docs: define RTL audit evidence policy - #6046

Problem

The RTL audit currently mixes implemented checks, proposed verified-N/A handling, and proposed known-debt handling. An all-N/A result can be mistaken for positive RTL evidence, while source lint, browser geometry, interaction checks, and screenshots prove different things.

Changes

Add a draft architecture:rtl-audit-evidence record that:

• distinguishes measured, verified-N/A, coverage-gap, behavior-failure, stale-evidence, and proposed known-debt states;
• keeps source lint, rendered geometry, and interaction evidence within their actual proof boundaries;
• makes the live component roster the coverage owner;
• states that behavior failures have no allowlist;
• separates affected-component PR scope from the full weekly sweep;
• requires mutation proof for new audit dimensions;
• records today's enforcement gaps instead of describing proposed behavior as shipped.

The draft explicitly records that D7 checks hit-testing and center alignment but does not yet enforce target size. It also records that the reason-only verified-N/A registry detects newly applicable dimensions but does not bind evidence to source/story bytes.

Decisions needed before promotion

1. Admit verified N/A with fail-closed evidence binding, or require executable rendered measurement for every component.
2. Admit a visible removal-only baseline for pre-existing unexplained gaps, or keep every gap failing.
3. Define the clean-state and stability prerequisites for making RTL a required merge check.

The draft also calls out the missing architecture:bidirectional-layout owner for shared RTL implementation mechanics; this evidence record does not claim that ownership.

This PR is intentionally draft and remains non-authoritative until those policy choices are approved.

Test plan

• node scripts/check-knowledge.mjs
• git diff --check
• Verified current behavior against the coverage builder, D7 browser check, verified-N/A registry, CI summary, and proposed fix(rtl): baseline pre-existing applicability gaps #5988 debt registry.
• Reproduced the fix(core): center coarse-pointer touch targets in RTL for Checkbox, Radio, and Switch (#5170) #5177 false-green: target centering remains measured when the native ×ばつ24 expansion is removed because D7 does not currently assert size.