docs(agents): require an evidence class on every review finding

[vsits-proxy-builder]

Problem

The global reviewer baseline (~/.codex/AGENTS.md:155) says to distinguish what is confirmed correct from what is assumed or hypothesized. That's a disposition, and a disposition is self-assessed — a reviewer who read a code path and found it plausible will honestly file it under "confirmed." Nothing about the output tells the next reader which it was.

Same shape as the bloat gap we closed in #287: "None" isn't checkable, a number is. "Confirmed" isn't checkable, a named command is.

The motivating case

On #270 the reviewer endorsed a claim about agent-id availability that measurement later destroyed. Two model families agreed it was true; the data disagreed. Cross-family review — our main structural guard — did not catch it, because correlated plausibility is exactly what agreement measures. One query did.

Re-measured today against a now-larger log while writing this: still 0 of 184,976 usage rows populated. The finding held, but I'd have been repeating a June number as a present-tense fact if I hadn't checked — which is the failure the rule exists to prevent, committed inside the rule itself.

Change

Every finding, blocking or not, carries one of three classes:

• Measured — name the command, paste the result
• Read — cite file:line, say plainly it's a code read
• Reported — name the source, say you did not reproduce it

Plus a hard rule: a load-bearing claim from a PR body may not be repeated as fact in a review without independent measurement.

Binds every reviewer — Codex, our own maintainer comments, and any third model we add later. Codex already did this well in places today unprompted ("I reproduced this locally: changing OLD to NEW … still classified as normalized"; separately "I could not verify the author's live-traffic claims"). This makes it consistent rather than occasional.

Why now

We're considering adding a third reviewer for complex work. More reviewers means more findings to cite-check, and that cost compounds if each arrives without provenance. Classes make reconciliation cheaper rather than 1.5× more expensive — Measured from one reviewer and Read from another on the same point is a signal, not a contradiction.

Non-Functional Requirements

• Size/complexity budget — 49 lines in AGENTS.md. Documentation only, no code.
• Threat model — n/a. No inputs, no trust boundaries, nothing on the wire.
• Maintainability — no new abstraction. Sharpens an existing global rule into a checkable form and records why, so it isn't softened back later.
• Performance/reliability — n/a.
• Load-bearing? No. Documentation governing review behavior.

Testing

Documentation only. Both figures cited in the new section were verified rather than quoted: the proxy-read-dedupe.test.mjs failure was reproduced on a merged worktree earlier today, and the agent-id figure was re-measured against the current log. Dates attached to both.

— Proxy Builder