feat(miner): contribution-profile-extract.ts never checks AGENTS.md/CLAUDE.md despite the schema defining agent_docs as a signal source

[JSONbored]

Context

packages/loopover-miner/lib/contribution-profile.ts (#6795, the ContributionProfile schema) defines ContributionSignalSource as one of four values:

export type ContributionSignalSource =
  "labels" | "contributing_md" | "pr_template" | "agent_docs";

with the design comment explicitly noting: "#6794 found the primary source differs per repo — some state rules only in agent docs, some only in labels." CONTRIBUTION_SIGNAL_SOURCES (the exported runtime vocabulary) freezes the same four values.

packages/loopover-miner/lib/contribution-profile-extract.ts (#6796, the extractor that actually populates a ContributionProfile) only ever fetches two of the four: repo labels (fetchRepoLabels, source "labels") and CONTRIBUTING.md/.github/CONTRIBUTING.md (fetchContributing, source "contributing_md"). It never fetches an AI-agent-facing doc (AGENTS.md, CLAUDE.md) and the extractor never emits a "agent_docs" provenance entry — despite the schema being explicitly designed around that source existing and #6794's own research (docs/ams-contribution-signal-inventory.md) motivating it.

AGENTS.md/CLAUDE.md are full contributor-instruction documents (the same size/nature class as CONTRIBUTING.md, not a short signpost), so the existing extractPrBody/signpost-size heuristic (CONTRIBUTING_SIGNPOST_MAX_BYTES) is directly reusable without inventing a new sizing rule — unlike a PR template (typically short by design), which is why this issue is scoped to agent_docs only, not pr_template.

Requirements

⚠️ Required pattern. Mirror fetchContributing's exact shape (probe a fixed list of repo-root paths via the GitHub contents API, decode with the existing decodeContents, return the first hit or null) for a new fetchAgentDocs that probes AGENTS.md then CLAUDE.md at the repo root, in that order.

• Add fetchAgentDocs(base, target, headers, fetchImpl, sleepFn): Promise<string | null> that probes AGENTS.md then CLAUDE.md (repo root only — do not add .github/-prefixed variants; neither of these files follows CONTRIBUTING.md's .github/ convention in real repos) using the same getJson/decodeContents helpers fetchContributing already uses.
• extractContributionProfile must call fetchAgentDocs alongside the existing fetchContributing call.
• Precedence rule (exact, not left to interpretation): CONTRIBUTING.md stays authoritative when it yields a non-absent prBody rule (unchanged behavior — zero regression risk for every repo that already has a real CONTRIBUTING.md). Only when extractPrBody(contributing) resolves to { confidence: "absent" } (no CONTRIBUTING.md at all) does the agent-doc candidate get evaluated, through the exact same extractPrBody logic (same CONTRIBUTING_SIGNPOST_MAX_BYTES floor, same LINKED_ISSUE_TERMS match), with its resulting provenance source tagged "agent_docs" instead of "contributing_md".
• Do not add a new field to ContributionProfile — the result still lands in the existing prBody: ContributionSignalRule<ContributionPrBodyRequirements> slot, exactly as CONTRIBUTING.md-derived results already do; only the provenance source differs.

Deliverables

• fetchAgentDocs added to contribution-profile-extract.ts, mirroring fetchContributing's structure.
• extractContributionProfile wired to fall back to the agent-docs source per the precedence rule above.
• extractPrBody (or a thin wrapper) tags provenance source: "agent_docs" for an agent-doc-derived result, matching CONTRIBUTION_SIGNAL_SOURCES's existing vocabulary.

Test Coverage Requirements

packages/loopover-miner/** is not Codecov-gated (only src/** is), but npm run test:ci must stay green with full branch coverage on the new logic in test/unit/miner-contribution-profile-cache.test.ts or a new/existing extract-focused test file — cover: (1) CONTRIBUTING.md present and non-absent (agent docs never fetched or ignored), (2) no CONTRIBUTING.md but AGENTS.md present and large enough (provenance tagged agent_docs), (3) no CONTRIBUTING.md, AGENTS.md missing but CLAUDE.md present, (4) neither CONTRIBUTING.md nor any agent doc present (stays absent), and (5) an agent doc present but below the signpost-size floor (stays unknown, matching extractPrBody's existing signpost branch).

Expected Outcome

A repo whose contribution rules live only in AGENTS.md/CLAUDE.md (no CONTRIBUTING.md) now produces a real, provenance-tagged prBody signal instead of the schema's own agent_docs source sitting permanently unused — the exact gap #6794's own research flagged as a real-world pattern.

Links & Resources

• packages/loopover-miner/lib/contribution-profile-extract.ts (the file to change — fetchContributing, extractPrBody, extractContributionProfile)
• packages/loopover-miner/lib/contribution-profile.ts (ContributionSignalSource, CONTRIBUTION_SIGNAL_SOURCES — the schema this closes the gap against)
• epic: AMS per-repo contribution-profile discovery (labels, docs, agent-facing eligibility signals) #6793 (epic), AMS contribution-profile: inventory real-world eligibility signals across a diverse repo sample #6794 (the real-repo signal inventory that named agent_docs), AMS contribution-profile: design the ContributionProfile schema and caching shape #6795 (schema), AMS contribution-profile: implement generic label/docs/agent-file extraction #6796 (original extractor issue)

[linear-code]

JSO-5781