fix(agentic-engineering): classify current-head CI safely

[devantler]

🤖 Generated by the Agentic Engineer

Description

The generic portfolio surveyor can misclassify current-head GitHub Actions health because its procedure does not provide one deterministic, fail-closed classifier contract for pagination, recovery state, managed-run identity, and equal timestamps. A consumer repair proved the behavior against live paginated payloads, but the portable logic belongs in this plugin before that consumer can roll it out.

Related: devantler-tech/monorepo#2173, devantler-tech/monorepo#2436, and #95.

Steps to Reproduce

1. Feed default-branch run pages where an older failure is followed by a non-success retry, or where independent dynamic/dependabot jobs share one workflow id.
2. Split the payload across raw gh api --paginate page documents or a --slurp envelope array.
3. Include two decisive runs with the same created_at second but different run ids.
4. Let a later page request fail after earlier pages were emitted.

Expected Behavior

The surveyor evaluates all pages globally, preserves a known failure until a successful run clears it, keeps managed jobs independent, breaks timestamp ties by run id, preserves fields needed for managed-run routing, and reports query failure as unknown rather than green.

Actual Behavior

The current generic prose selects only by workflow id and timestamp, treats non-red latest states as clear, and does not bind classifier success to successful completion of the paginated producer. The consumer-local repair cannot be canonical under the definition-routing contract.

Environment

• Plugin: agentic-engineering 4.3.3 / reviewed main 564a6a0f
• Consumer: devantler-tech/monorepo PR #2436

Acceptance Criteria

• Canonical plugin agent/procedure owns the full classifier contract and any executable helper.
• Fixture tests RED-prove raw and slurped pagination, producer failure propagation, success-only recovery, managed logical identity, deterministic run-id tie-breaking, and identity-field output.
• The generic agent fails closed to unknown on incomplete API pagination.
• The consumer can remove or reduce its duplicated classifier/prose after advancing the reviewed plugin pin.
• Plugin validation and current-head review are green.

[coderabbitai]

🔗 Related PRs

#87 - fix(agentic-engineering): retire human promotion gate [merged]
#120 - security(agentic-engineering): make the surveyor's read-only boundary executable [merged]
#130 - feat(agentic-engineering): measure Improver effectiveness [merged]
#132 - feat(agent-improver): research on evidence-clean runs [merged]
#134 - chore(deps): update agent skills [merged]

---

📝 Issue Planner

_{Check the box below or use the @coderabbitai plan command to generate an implementation plan and prompts that you can use with your favorite coding assistant.}

• Create Plan

---

🧪 Issue enrichment is currently in open beta.

You can configure auto-planning by selecting labels in the issue_enrichment configuration.

To disable automatic issue enrichment, add the following to your .coderabbit.yaml:

issue_enrichment:
  auto_enrich:
    enabled: false

💬 Have feedback or questions? Drop into our discord!