feat(review): shorten and simplify review comments

[zfarrell]

Adds a ## Comment Style section telling the reviewer to write its comments in simplified technical English, one finding per comment, with the proof scaled to severity while the consequence is always stated.

Measured by rewriting 14 real posted comments from this repo (#38) and runtimedb (#1242, #1236) under each candidate ruleset:

Ruleset	Total prose words
As posted today	2,338
Sentence caps + <details> folds	−2%
Proof-scaling, first draft	−60%
This PR	−46%

Two earlier approaches were dropped with cause. Sentence caps failed three independent ways (a 4-sentence super nit cannot hold claim + evidence + consequence + fix; the count is gameable, since a fix costs 0 sentences as code and 1 as prose; and the word cap and sentence cap compete). <details> folds cut only 2% of real volume — they relocate text, and folded text returns whole through <prior_review_comments> on later cycles — while hiding scope-limiting facts, which made comments read more severe than the originals. A first draft of proof-scaling coupled proof to convention-existence and dropped the consequence on the two highest-impact findings in the set; that clause is gone and the consequence is now unconditional.

Known limits, worth watching rather than fixing blind:

• Severity tiers conflate likelihood with blast radius, so an improbable-but-catastrophic nit gets the thinnest argument. The consequence is stated, so the gap is persuasiveness, not information.
• Absence-claims ("X is not what this code does") structurally need two cited facts where a nit allows one.
• One test comment lost its scope-limiting preconditions and read slightly more severe; the two reviewers read that rule differently.

The prompt document is measured at runtime and subtracted from the diff's context allowance, so this section is sized deliberately: 1,193 bytes, leaving 107 before tests/context-step-test.sh starts omitting diffs that currently fit.

Deploys org-wide from main on merge. Rollback is a one-file revert.

[pullfrog]

Important

The one-finding rule conflicts with required blocking summaries and should be scoped before merge.

Reviewed changes Reviewed the new organization-wide review-comment style contract and its interaction with the existing output rules.

• Controlled language — Adds sentence length, voice, terminology, uncertainty, tense, and pronoun constraints for inline and summary comments.

• Comment structure — Requires claim-first feedback, fix-before-derivation ordering, and one finding per comment.

• Severity-scaled evidence — Requires every finding to state its consequence while varying supporting proof by severity.

  ^{｜ Fix all ➔ ｜ Fix 👍s ➔ ｜ View workflow run ｜ Using GPT Sol ｜ 𝕏}

[claude]

Prior nits are addressed: line 155 scopes the one-finding rule to inline comments, and line 149 exempts the context disclosures. No new findings.

CI has not reported yet, so the prompt-document byte budget in tests/context-step-test.sh is unverified here.

[pullfrog]

✅ No new issues found.

Reviewed changes Reviewed the prompt-contract fixes added since the prior Pullfrog review.

• Inline finding scope — Limited the one-finding rule to inline comments, preserving the required aggregate blocking summary.
• Context disclosures — Exempted mandatory partial-context disclosures from the subject-deletion rule.
• Proof wording — Simplified the blocking-proof instruction without changing its requirement.

  ^{｜ View workflow run ｜ Using GPT Sol ｜ 𝕏}