feat(diff): semantic diff between agent versions

[code-saksham-hash]

Adds opengap diff [from] [to], closing #4.

Compares two agent snapshots — git refs (resolved via git archive) or plain directories — and reports what actually changed instead of a raw text diff: agent.yaml fields, SOUL.md/RULES.md/DUTIES.md size deltas, skills/tools/workflows added/removed/modified, compliance changes (flags risk tier escalation), hooks (flags a newly-added fail_open: false entry), and which files under memory/ changed. --json for scripting/CI.

A few notes on how it's built:

• Either side of the comparison can be a git ref or a directory. If it's not an existing directory, it's treated as a ref and materialized into a temp dir via git rev-parse + git archive — no working-tree checkout, so no lock conflicts, and it still works when --dir isn't a git repo at all (both-directories case).
• A single A..B argument works too, split on the literal ...
• Bare opengap diff with no args compares HEAD against the working directory, same idea as plain git diff.
• SOUL.md is diffed as presence + line-count delta only, no NLP — per your comment on the issue.
• agent.yaml and compliance are diffed generically (flatten + compare) instead of hand-listing every field, so new schema fields get picked up without touching this code.
• risk_tier escalation only fires on a strict increase (low < medium < high < critical), not on a downgrade.
• SOD conflict pairs compare order-insensitively ([a, b] == [b, a]).

Tested with npm run build + npm test (75/75, 24 new tests in agent-diff.test.ts and git-ref.test.ts), plus manual runs against examples/standard vs examples/full and a throwaway two-commit repo — ref diffing, A..B syntax, bare diff, an invalid ref, and confirming temp dirs actually get cleaned up.

Sample output (`opengap diff ./examples/standard ./examples/full`)

gitagent diff: ./examples/standard → ./examples/full
────────────────────────────────────────────────────────────
  Identity (SOUL.md): +14 added, -10 removed
  Manifest.description: Automated code review agent with best-practice enforcement → Financial compliance analysis agent for FINRA and Federal Reserve regulated environments
  Manifest.name: code-review-agent → compliance-analyst
  ...
  Rules (RULES.md): +69 added, -15 removed
  Duties (DUTIES.md): added; conflict added: [analyst, reviewer]; conflict added: [analyst, auditor]; conflict added: [reviewer, auditor]
  Skills: +2 added (document-review, regulatory-analysis), -1 removed (code-review)
  Tools: +2 added (generate-report, search-regulations), -2 removed (complexity-analysis, lint-check)
  Workflows: +1 added (regulatory-review), 0 removed
! Compliance.risk_tier: unset → high  ⚠ tier escalation
  Compliance.frameworks: +finra, federal_reserve, sec, cfpb
  ...
! Hooks: +on_session_start:scripts/load-compliance-context.sh (fail_open=false), +pre_tool_use:scripts/audit-tool-call.sh (fail_open=false), ...  ⚠ enforcement added
! Memory: MEMORY.md, context.md, daily-log/2026-02-18.md, daily-log/2026-02-20.md, key-decisions.md, memory.yaml changed  ⚠ review needed
────────────────────────────────────────────────────────────

Known gaps: nothing tests the Commander wiring directly (same as the other files in src/commands/, the logic underneath is what's covered), and array-valued compliance fields like escalation_triggers/handoffs are diffed as whole values rather than per-item add/remove. Can break those out further if you want it more granular.