[mzored]

Summary

A 7.0.2 field session iterating on an interface left the owner repeating one correction on instance after instance, and three times a shared piece changed for one screen broke another. 7.0.0 had narrowed the kernel's class rule to "when a failure repeats, fix the shared cause", which said nothing about other instances or an owner's design correction, while docs/decisions.md still promised the other instances.

• Kernel: a defect or an owner's correction covers every place that shares its cause or role, fixed where the cause lives with no guard per place.
• Fast-fixes: feedback on one element is feedback on its role, applied through the shared piece and reported with the places it reached, unless the owner limits it or the others serve a different purpose. An interface change is inspected on every screen that renders it, against the state it replaces.
• Offset: the kernel drops "re-review only the fixes and stop when a round returns nothing that qualifies", which review.md carries before any review. Kernel 9,808 → 9,790 characters.
• Decision record, version files, and changelog follow.

The field read, prior-art check, and the one-round Codex review are in CHANGELOG.md.

Validation

Capability	Status
Deterministic package gate (scripts/check.py, git diff --check)	PASS locally
Codex and Claude schema validation	CI
Clean install on each host	UNVERIFIED (no layout or manifest-shape change)
Explicit invocation, implicit activation, continuity	UNVERIFIED
Behavioral suite	UNVERIFIED, see docs/evidence.md; the owner declined paid runs

🤖 Generated with Claude Code