Fix/129 label events cross path dedup

[enjoyandlove]

Summary

Fixes duplicate rows in label_events caused by the live webhook and backfill paths recording the same label action under two different identities.

The root cause: the live webhook payload carries no GraphQL LabeledEvent/UnlabeledEvent node id, so the live path could only key a row on timestamp — and it stored the mirror-receive time, while backfill stored GitHub's createdAt. The two clocks never coincide for the same action, so the old uq_label_events_natural_key (which includes timestamp) never collapsed the pair: every action seen by both paths produced two rows.

This PR introduces a path-independent identity and reconciles the two paths around it:

• github_node_id column on label_events — the globally-unique GraphQL node id, the only path-independent event identity. Set on authoritative backfill rows; NULL on provisional live-webhook rows.
• Live path writes a provisional row (github_node_id NULL, timestamp = mirror-receive time).
• Backfill writes the authoritative row (github_node_id set, timestamp = GitHub createdAt) and then reconciles: it deletes each provisional duplicate the authoritative row supersedes, pairing 1:1 and nearest-in-time within a 120s delivery-latency window. A provisional row with no authoritative row to claim it survives, so events backfill hasn't captured yet are never lost.
• Dedup guard repointed onto the stable identity: drop uq_label_events_natural_key, add a partial uq_label_events_github_node_id (WHERE github_node_id IS NOT NULL) so backfill↔backfill collapses by true identity while provisional rows are never constrained against each other.
• One-time historic dedup in the migration, gated on the presence of the old index so it runs exactly once and is a no-op on every later (re)deploy.

The 120s reconcile window is defined in both github-fetcher.service.ts (LABEL_EVENT_RECONCILE_WINDOW_SECONDS) and 07_label_events.sql, and the two must stay in sync.

Related Issues

Fixes #129

Type of Change

• Bug fix
• New feature
• Refactor
• Documentation
• Other (describe below)

Testing

• Verified the live path inserts a provisional row (github_node_id NULL, mirror-receive timestamp) and backfill inserts the authoritative row (github_node_id set, createdAt timestamp).
• Confirmed reconciliation deletes the provisional duplicate once the authoritative row lands, with pairing limited to the closest provisional row inside the 120s window.
• Confirmed a provisional row with no matching authoritative row is preserved (events not yet/never backfilled are not lost).
• Confirmed genuine repeat actions (add → remove → re-add) spaced beyond the window are not merged.
• Confirmed the one-time historic dedup runs only when uq_label_events_natural_key is present and is a no-op on re-deploy.
• Confirmed backfill re-runs / BullMQ retries collapse to a no-op via the partial uq_label_events_github_node_id index.

Checklist

• I have read the Contributing Guide
• Code builds without errors
• New and existing tests pass (if applicable)
• Documentation updated (if applicable)
• No unnecessary dependencies added