feat(review): add /review:quality-gate code-design mode - #8388

The phase-boundary inputs are passed explicitly by an orchestrator, so the
description needs no new trigger phrase.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…-gate-phase-boundary-inputs

…d grants

Restore the argument-hint to the mode set so it fits the 100-character
budget and matches the lead of the Arguments line; the phase-boundary
flags stay documented in the Arguments line and context/phase-boundary.md.

Remove the allowed-tools entries that put a wildcard before the git
subcommand (git -C * ...). A wildcard there also approves options such as
-c core.fsmonitor=<cmd>, which runs a program, in a skill that reviews
untrusted diffs. Worktree-targeted commands go through normal permission
handling.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Under --base or --worktree, code mode skips the pr-review-toolkit path
for its code-reviewer fallback, and downstream mode skips the remote
downstream_probe layer and treats the floor as report. Under
--report-only the size-crossing loop reads git diff -z instead of a
written path file. Name the modes --quality-bar applies to, pass spec
text to the spec reviewer, and fail the eval grader only on a finding
in admin.js, not on a note that it predates the base.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

- Name pr and close-out in the --quality-bar row; brief reviewers with
  each mode's own required inputs; restatement dispatches under --base.
- Under --report-only the standards ladder takes its non-interactive rung.
- code skips size crossings under --worktree; define the report-only base.
- Step 0 skips the PR-exists signal with --base; drop restated upstream.
- Eval: fail on any Write, add a no-Edit grader.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

code-reviewer, security-reviewer and architecture-guardian run only the
change-set commands their brief supplies (worktree path and base SHA),
with no gh lookup, fetch or ls-remote; without one, step 2 is unchanged.
phase-boundary.md says the brief replaces that step. criteria mode under
--base or --worktree dispatches a fresh-context review:brief-reviewer
instead of reviewing inline.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Restatement mode now dispatches its batch reviewer under --worktree as
well as --base, so a worktree-only run no longer reviews inline from the
invoking checkout.

The phase-boundary reviewer brief names the worktree's absolute path as
the file-read root, and code-reviewer, security-reviewer and
architecture-guardian resolve every Read, Grep, Glob and listed
untracked path under a supplied worktree. The advisory line no longer
reads as if per-slice passes the plan's rationale.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Replace the per-command worktree anchoring with one rule: when the brief
supplies a worktree, every command and tool call in the review targets it
(git via git -C, Read/Grep/Glob, untracked paths, and any other command
such as a dependency audit), never the invoking checkout. Stated
identically in code-reviewer, security-reviewer and architecture-guardian,
and in the quality-gate phase-boundary brief for every dispatched
reviewer, review:brief-reviewer included. The downstream probe's
repository root is the worktree path under --worktree.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…puts

…alternation

typos read the [Ss]earch character class as the misspelling "earch" and
failed lint-repo. (?:search|Search) matches the same strings.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

The mode loads the consumer's review criteria through the standards
resolution ladder, falls back to small bundled criteria (Duplicated Code,
Shotgun Surgery, Speculative Generality), and dispatches a fresh-context
brief-reviewer. Repetition is CRITICAL only when the diff introduces an
unsynced copy in code, naming the fact and the diverging change; Markdown,
cascades and the prototype bar stay SUGGESTION. A contradicting consumer
standard wins and the report names the overridden default; with no
quality bar the review assumes production and says so.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Five plugin-eval cases cover the mode's severity rules: an introduced,
unsynced code copy is CRITICAL with the production assumption named; a
declared-registry entry and a file move are not CRITICAL; Markdown
repetition and the prototype bar stay SUGGESTION; a contradicting
docs/standards file produces the override line.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

The no-critical and repetition-is-suggestion graders now fail only on a
rated finding (a severity table cell, a line that starts with the
severity, a (CRITICAL) or [CRITICAL] label, or Severity: CRITICAL) and
pass a negated mention such as "Not CRITICAL: setup.md is Markdown".
The cascade grader requires the CRITICAL finding to name
FREE_SHIPPING_THRESHOLD or 50, not the word threshold. code-design.md
states that its repetition rule overrides severity.md's "code
duplication" IMPORTANT example.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…e-code-design-mode