Clarify verification receipt reuse across roles

[mzored]

Summary

Delegates, leads, and reviewers can inspect and reuse a passing verification record when the command, tree fingerprint, and relevant environment match. Role changes and reporting stages do not require another execution. Changed inputs, missing or unverifiable evidence, or a concrete named doubt justify reruns; a different fingerprint never proves the current tree passed.

The kernel carries the invariant, verification defines the minimal record, and delegation and integration reference it. Five original arm-aware probes cover matching receipts, changed trees and environments, scratch inputs, unavailable records, uncertain capture, and conflicting results. Required coverage, independent review, destination inspection, and agent-verify inherited remain intact. No host implementation changed.

Verification

• Full local python scripts/check.py passed on 7add235 with fingerprint 85dfdc3f8d6f0e79 unchanged before and after execution. Its headroom lint required a bounded editorial correction to delegation.
• Final candidate: 19 affected package/document/contract checks passed after that correction; git diff --check is clean and reference headroom lint is empty. The earlier full result is scoped to its own fingerprint, not presented as an exact-tree pass for this revision. PR and main CI provide current-revision full evidence.
• Independent isolated Codex source review approved with no qualifying findings. The later correction only tightens delegation wording to preserve its read reserve. Control and review transcripts showed no personal skills or installed plugin; review resources were removed.
• Local strict Claude schema validation passed before the editorial-only correction. Codex schema validation is unavailable locally and is required in CI. Clean installs, activation, continuity, model behavior, and savings remain UNVERIFIED.

Measurement and release

The fixed sample contains 28 same-label, same-fingerprint candidates among 116 completed runs, 24.1% of runs and 9.6% of execution time. This is a proxy: command identity and session/role correlation are absent from the host journal. Private transcript inspection identified output/status recovery and legitimate changed selections or external inputs. The exact same-command repeat rate and causal attribution remain UNVERIFIED.

This is one PATCH release, 6.3.3, clarifying existing evidence reuse. Package structure and installation are unchanged. The evidence ledger records host telemetry and stable-capture gaps without publishing private logs or changing host configuration. The unrelated contributor-skill edit is excluded.