feat: port ADLC guardrails, personas, and templates from harbor

[sauravaman02]

Summary

Ports the production-proven ADLC patterns from the harbor-pricing-service .cursor/ setup into MiniGit's AI SDLC factory. All additions are generic — no product-specific or internal references.

What's added

4 new skills (keyboard-level guardrails):

• diff-preview — per-file diff approval before edits; never batch-apply
• commit-workflow — keyword gates ("commit" / "push"); explicit-path staging
• local-validation — pre-declare expected outcomes, run make check, compare
• investigator — evidence-first debug ladder (reproduce → tests → code → data → env) + RCA template

4 mode activation rules (.cursor/rules/*.mdc):

• architect-mode — triggers on "use architect", loads idea/spec/plan skills
• developer-mode — triggers on "use developer", loads impl + safety skills
• reviewer-mode — triggers on "use reviewer", loads code-review skill
• investigator-mode — triggers on "investigate", loads investigator skill

5 templates (Skills/_templates/):

• handoff.md — cross-session state bus between agents (works without Jira)
• design-doc.md — 10-section architecture template with Decision Log + Open Questions
• task-breakdown.md — phased task list with complexity and checkpoints
• phase-report.md — phase completion receipt (Completed / Validation / Deferred / Ops)
• rca.md — Root Cause Analysis document template

What's enhanced

• code-review-and-quality: P0–P3 severity taxonomy (BLOCKER / SHOULD FIX / SUGGESTION / NOTE) + APPROVE / REQUEST CHANGES / NEEDS DISCUSSION verdicts + structured output format
• All 12 skills: "What You Are NOT" role boundaries to prevent role bleed
• incremental-implementation: References to diff-preview, commit-workflow, and local-validation skills

What's fixed

• agent_space/README.md created — make eval-skills was broken without it
• De-identified all internal references (Jira project keys, org URLs, personal paths) for public repo safety
• repos/repos.json — replaced absolute local path with relative path

Eval harness

Extended run_skill_evals.py to validate new skills, templates, rules, and symlinks:

OK: skill registry evals passed
  skills=12 rubrics=7 templates=7 rules=4

Test plan

• make eval-skills passes (12 skills, 7 rubrics, 7 templates, 4 rules)
• No internal/org references (rg -i "redhat|BLRID|/home/krastogi" → zero matches)
• All .cursor/skills/ symlinks resolve correctly
• All skill files have YAML frontmatter + Deterministic section + "What You Are NOT"
• No source code (src/, tests/) was modified

Made with Cursor