Add dual-pool-review: adversarial review methodology with dual-pool rotation

[YuhaoLin2005]

Problem

The marketplace has excellent QA tools (qa-expert for testing infrastructure, marketplace-health-check for repo audits, code-reviewer for single-pass review), but no adversarial review methodology that prevents both reviewer blind spots and confirmation bias. A single reviewer — no matter how skilled — has domains they don't know, failure modes they're blind to, and assumptions they share with the author. Running the same reviewer multiple times doesn't help: they miss the same things every pass.

Solution

dual-pool-review fills this gap with a structured multi-round adversarial review system:

• Dual-pool rotation: Fixed pool (4-6 curated personas that learn your patterns) + Random pool (1 fresh reviewer per round, recruited via web search with verifiable expertise)
• Quote-First Rule: Every reviewer extracts verbatim quotes from web search BEFORE seeing the target — findings without traceable quotes are rejected
• Symmetric Burden: Reporting zero findings is as expensive as reporting one — must cite 3+ quotes the output successfully satisfies (prevents lazy "LGTM")
• Evidence Table: Mandatory per-round table linking every finding to a specific, searchable source quote — the non-negotiable quality gate
• Adaptive Stop Check: Each round asks "did this find issues the previous round missed?" — self-terminates when value stops
• Concurrence Promotion with Independence Requirement: 2+ reviewers finding the same issue promotes severity, but at least one must be from the random pool (prevents model-bias corroboration)
• Lightweight Path: Single-file low-risk changes get 1 reviewer + web search; escalate to full dual-pool only if CRITICAL/WARNING found

How it complements existing marketplace skills

Existing Skill	dual-pool-review adds
qa-expert	Adversarial methodology for test plan review, multi-persona blind-spot coverage
code-reviewer	Multi-round rotation with fresh perspectives, evidence-backed findings
marketplace-health-check	Same adversarial spirit, but reusable for any output (code, docs, config, identity)
fact-checker	Structural review methodology with rotating expertise lenses

Changes

• dual-pool-review/SKILL.md — 152 lines, follows daymade format conventions (YAML frontmatter with keywords, ## When to Use This Skill section, imperative style)
• .claude-plugin/marketplace.json — Added plugin entry (category: productivity, 14 keywords), bumped version to 1.79.0
• README.md / README.zh-CN.md — Added listing as skill feat: Add standalone zero-dependency validator to skill-reviewer #83 with install instructions, use cases, and key features

Prior validation

The methodology was reviewed by 3 adversarial expert personas (Torvalds, Carmack, Cagan) and all 6 findings were incorporated before this submission.

[YuhaoLin2005]

✅ Validation Results

Ran the full marketplace validation pipeline from CLAUDE.md §Validation:

quick_validate.py — PASS ✅

Skill is valid!

Had to remove keywords from SKILL.md YAML frontmatter — quick_validate rejects it as an invalid frontmatter field. The allowed fields are: agent, allowed-tools, argument-hint, compatibility, context, description, disable-model-invocation, hooks, license, metadata, model, name, user-invocable. Keywords live in marketplace.json and README entries only.

check_marketplace.sh — PASS ✅ (2 warn-only items)

Check	Result
[1/4] JSON syntax	PASS — marketplace.json is valid JSON
[2/4] claude plugin validate	SKIP — claude CLI not installed (Windows)
[3/4] source+skills resolution	PASS — all plugin source paths resolve to real SKILL.md
[4/4] reverse sync	WARN — 2 pre-existing unregistered skills (daymade-claude-code/claude-migrate-memory-to-doc, design-style-picker), not introduced by this PR

check_doc_skill_lists.py — PASS ✅

Authoritative skills in marketplace.json: 81
CLAUDE.md: 81 listed — OK
README.md: 81 listed — OK
README.zh-CN.md: 81 listed — OK
README.md version badge: 1.79.0 — OK
README.zh-CN.md version badge: 1.79.0 — OK
Result: all doc skill lists are in sync with marketplace.json.

security_scan.py — BLOCKED ⚠️

gitleaks not available on Windows. This skill is a pure methodology (SKILL.md only, no scripts, no dependencies, no bundled resources), so the attack surface is minimal.

---

Summary: 4/4 applicable checks pass. The 2 warnings are pre-existing issues in the repo, not introduced by this PR.