PR #2081 — A Conformance Suite That Names Adapter Drift Instead of Discovering It Live

The Broader Goal at the End of the Previous PR

v0.24.0 still needs dispatch to be trustworthy across every registered harness. The GOVERNS preflight now names a missing --issue/--pr/--story before #1990 can hard-fail. The remaining gap called out in the three-lens review was adapter quality: receipt markers, cost attribution, and failure classification were only checked live, after a regression had already shipped.

Strategic Shifts in This PR (if any)

No strategy change. #2062 still owns the incomplete strangler retirement. This PR only ships the narrowed Task D from the 2026-10-06 three-lens follow-ups plan: a parametrized conformance suite over the five registered adapters. Adapter implementation drift is tracked, not patched here.

What This PR Shipped

• tests/test_adapter_conformance.py parametrizes claude, codex, agy, grok, and local for Protocol method presence, translate_permissions signature match against HarnessAdapter, receipt-marker parsing through jobs._check_task_receipt, classify_failure returning FailureKind or None, and one extract_tokens fixture per harness.
• Claude/Codex/Agy translate_permissions is missing the Protocol's skip_permissions parameter. Those three cases are xfail against #2078.
• LocalAdapter.parse_output leaves compatibility_evidence unset. Those two receipt cases are xfail against #2079.
• No adapter source files were changed.

Plan: docs/superpowers/plans/2026-10-06-three-lens-followups-plan.md Task D. Design: docs/superpowers/specs/2026-10-06-three-lens-followups-design.md.

Brainstorm Visuals Used

None. This follows the committed design and plan.

What This Achieved on the Path to Autonomy

A new adapter, or a signature change on an existing one, now fails in pytest instead of in a live dispatch. The suite is green because the known drift is named and ticketed, not because it was papered over.

Strategic Note: The Goal at the End of This PR

#2078 and #2079 are the next adapter-side fixes. The strangler-path retirement still listed on #2062 stays a separate implementation task.