Automate a review finding as a lint or test: The line and the tints work and the tests are real, but three things need fixing #11894

The gate now refuses three things that reviewers kept asking for, and each refusal names its rule. I ran the gate, tsc and test/review-rules.test.ts locally and they pass. I did not run the full suite or the fast-test selection (scripts/fast.ts); the forge's check runs them.

What changed

• New rules: src/kernel/review_rules.ts gets parseAdded and addedProblems, beside the existing #8888 rules. They judge only the lines a pull adds, because the tree already has about 600 as never casts and thousands of long comment lines. Main's own lines are never refused.
• Gate: scripts/gate.ts gains addedRefusal, which diffs $FORGE_CHECK_BASE (or the merge base with main) to HEAD over src/ and test/. It runs with the other gate checks, so the check refuses a pull that adds these. If no base resolves, it passes, as the other diff-based gate checks do.
• The rules, each from a review comment that came back:

- a test builds the real input, not as never` (#11074): in test files, an added as never fails. - a test does not wait on a real timer (#11723): in test files, an added Bun.sleep or a setTimeout-based wait fails. A line that ends in // real-timer: <why> is allowed. - a comment line is wrapped at 100 columns (#11074): in src/ and test/`, an added comment line over 100 characters fails.

• Tests: test/review-rules.test.ts adds four tests. They cover the diff parsing, each rule's message with file and line, and the // real-timer: escape. The as never rule is checked not to fire in src/, and the 100-column rule not to fire in .md files.
• Gate on its own commit: the first commit failed the gate on its own test file, because the fixtures contained the banned text. I built those strings apart and rephrased one test name.

What I chose not to do

• Rules left out: two things from the 11074/11532 reviews have no rule, because a regex can't judge them.

- "Don't assert exact CSS source text". - "A link needs a 44px tap target or an accessible name". design-lint is the place for that, and the pull touches no src/web/.

• #11761: its finding was a behaviour (chapter without view) in that one tool, so I made no general rule for it.
• Old uses: I did not clean up the existing as never casts, sleeps or long comments.

To try it: on a branch, add x as never or await Bun.sleep(1) to a test file, commit, and run bun scripts/gate.ts. It prints the rule, file and line.

🤖 Generated with Claude Code

Closes #11884

The gate now refuses three things that reviewers kept asking for, and each refusal names its rule. I ran the gate, tsc and test/review-rules.test.ts locally and they pass. I did not run the full suite or the fast-test selection (scripts/fast.ts); the forge's check runs them.

What changed

• New rules: src/kernel/review_rules.ts gets parseAdded and addedProblems, beside the existing #8888 rules. They judge only the lines a pull adds, because the tree already has about 600 as never casts and thousands of long comment lines. Main's own lines are never refused.
• Gate: scripts/gate.ts gains addedRefusal, which diffs $FORGE_CHECK_BASE (or the merge base with main) to HEAD over src/ and test/. It runs with the other gate checks, so the check refuses a pull that adds these. If no base resolves, it passes, as the other diff-based gate checks do.
• The rules, each from a review comment that came back:

- a test builds the real input, not as never` (#11074): in test files, an added as never fails. - a test does not wait on a real timer (#11723): in test files, an added Bun.sleep or a setTimeout-based wait fails. A line that ends in // real-timer: <why> is allowed. - a comment line is wrapped at 100 columns (#11074): in src/ and test/`, an added comment line over 100 characters fails.

• Tests: test/review-rules.test.ts adds four tests. They cover the diff parsing, each rule's message with file and line, and the // real-timer: escape. The as never rule is checked not to fire in src/, and the 100-column rule not to fire in .md files.
• Gate on its own commit: the first commit failed the gate on its own test file, because the fixtures contained the banned text. I built those strings apart and rephrased one test name.

What I chose not to do

• Rules left out: two things from the 11074/11532 reviews have no rule, because a regex can't judge them.

- "Don't assert exact CSS source text". - "A link needs a 44px tap target or an accessible name". design-lint is the place for that, and the pull touches no src/web/.

• #11761: its finding was a behaviour (chapter without view) in that one tool, so I made no general rule for it.
• Old uses: I did not clean up the existing as never casts, sleeps or long comments.

To try it: on a branch, add x as never or await Bun.sleep(1) to a test file, commit, and run bun scripts/gate.ts. It prints the rule, file and line.

🤖 Generated with Claude Code

This round fixes the three review points. tsc, scripts/gate.ts and the two test files I touched pass. I did not run the full suite; the forge's check does.

• Gate wiring test: test/gate-added-refusal.test.ts builds a temp git repo under TMPDIR with a base commit and a head commit that adds a cast in a test file.

- With FORGE_CHECK_BASE set to the base, addedRefusal(cwd) returns the as never rule line, naming test/a.test.ts:2. - With no base resolving, it returns []. The repo is created with -b trunk so a main branch can't resolve by accident. - This now fails if the gate stops calling addedRefusal, if the base resolves wrong, or if the git diff arguments break.

• *Bare in comments:** COMMENT_LINE no longer matches a bare . A line counts as a comment only inside a block comment opened by / on an added line just before it. A test covers a long 1000 * rate … line, which now passes. A long * word … line inside a /** … */ block is still refused.
• ++ x as a file header: parseAdded reads +++ as a header only before the first @@ of a file, and resets on diff --git. A test covers an added ++ x line.
• Block comment limitation: with -U0, a block comment that opened before the hunk can't be seen, so a long * line in an added run that begins mid-comment is not caught. I chose to miss it rather than refuse code.
• Issue boxes 1 and 2: I left them unticked and did not edit the issue. The four findings covered by rules are as never, a real timer in a test, an unwrapped comment line, and a test importing another test. The set also holds the trailing-underscore name rule from the earlier work. I haven't listed which two review findings have no rule, since I can't tell which they are from here.

The check failed because my new gate test tripped the as never rule it was testing. Line 42 spelled out the rule's own name, ` as never , in an assertion, and the gate reads every added line. The assertion now matches only rule "a test builds the real input`, so the file no longer contains the cast text.

bun scripts/gate.ts is clean against the merge-base, and test/gate-added-refusal.test.ts passes (2 tests). The pushed commit is dec97898d. The tsc run printed nothing.

The earlier changes in this pull are unchanged:

• A temp-repo test now covers the gate's added-lines wiring (addedRefusal).
• A bare * counts as a comment only inside a block comment.
• File headers in a diff are read only outside a hunk.

Issue boxes 1 and 2 stay unticked. The two review findings that still have no rule are not named in the issue yet.

Left #11884 open: 1 unticked item remain (A lint (scripts/) or a test that fails on exactly what these comments ask for, so a reviewer need not say it again.).