fix(dataplane): close idempotency dedup race with transactional advisory lock claim

[mekilis]

Summary

• The idempotency duplicate check ran in the async workers as check-then-insert, so two near-simultaneous creates with the same key could both pass the check and both be persisted and delivered as non-duplicates.
• The authoritative duplicate check now runs inside the event repository's insert transaction under pg_advisory_xact_lock(hashtextextended('events-idempotency:<project>:<key>', 0)); the re-check under the lock sees rows committed by earlier holders and flips IsDuplicateEvent, so exactly one row per key persists as non-duplicate. Lock or lookup errors fail closed and the worker retries.
• A unique index is not an option because the events table is partitioned and its unique indexes must include the partition key, which (project_id, idempotency_key) cannot satisfy.

Design notes

• The fanout service's novelty predicate (FindFirstEventWithIdempotencyKey, non-duplicate rows only) diverges from the claim's (FindEventsByIdempotencyKey, any row). They only disagree if a key's original non-duplicate row was deleted while duplicate rows survive; the claim errs toward flagging the new event duplicate, which is the safe direction.
• The advisory lock is held through the endpoint batch inserts. This only serializes writers of the same project + idempotency key; unrelated writers are unaffected.

Test plan

• go test ./internal/events/ (concurrent same-key creates yield exactly one non-duplicate row; lock/lookup error path fails closed)
• gofmt, go vet, golangci-lint clean on touched packages

Fixes PDE-929

[linear]

PDE-929

[cursor]

Current version of PR was reviewed by /review-bugbot on Jul 20, 15:50 GMT+1. It flagged 0 findings.

^{Bugbot on commit 7fc0057 is skipped.}