#36000 validation: prefetch blocks while connecting

• Overlap block read I/O with validation during initial sync, reindexing, and reorgs
• Significantly reduces sync and reindex times for node operators

Adds a Chainstate-owned BlockFetcher with a 2-worker thread pool that asynchronously reads and deserializes upcoming blocks from disk up to a depth of 4 while the validation thread is busy connecting earlier blocks. It also handles prefetching the first sibling block on a competing fork during reorgs while disconnecting the existing tip.

Problem: Block reading and deserialization from disk occur synchronously just before connecting each block, leaving CPU validation cores idle during disk I/O and disk readers idle during validation.

P2 · speedup

• P2 because it brings double-digit performance gains to core block connection
• Reduces wall-clock time by 13-15% for IBD and 35-50% for reindexing for node operators

Delivers a substantial and empirically proven speedup for IBD and chain reindexing (e.g. 13% to 35% time reduction in benchmarks). Users directly feel faster synchronization times, and the abstraction also paves the way for index sync parallelization.

Membership: Directly modifies Chainstate::ActivateBestChainStep and block connection pipelines in src/validation.cpp.

Factors: security/stability 0, bug 0, performance 2, user value 2, leverage 1

• Ready for review

The implementation is complete, well tested, and stable; w0xlt's recent comment is a straightforward 2-line fix that does not invalidate review.

Author status: Active; addressed initial design feedback and provided comprehensive multi-hardware benchmarks, pending response to w0xlt's recent comment.

Open concerns:

• w0xlt pointed out that if a block fails validation, prefetched descendant blocks could remain in memory indefinitely unless Clear() is invoked on failure.

Resolved concerns:

• Thread pool creation spam on every ActivateBestChain call was fixed by moving BlockFetcher lifetime onto Chainstate.
• Potential circular dependency between node/blockfetcher, blockstorage, and validation was resolved using functional callback injection.
• Prefetch queue sizing and sliding window dynamics were tuned down to 2 threads and queue depth of 4 based on benchmarks.

• Strong support with independent benchmarks verifying speedups (andrewtoth)
• Verified double-digit speedups for IBD and up to 35% on reindexing (andrewtoth)
• Minor suggestion to clear the queue when an invalid block is found (w0xlt)

Strong approval and benchmarks from andrewtoth; one open nonblocking memory retention edge case noted by w0xlt.

andrewtoth provided a full ACK backed by extensive benchmarks. An open nonblocking observation from w0xlt regarding clearing the queue on an invalid block remains unanswered.

• andrewtoth: 'ACK 3547915bfb... Managed to see speedup in IBD benchmarks... Nice speedups, especially for reindex-chainstate.'
• w0xlt: 'When a block was rejected, blocks read ahead for its branch could remain in memory indefinitely.'

Objections:

Support:

• andrewtoth: Measured significant speedup on IBD (3.8% to 13.4%) and reindex-chainstate (12.7% to 35.8%) across machines

Participants: fanquake (objection), andrewtoth (support), w0xlt (objection)

State derived from the lists: nonblocking objection open (w0xlt)

Review verdicts (DrahtBot): 1

• ACK: andrewtoth

Enables:

• Parallel block reading in BaseIndex synchronization

47 lines under test/bench/ci.

• src/node/blockfetcher.cpp +59/-0
• src/node/blockfetcher.h +52/-0
• test/functional/feature_reindex.py +44/-3
• src/validation.cpp +19/-5
• src/validation.h +6/-1
• doc/release-notes-36000.md +6/-0
• doc/developer-notes.md +3/-0
• src/CMakeLists.txt +1/-0
• src/kernel/CMakeLists.txt +1/-0

Adds a background BlockFetcher that prefetches and deserializes disk blocks ahead of block connection during chain activation and reorgs. This resolves the bottleneck where CPU validation and disk I/O do not overlap during reindexing and IBD. Extensive benchmarks demonstrate substantial elapsed time speedups between 13% and 35%. The PR has an ACK from andrewtoth, with one minor nonblocking comment from w0xlt regarding memory clearing on invalid blocks pending author attention.