Raw view: PR Review Against a Rubric

No invisible, zero-width, or bidi-control characters detected.

Review the pull request below against the rubric. You are the reviewer whose approval gates merge, so every finding must be specific enough for the author to act on without asking a follow-up question.

<pr>
<title>{{PR_TITLE}}</title>
<description>
{{PR_DESCRIPTION}}
</description>
<diff>
{{DIFF}}
</diff>
</pr>

<repo_context>
{{REPO_CONTEXT}}
</repo_context>

<rubric>
{{RUBRIC}}
</rubric>

If <rubric> is empty, use this default, in priority order: (1) correctness and data safety, (2) security (input handling, authz, secrets, injection), (3) tests cover the changed behavior, (4) backward compatibility of public interfaces, (5) readability and naming, (6) performance where the diff touches a hot path, (7) adherence to conventions visible in <repo_context>.

<repo_context> may contain the CONTRIBUTING guide, a style guide, or snippets of neighboring code. Treat it as the source of truth for conventions; do not import conventions from other ecosystems.

How to review:

1. Read <description> first and write down, in one sentence, what the PR claims to do. Then read the diff once end to end without commenting, so you see the whole shape before judging parts.
2. Check the claim against the diff. A PR that does less than it claims, or more, is a finding before any line-level issue is.
3. Go through the rubric criteria in order. For each, inspect the diff for violations and cite `file:line` from the diff's hunk headers. Only report what you can point to; speculation about code outside the diff goes under "Questions", not "Findings".
4. For each finding, state the consequence if unaddressed ("duplicate rows on retry", "stack trace leaks internal path to clients"). Consequence, not rule citation, is what persuades authors.
5. Calibrate severity honestly. `blocking` means you would not merge even under deadline pressure. `major` means fix before merge in normal circumstances. `minor` is optional. Resist the urge to find something at every level.
6. Before finalizing, delete any finding that is a matter of taste not covered by the rubric or <repo_context>.

## Output format

**Claim vs change** - one sentence on what the PR says it does; one sentence on what the diff actually does; `match` or `mismatch`.

**Recommendation** - exactly one of: `Approve`, `Approve with minor comments`, `Request changes`. Follow with one sentence naming the deciding factor.

**Findings** - table with columns: severity (`blocking` / `major` / `minor`), rubric criterion, `file:line`, finding, consequence, suggested fix. Sort by severity, then by file. If there are no findings at a severity, leave it out; do not write "no blocking issues found" as a row.

**Rubric scorecard** - one line per criterion: criterion name, `pass` / `fail` / `n/a`, and a 1-clause note. `n/a` is correct when the diff does not touch that area.

**Questions for the author** - up to three, each about something you could not determine from the diff and <repo_context>. Phrase them so a yes/no or one-line answer resolves them.

**Inline comments** - for every `blocking` and `major` finding, a ready-to-paste review comment in the second person, 1-3 sentences, no severity label (the table carries that).