feat: drop-in observability kit with audit comparison and behavioral signals

[mnkiefer]

• Adds the foundation of a drop-in observability kit that consumers can adopt via gh aw add or gh aw add-wizard to get recurring repo-wide reports on workflow health, drift, and risk.
• This includes:
  • Audit comparison: finds a successful baseline run, computes deltas (turns, posture, blocked requests, MCP failures), and classifies changes as stable, changed, or risky.
  • Task domain and behavior fingerprinting: infers run type and builds a compact execution profile for cross-run comparison.
  • Agentic assessments: flags resource-heavy, weakly-controlled, or unnecessarily agentic runs.
  • Observability insights: per-run and fleet-level signals surfaced in audit and logs JSON output.
  • Deterministic execution episodes and lineage edges: gh aw logs now emits structured episodes[] and edges[] data so orchestrated runs can be grouped and rolled up consistently.
• The bundled workflow now follows the enterprise-style reporting model of one full discussion report per run, plus at most one escalation issue when the findings warrant action.
• Opt-in job summary remains available via frontmatter for in-run visibility:

observability:
  job-summary: on

[pelikhan]

@copilot review comments, lint go

[Copilot]

@copilot review comments, lint go

Addressed all review comments in commit 5abdb23:

1. countBlockedRequests() — now sums across all gatewayEventPaths files instead of returning after the first match.
2. Audit comparison classification — added reason codes for beneficial changes (turns_decrease, blocked_requests_decrease, mcp_failures_resolved) and explicit switch cases so runs that differ from baseline are no longer classified as stable.
3. renderObservabilityInsights — added critical and low severity icons.
4. min_blocked_requests evidence — now includes both threshold and actual count, e.g. blocked_requests_gte=5 actual=7.
5. warn action — added explicit case "warn" using FormatWarningMessage so it renders distinctly from info messages.

[github-actions]

Hey @mnkiefer 👋 — great work on this observability overhaul! Strengthening gh aw audit with baseline comparisons, behavioral fingerprinting, and agentic assessments, while also levelling up gh aw logs with trend-aware reporting, is a really cohesive and valuable addition to the toolkit. The drop-in observability workflow kit is a particularly nice touch for teams who want to operationalize monitoring across repos without starting from scratch.

The PR is well-structured, thoroughly described, and backed by solid test coverage across all the new modules. It looks ready for maintainer review. 🟢

Generated by Contribution Check · ◷