[AGT-04] GitHubEventResolver — concrete event-to-claim resolution adapter

[an0mium]

Slice

Adds GitHubEventResolver — the concrete resolution adapter that converts pre-fetched GitHub webhook event payloads (PR merge, issue close, CI check/workflow) into StakeableClaim resolutions. This is sub-deliverable 2 of AGT-04; sub-deliverable 1 (StakeableClaim schema and GithubResolutionAdapterStub) is already on main.

Gating

• Default: OFF (flag: ARAGORA_PREDICTION_MARKETS_ENABLED — same gate as the existing store)
• Live queue effect: none
• Advances issue: [AGT-04] Synthetic GitHub prediction markets #6065 (AGT-04 — Synthetic GitHub prediction markets)

What this adds

Symbol	Purpose
GitHubEventPayload (frozen dataclass)	Minimal webhook event representation: event_type, action, target_ref, merged, conclusion, raw
ResolutionResult (frozen dataclass)	Outcome of a resolution attempt: claim_id, resolved, resolution_value, evidence
GitHubEventResolver	can_resolve(claim, event) → bool + resolve_from_event(claim, event) → ResolutionResult; dispatches to per-type resolvers for PR_MERGE, ISSUE_CLOSE, CI_PASS

All logic is deterministic; no live GitHub API calls. The adapter consumes pre-fetched payloads, making it testable offline and safe to compose with the existing InMemoryStakeableClaimStore.

Tests

• TestCanResolve (7 tests) — matching/mismatching event types, target refs, DEPENDENCY_RELEASE unsupported type
• TestFlagGate (2 tests) — resolve_from_event raises when flag off; can_resolve is flag-free
• TestPRMergeResolution (5 tests) — merged→YES, closed-no-merge→NO, non-terminal action, already-resolved skip, target-ref mismatch
• TestIssueCloseResolution (3 tests) — closed→YES, reopened/labeled non-terminal
• TestCIPassResolution (5 tests) — success/failure/cancelled check_run, success workflow_run, queued non-terminal
• TestResolverWithStore (2 tests) — end-to-end roundtrips through InMemoryStakeableClaimStore

Validation

• pytest tests/prediction/test_github_event_resolver.py --noconftest — 24 passed
• Syntax: all three files parse without errors (ast.parse)
• Ruff / mypy: no new dependencies introduced; frozen dataclasses with explicit type annotations

Out of scope

• Live GitHub API polling (sub-deliverable 3; will use PyGithub or REST client behind an adapter interface)
• Durable JSONL/PostgreSQL claim store (covered by agt-04-jsonl-claim-store branch)
• AGT-05 reputation wiring — ResolutionResult is the feed point; wiring waits for AGT-05

---

Generated by Claude Code

[github-actions]

Aragora Code Review

Advisory-only review. No issues found.

[scarmani]

Claude independent model review

Reviewer: claude (anthropic) — independent adversarial model review via the Aragora Claude reviewer, grounded on the exact PR head.
Head: 3216f48 (3216f48), committed 2026-06-30T16:19:14Z.
PR: #8519.
Model family: claude

Verdict: PASS

• [P3] aragora/prediction/github_event_resolver.py _resolve_ci_pass (~L191-206): a single check_run with conclusion="success" resolves the whole CI_PASS claim as pass. A repo has many check_runs; one passing check ≠ all CI green. Correctness depends entirely on the caller pre-aggregating and setting target_ref granularity — worth a docstring caveat, since check_run and workflow_run have very different aggregation semantics yet share one resolver.
• [P3] github_event_resolver.py resolve_from_event (~L96): no claim.expiry check — an event arriving after expiry still resolves an expired market. If expiry isn't enforced elsewhere, this lets stale events settle closed markets.
• [P3] github_event_resolver.py _resolve_issue_close (~L178-189): any issues/closed event yields resolution_value=True, ignoring state_reason (not_planned). For a claim meaning "issue gets fixed," a not-planned close is a false-positive YES.
• [P3] docs/METRICS.md: trailing whitespace appended after "...regardless of threshold." (L52), and the "Unique permission strings" command now renders \\1 (was \1) inside a code span — copy-pasting the doc'd sed gives a broken backreference. Both look like regeneration artifacts; harmless but cosmetic regressions in a generated file.

Note (non-finding): the metrics deprecated entry sets test_file_count: 0 despite tests/metrics/ containing ~10 files, but this matches the existing operations deprecated convention (also 0), so it appears intentional rather than a defect.

dogfood: yes