follow-up: durable occurrence history, restart-safe L2, and V1 migration

[AllenMuu]

Parent: #3

This child tracks the work intentionally deferred from Issue #3 after its P0 deterministic identity layer is complete.

Preconditions

• feat: evolve error grouping with normalization, cause-chain fingerprints and persistent history #3 delivers ThrowableInfo, explicit normalization, V2 strict/loose fingerprints, a durable ErrorGroup source of truth, and versioned composite lookup keys.
• P0 records a minimal idempotency key for accepted events so group counters are not inflated by retries. This issue expands that into retained occurrence history; it must not weaken the P0 counting guarantee.

Scope

• Add durable, retention-controlled ErrorOccurrence records for trace/drill-down and environment trends.
• Make vector-cluster similarity survive process restarts via a queryable durable index or a tested startup reconstruction; do not rely only on a process-local clusterIndex.
• Define and implement the V1-to-V2 transition: V2-primary dual read, V1 read-only behavior, observability, retirement criteria, and any backfill needed for persisted legacy groups.

Out of scope

• Replacing the LLM RCA pipeline.
• Changing feat: evolve error grouping with normalization, cause-chain fingerprints and persistent history #3 strict fingerprint semantics or auto-reusing RCA for a loose-fingerprint match.

Acceptance criteria

• Duplicate (appName, eventId) submissions are idempotent; accepted distinct events preserve first/last-seen and accurate group counts.
• Occurrence retention or sampling never changes group-level totals.
• After restart, a qualifying historical vector match still returns the existing cluster without an LLM call.
• V1 groups remain readable during the declared dual-read period; new groups use V2; retirement is measurable and reversible.
• Integration tests cover restart behavior and migration/dual-read behavior.