feat(miner-discovery): per-repo AI-policy map — parse CONTRIBUTING/AI-USAGE docs for AI-PR bans

[JSONbored]

Before a miner spends any effort on a repo, it must check whether that repo bans AI-generated PRs. Implement a deterministic, best-effort text scanner over a repo's CONTRIBUTING.md / AI-USAGE.md-shaped documents (fetched as public metadata — raw file content via the GitHub contents API, never a clone) that returns a simple allow/deny verdict, so downstream discovery can hard-skip banned repos entirely. Conservative-by-construction: ambiguous or absent policy text defaults to "allowed" (never invent a ban), while an explicit ban phrase always wins.

Deliverables

• packages/gittensory-engine/lib/ai-policy-map.ts — export type AiPolicyVerdict = { allowed: boolean; matchedPhrase: string | null; source: "CONTRIBUTING.md" | "AI-USAGE.md" | "none" } and export function scanAiPolicyText(content: string | null, source: AiPolicyVerdict["source"]): AiPolicyVerdict — a case-insensitive phrase scanner for a small, explicit ban-phrase list (e.g. /no ai[- ]generated (pull requests|prs|contributions)/i, /ai[- ]generated (prs?|contributions?) (are|will be) (banned|rejected|not accepted)/i, /do not (use|submit) ai[- ](written|generated) code/i) — matching returns allowed: false; no match returns allowed: true, matchedPhrase: null.
• export function resolveAiPolicyVerdict(docs: { contributing: string | null; aiUsage: string | null }): AiPolicyVerdict — scans AI-USAGE.md first (more specific), falls back to CONTRIBUTING.md, defaults to { allowed: true, matchedPhrase: null, source: "none" } when both are absent.
• Unit tests: each ban-phrase pattern individually matches on a realistic sentence fixture; a document with no ban language returns allowed: true; a document present but empty-string returns allowed: true; AI-USAGE.md taking precedence over a conflicting CONTRIBUTING.md when both are supplied; both docs absent (null, null) returns the source: "none" default.
• A doc-comment explicitly stating the conservative bias: false-negatives (missing a real ban) are acceptable and expected to shrink over time as phrases are added; false-positives (banning a repo that doesn't actually ban AI) are the worse failure mode and phrases must stay literal/explicit, never a fuzzy heuristic.

References

• src/signals/focus-manifest.ts isFocusManifestPublicSafe (line 300) — sibling pattern of a small explicit-phrase text scanner used defensively (public-safety scanning) as a style reference
• CONTRIBUTING.md (repo root, 17804 bytes) — an example of the document shape being scanned (not itself scanned by this code — it's gittensory's own contributing doc, used only as a realistic test fixture)
• new path: packages/gittensory-engine/lib/ai-policy-map.ts (this issue creates it)

[linear-code]

JSO-2402