fix(webhook): self-heal PR state drift from missed merge webhooks

[LandynDev]

Problem

A merged PR was showing as OPEN in the mirror (and therefore on the dashboard, and therefore mis-scored by the validator). Confirmed on phase-rs/phase for miner 156195510: PRs #2745, #2751, #2752, #2753, #2756, #3095 are MERGED on GitHub (with merge commits) but sit at state=OPEN, merged_at=NULL in the mirror DB.

Root cause

PR state/merged_at is written by exactly one runtime path — the pull_request webhook handler. If the pull_request.closed (merged) delivery is missed/dropped, nothing ever corrects it:

1. handlePrMetadata runs frequently (every synchronize/edited/closed) but only wrote closingIssueNumbers/body/lastEditedAt — never state. (Logs show #3095 metadata-fetched 9× on 06-12, the last 2s after its merge, yet it stayed OPEN.)
2. No reconciliation existed — the only periodic job was webhook-delivery pruning.
3. Backfill only ran on manual admin trigger / repo register.

So a single missed merge webhook = permanent false-OPEN. The validator reads PR state verbatim from the mirror, so the miner is genuinely under-credited (those PRs never enter merged_prs).

Changes

• github-fetcher.ts — fetchPrMetadata now also returns authoritative GraphQL state/mergedAt/closedAt/mergedBy. (GraphQL state is the source of truth: OPEN/CLOSED/MERGED.)
• fetch.processor.ts — handlePrMetadata re-asserts that state. MERGED is treated as terminal so an in-flight stale fetch can't revert it; logs a State drift corrected … warning whenever it flips a PR (observability — this drift was previously invisible).
• pr-reconcile.service.ts (new) — hourly sweep enqueues a metadata fetch for every still-open PR in registered repos within the scoring window (PR_RECONCILE_WINDOW_DAYS, default 45). This is the actual self-heal for missed merge webhooks. Runs once at startup too. Interval/window env-tunable.
• repo-backfill-schedule.service.ts (new) — daily full backfill per registered repo as a coarse safety net (also catches issue/label drift). Kill-switch: NIGHTLY_BACKFILL_ENABLED=false. Heavier than the sweep, hence daily + disableable.
• pull-request.handler.ts — derive merged state from pr.merged alone (synthesize merged_at from closed_at if absent), so a closed event arriving before merged_at is populated isn't pinned to OPEN.
• .env.example — documents the new tunables.

Layering

pr-reconcile (cheap, hourly, open-PRs-only) is the primary fix; nightly backfill is belt-and-suspenders. Both reuse existing queue jobs (meta-… jobId dedupes against webhook-triggered fetches; concurrency-5 worker provides backpressure).

Effect on the stuck PRs

Once deployed, the startup + hourly reconcile sweep re-fetches those 6 PRs, flips them to MERGED, and the next validator round (~2h) credits the miner. (A one-time manual admin/backfill for phase-rs/phase still recommended to correct them immediately rather than waiting for the first sweep.)

Verification

npm run build, npm run lint, prettier --check all pass locally. No existing tests reference the changed APIs.