fix(cnight-observation): chunk the bulk cache catch-up pull

[skylar-simoncelli]

Overview

Investigating the "slow db-sync query" spikes on devnet/stagenet (node 2.0.x) traced the load to the cNIGHT observation path, not to db-sync resourcing. Two facts pin it to the node code, not the DB:

• devnet (2.0.x) and qanet (1.0.0) read the same shared Aurora; only the 2.0.x one is slow. stagenet (2.0.x) has its own dedicated Aurora and is slow too. The variable that tracks the symptom is the node version, not the database.
• An EXPLAIN ANALYZE of the cNIGHT observation join shape ran ~53 s on an otherwise idle DB. The plan is sensible (index scans, accurate estimates) — the cost is the range width scanned, not a missing index or contention.

This PR fixes one concrete driver of that width: the bulk cache's catch-up pull.

Change

RefreshContext::refresh warms/slides the in-memory cNIGHT window in the background. It pulled the whole [from_block, target_end] range in a single bulk_pull with LARGE_LIMIT. In steady state that range is small ([old_end + 1, target_end]), but when a node is far behind tip (cold start, or after a lag) the span can be hundreds of thousands of Cardano blocks, and one query over it is a multi-minute db-sync scan that stalls block import.

The refresh now pulls the range in bounded block-span chunks (REFRESH_CHUNK_BLOCKS = 5000), committing snapshot_end_block after each chunk so an interrupted catch-up resumes from the last committed block instead of restarting the giant query.

Consensus-safe / scope: this is a cache-warming detail only. The union of chunks covers exactly the same range, so no inherent payload changes. The per-block consensus path (MidnightCNightObservationDataSourceImpl::get_utxos_up_to_capacity) is untouched and still queries through tip — deliberately, per its existing replay-semantics comment.

Relationship to other PRs

• Supersedes the approach in fix(cnight-observation): bound IDP queries by CardanoBlockWindowSize #1432 (which clamped the inherent query upper bound — no longer viable, since the surviving path documents that clipping the SQL range changes the inherent payload and breaks import of already-authored blocks). fix(cnight-observation): bound IDP queries by CardanoBlockWindowSize #1432 should be closed as superseded; its root-cause writeup remains accurate.
• Complements fix(node): re-anchor cNIGHT observation window near tip on from-genesis sync #1836 (re-anchor the window near tip): fix(node): re-anchor cNIGHT observation window near tip on from-genesis sync #1836 reduces how often a node ends up far behind / oversized; this PR bounds the query cost when a catch-up does happen. They stack.

Testing Evidence

• rustfmt --edition 2024 clean.
• Not compiled locally (dev-machine build policy); relying on CI for cargo check/clippy/tests.
• Follow-up wanted: a unit test around the chunk boundary in refresh (chunk_end clamping, resume via committed snapshot_end_block). Happy to add once a reviewer confirms the approach.

Submission Checklist

• Changes are backward-compatible
• PR description explains why the change is needed
• Self-reviewed the diff
• Included a change file
• Bump node minor version (n/a — behavior-preserving perf fix)
• Additional tests (see follow-up above)