feat(observability): /sprint --health per-sprint outcome aggregator

[hx2ryu]

Summary

• Adds scripts/lib/sprint-health.py — read-only deep-dive on a single sprint's outcomes (verdicts, fix-loop counts, failure-mode guard activity, contributed patterns + curator transitions)
• Workflow integration: /sprint <sprint-id> --health mirrors the existing /sprint <sprint-id> --status shape — --status shows progress, --health shows outcomes
• Companion to sprint-fleet-status.py (cross-sprint summary). With this PR, the three observability surfaces line up:

Surface	Question	Tool
Per-sprint progress	Where is the sprint right now?	/sprint <id> --status
Per-sprint outcomes	How is the sprint going?	/sprint <id> --health (new)
Cross-sprint summary	What's running across my sprints?	/sprint --status (fleet)

Data sources

The aggregator reads five sources, all read-only:

1. runs/sprint/<id>/contracts/group-*.md — group count
2. runs/sprint/<id>/evaluations/group-*.md and group-*.adversarial.md — standard + adversarial verdicts
3. runs/sprint/<id>/logs/evaluator.jsonl — eval rounds per group (proxy for fix-loop count)
4. Project-scoped logs/guards.jsonl filtered by sprint_id — drift / self_deception / context / regression buckets
5. .zachflow/kb/learning/patterns/*.yaml filtered by source_sprint + project-scoped logs/curator.jsonl for transitions on those patterns

Output formats: markdown (default, post-to-PR friendly) or JSON (for piping into external trend tracking).

Files

Type	Path
New	scripts/lib/sprint-health.py (315 LOC)
New	tests/sprint-health-test.sh (6 cases — empty / mixed verdicts / eval rounds / guards filter / patterns filter / JSON parse)
Mod	workflows/sprint/SKILL.md (CLI reference table + --health)
Mod	workflows/sprint/phase-modes.md (new "## --health Mode" section)
Mod	docs/roadmap.md (v1.x entry marked shipped)

Test plan

• New sprint-health-test.sh passes (6 cases)
• All existing tests still pass (11 suites: 4 guards + curator + migrate + schema + kb-smoke + jsonl-chain + bump-rubric + fleet-status)
• 3-OS CI matrix green (cygpath compat already in place for jsonl/yaml reads)
• First real sprint run that hits multiple groups — eyeball that the rendered markdown matches the actual run state

Design notes

• No snapshots persisted by default — the script computes the snapshot on demand. Trend lines (last N snapshots, alerts) deliberately deferred so this PR stays tight; the JSON output is the contract for any future trend-tracking layer
• sprint_id is the filter key everywhere — guards.jsonl events tag their sprint; curator.jsonl events don't (curator is kb-scoped) so we filter via the pattern's source_sprint instead. Both filters trivially exclude other sprints' data
• Fix-loop count is a proxy, not a count — derived from evaluator.jsonl rounds (= rounds - 1). If the workflow caller adds a dedicated fix_loop_started event later, the script can switch sources without changing the API

🤖 Generated with Claude Code