feat(skills): add mssql-performance-review orchestrator (tier 1 — agentic core)

[vanterx]

Summary

Adds mssql-performance-review — an agentic offline orchestrator that routes mixed SQL Server artifacts (or symptom descriptions) to the 15 specialised review skills, runs an adversarial root-cause check, and emits a single consolidated report.

This is tier 1 of the v4 roadmap (see backlog/performance-review-orchestrator-plan-v4.md in local backlog). Tier 1 delivers the trustworthy agentic core. Tier 2 (multi-model routing, DAG, domain memory, follow-up Q&A) and tier 3 (capture-bundle generator, verification checklist, baseline-diff loop) ship in subsequent PRs.

What's new

Primitive	What it does
Evidence chain	Every Critical finding cites at least 3 check IDs from at least 2 skills, with source artifact + observed value + threshold. Reproducible from input set.
Risk-aware recommendations	Every fix carries action, effort, blocking window, risk class, side effects, exact rollback, and post-deployment verification capture
Adversarial root-cause check	After the primary hypothesis is identified, a deliberate pass tries to disprove it. Strong contradiction escalates the alternative.
Confidence-driven early termination	Stops dispatching once 3+ skills converge HIGH with no contradiction. --exhaustive to override.

Strictly offline

The orchestrator never opens a connection to SQL Server. All execution against the database is the user's action. Works in air-gapped, regulated, or third-party-DBA scenarios.

Files

New:

• skills/mssql-performance-review/SKILL.md (304 lines)
• skills/mssql-performance-review/references/ — README, check-explanations, evidence-schema, risk-rubric, adversarial-prompts
• skills/mssql-performance-review/evals/evals.json — 4 test prompts (artifact-first, symptom-first, adversarial, baseline-diff)
• skills/mssql-performance-review/scripts/.gitkeep, assets/.gitkeep
• example/mssql-performance-review/mixed-artifacts/ — .sql + .sqlplan + stats + wait stats
• example/mssql-performance-review/mixed-artifacts-analysis.md — reference unified report
• example/mssql-performance-review/symptom-first-analysis.md — symptom-driven triage walkthrough

Updated:

• All 15 existing SKILL.md files (Companion Skills section)
• CLAUDE.md — Key Files, Skills tables, prefix map (dispatcher exemption noted)
• README.md — install line, full ## mssql-performance-review section
• PERFORMANCE_TUNING_GUIDE.md — Skills at a Glance (+ orchestrator row, + missing hadr-health-review row), new "I have a pile of mixed artifacts" scenario
• LLM_COST_ESTIMATION.md — orchestrator size row (~7,500 tokens)

Test plan

• bash scripts/verify-docs.sh — 31 PASS, 0 WARN, 0 FAIL
• Orchestrator SKILL.md under 1000-line cap (304 lines)
• Every Critical finding in the example analysis cites at least 3 check IDs from at least 2 skills
• Every recommendation in the example analysis has action + effort + window + risk + side effects + rollback + verification
• Adversarial check section present in the example analysis with no_contradiction result
• Recommendation Conflicts section explicit (states "None detected" rather than omitting)
• Skills Skipped table explicit (10 skills skipped with reasons, only 5 ran on the example set)
• Symptom-first example clearly states the orchestrator will not contact SQL Server
• All 15 existing skills now reference mssql-performance-review as a Companion Skill

Roadmap

Tier	Branch	Adds
tier 1 (this PR)	feat/perf-review-orchestrator-v4-tier1	Evidence chain, risk-aware fixes, adversarial check, early termination
tier 2	feat/perf-review-orchestrator-v4-tier2	Multi-model routing, skill-graph DAG, domain memory, follow-up Q&A
tier 3	feat/perf-review-orchestrator-v4-tier3	Capture-bundle generator, verification checklist, baseline-diff feedback loop

🤖 Generated with Claude Code

[vanterx]

Fixed the off-by-one: two locations in skills/mssql-performance-review/SKILL.md said "14 specialised review skills" but the Companion Skills section lists 15. Changed both to "15" (line 15 in Purpose and line 199 in Output Format). verify-docs.sh still passes 31/31.