintained Claude adapter passed a live first-attempt canary with exact step arithmetic: 3 + 12 + 0 = 15 seconds.
• Release PR Release Code Mower v1.2.1 #857 passed both peer audits, the full test suite, package rehearsals, and Python 3.12/3.13/3.14 CI.
• v1.2.1 is the GitHub latest release and is published on PyPI.
• The metadata-only CodeMower.com dogfood bundle was uploaded successfully.