docs(review): falsification + shared-field-semantics guards in pre-PR checklist

[alonbilu]

Two guards for the /review checklist, distilled from a real miss on a downstream project (a validator rule read an LLM-echoed field as advisory while sibling rules read the same field as committed — every isolated test passed, the two were contradictory together; it survived several spec revisions before review caught it).

Added under Testing:

1. Falsify the acceptance test — confirm a new rule/acceptance test goes RED without the change under test. A test that passes because it never exercised the case is worse than none; fixtures shaped to match the implementation (single-mode / edge-case seed data) are the tell.
2. Shared-field semantics — when a new check reads a field another rule/consumer already reads, verify they agree on what it means. Divergent interpretations pass isolation yet contradict each other. Grep the field name across rules/consumers.

Guard (2) is the one that would actually have caught the miss — falsification alone wouldn't have, because the original (wrong) test and the bug agreed.

🤖 Generated with Claude Code

[alonbilu]

Updated (force-push):

• Generalized the shared-field guard so it reads as a pattern for any shared field (denormalized copy, user-supplied value, upstream/LLM echo, cached flag) with the committed-vs-advisory case as one illustration — not a domain-specific footnote a reader pattern-matches too narrowly.
• Added a backlog note (docs/FUTURE-ROADMAP.md, commit 8779350): adding these guards surfaces a meta-question — each seeded repo vendors its own copy of review.md et al., which silently drifts behind the canonical starter (the same canonical-value-in-two-places failure class the guard itself is about). Deferred; flagged to decide before the repo count grows (vendor-per-repo + drift check, or reference a single source).