feat: store state and combined snapshot+state in S3 with configurable retention

[marcus-girolneto]

Summary

Snapshot-streaming currently uploads only the raw incremental snapshot to S3. That is not sufficient to support node rollback: both the GlobalSnapshotInfo (state) and the combined snapshot+state bundle are required to validate and restore. Two production incidents — nodes missing snapshots, and the streaming DB getting stuck on inserts — both benefit from having the last N states reliably retrievable from S3.

• Uploads state (GlobalSnapshotInfo) and combined (SnapshotWithState) to S3 alongside the existing snapshot upload. Each upload is independently toggled.
• Keeps the last retentionCount of each (default 1000) via exact-prefix pruning on the expired ordinal — O(1) per upload, no full bucket scan.
• Key scheme states/{ordinal}-{hash} and combined/{ordinal}-{hash}: ordinal enables cheap retention, hash prevents fork collisions. S3 object user-metadata carries both ordinal and hash for tooling.
• Payloads are gzipped JSON (state is ~9.5 MB uncompressed).
• downloadState(ordinal, hash) / downloadCombined(ordinal, hash) exposed on S3DAO for future rollback verification callers.

Config

New fields under snapshotStreaming.s3 (all default to off / safe values):

s3 {
  uploadEnabled = false          # existing — raw snapshot
  uploadStateEnabled = false     # new — GlobalSnapshotInfo
  uploadCombinedEnabled = false  # new — SnapshotWithState (rollback bundle)
  retentionCount = 1000          # new — lastN kept for state + combined
}

Files changed

• src/main/scala/org/constellation/snapshotstreaming/Configuration.scala
• src/main/scala/org/constellation/snapshotstreaming/s3/S3DAO.scala
• src/main/scala/org/constellation/snapshotstreaming/SnapshotProcessor.scala
• src/main/resources/reference.conf

Out of scope / follow-ups

• Rollback verification call site. downloadState / downloadCombined are exposed but no consumer wires them in. The ticket's "ensure via S3 download that state/ordinal/hash match" is presumably handled by the network-monitoring service that triggers the rollback.
• Backfill of pre-existing ordinals. Retention only kicks in on new uploads, so flipping the flags on a running stream at ordinal N populates S3 forward only; nothing below N - retentionCount is ever written. Matches "store lastN going forward." Happy to add a one-shot backfill if needed.

Test plan

• CI passes (develop pulls a snapshot tessellation artifact that is not locally cached; CI resolves from the internal repo)
• Deploy to a non-production env with uploadStateEnabled = true and uploadCombinedEnabled = true, retentionCount = 100 for quick verification
• Confirm states/{ordinal}-{hash} and combined/{ordinal}-{hash} objects appear with correct user-metadata
• After streaming >100 snapshots, confirm the oldest ordinals are pruned from both prefixes
• Download a combined object, gunzip, confirm JSON round-trips into SnapshotWithState and that ordinal/hash match the object key

[gclaramunt]

LGTM