Why

author_association/reviewer_association are snapshotted at ingest and never refreshed, so role-based scoring (issue-bonus tier, PR-drop gate, maintainer review gate) reads stale roles. PR #190 patched this by rewriting the stored columns across 4 tables hourly — but that's the wrong layer: write-amplification, an hour staleness window, registered+installed-only coverage, and it corrupts the event-time meaning of the stored snapshot.

Separately, the per-row label/review association resolution went through contributor_repo_roles — a plain view re-derived per row (EXPLAIN ANALYZE: ~30ms each, a sort + DISTINCT ON over ~8k rows across pull_requests/issues/reviews/comments, 1–2× per PR). That's the ~70ms/row that times the validator out on high-volume miners.

Both collapse into one primitive: a live, indexed maintainers table.

What

• New table maintainers (repo_full_name, github_id, login, association, refreshed_at), PK (repo_full_name, github_id). repo_full_name stored lowercased so every read joins m.repo_full_name = LOWER(<src>.repo_full_name) and still uses the PK index.
• Replace feat(maintainer): reconcile stored author_association to live GitHub roles #190's column rewrite with MaintainerPopulateService: reuses the same fetcher (fetchRepoCollaborators + fetchOrgMembers) and safety rules (fetch-before-write, fail-closed-per-repo, skip-on-empty), but atomically upserts maintainers per repo (transactioned) instead of mutating 4 tables. Hourly @Cron + once on OnModuleInit.
• Repoint off contributor_repo_roles onto maintainers: pr_labels_by_actor/issue_labels_by_actor (actor_association), pr_linked_issues (issue_author_association), and pr_review_summary (the maintainer CHANGES_REQUESTED filter). Indexed lookup instead of a per-row re-derivation → ~10–20× on the hot path.
• Serve-time author resolution in miners.service.ts via COALESCE(maintainers.association, stored) on PR/issue authors — resolves onto the existing author_association field, no new payload field, no gittensor change. Safe because the validator only ever tests … in MAINTAINER_ASSOCIATIONS, so maintainers-only (present → role, absent → NULL/stored) is lossless.
• Repoint /repos/:repo/maintainers to the table. Stored association columns marked as ingest snapshots; contributor_repo_roles left in place (no remaining hot-path consumers).

Validation

Loaded the full packages/db schema (incl. 11_maintainers + repointed views 21/22/24/25) into a throwaway Postgres and seeded a stale-snapshot case. All pass:

npm run build, lint, format:check clean in packages/das. No new deps (no lockfile change). No test files (team rule).

SQL migration checklist (prod) — run in this order

packages/db/*.sql only auto-runs on fresh Docker volumes. Prod has data, so run these manually against the mirror DB. Ordering matters: the table must be populated before the views are swapped, or views 24/25 emit NULL actor_association (every label actor reads as non-maintainer) during the gap.

1. Create the table (additive, safe anytime) — body of packages/db/11_maintainers.sql:

CREATE TABLE IF NOT EXISTS maintainers (
    repo_full_name  VARCHAR(255) NOT NULL,
    github_id       VARCHAR(255) NOT NULL,
    login           VARCHAR(255),
    association     VARCHAR(20)  NOT NULL,
    refreshed_at    TIMESTAMPTZ  NOT NULL DEFAULT NOW(),
    PRIMARY KEY (repo_full_name, github_id)
);

2. Deploy the app build. Safe with an empty table: serve-time COALESCE(m.association, stored) falls back to stored; views not yet swapped. OnModuleInit triggers an immediate populate.

3. Verify the table populated (seconds after boot; else await/trigger one hourly run):

SELECT repo_full_name, COUNT(*) FROM maintainers GROUP BY 1 ORDER BY 2;   -- each registered+installed repo >= 1 (owner)
SELECT * FROM maintainers WHERE repo_full_name='phase-rs/phase' AND github_id IN ('1388610','59729252');  -- expect MEMBER/COLLABORATOR

4. Cut over the views (instant CREATE OR REPLACE VIEW; paste the new bodies from this PR):

• 21_view_pr_review_summary.sql
• 22_view_pr_linked_issues.sql
• 24_view_pr_labels_by_actor.sql
• 25_view_issue_labels_by_actor.sql

5. Verify serving + perf:

EXPLAIN ANALYZE <windowed pulls query for 156195510>;   -- no contributor_repo_roles node; per-row cost collapsed

• GET /repos/phase-rs/phase/maintainers returns the live set.
• Row payload for github_ids 1388610 / 59729252 shows MEMBER/COLLABORATOR; SELECT author_association … still shows the unmutated snapshot.

6. Rollback (if needed): CREATE OR REPLACE VIEW the four views back to their contributor_repo_roles/stored bodies (originals are on origin/test). Table + app can stay (harmless).

Follow-ups

• After deploy, close gittensor #1485 (routes to the now-superseded stale maintainers path).
• Independent of the pagination PRs (mirror feat(miners): add cursor pagination to windowed POST pulls/issues endpoints #191 merged, gittensor #1487 draft) — they stack; with this landed the page limit could be raised to 200.