Closes #898.

Summary

processNewBlockAnnouncement previously routed any gap ≥ 2 between the assembler's current block and the blockchain tip through handleReorg. When moveBack == 0 (pure catch-up after downtime), handleReorg would still evaluate the large-reorg reset condition; with moveForward ≥ CoinbaseMaturity after a long downtime, this incorrectly triggered a full reset of block assembly state.

Fix

Pre-fetch reorg blocks at the top of the gap ≥ 2 case, then branch on moveBack length:

• moveBack == 0 → new handleCatchUp helper: delegates to subtreeProcessor.Reorg([]*model.Block{}, blocks). reorgBlocks already has a fast path for len(moveBackBlocks) == 0 && len(moveForwardBlocks) > 0 (SubtreeProcessor.go:2794) that iterates moveForwardBlock per block with skipNotification = true and skipDequeue = true for every block except the last, and rolls back the subtree processor to pre-catchup state on any mid-loop failure (lines 2820-2828). On error, processNewBlockAnnouncement returns early so BlockAssembler.setBestBlockHeader is not called either; BA and the subtree processor both stay at the pre-catchup tip, aligned. Note: pass an empty (non-nil) []*model.Block{} for moveBack — reorgBlocks rejects nil at line 2778 but takes the fast path at len == 0.
• moveBack > 0 → handleReorg(ctx, header, height) unchanged. The function body is byte-identical to upstream; the only cost on the reorg path is one additional getReorgBlocks call, accepted as the price of zero blast radius on the reorg implementation.

New prometheusBlockAssemblerCatchUp counter mirrors the existing reorg counter for ops visibility.

Test plan

TestProcessNewBlockAnnouncement_CatchupVsReorg with 7 subtests:

• gap=0 no-op — same hash returns early without state change.
• gap=1 moving up — existing default branch, single MoveForwardBlock.
• gap=2 catch-up moveBack=0 — new catch-up path; stubs Reorg, asserts called with len(moveBack) == 0 and the expected moveForward slice.
• large catch-up moveBack=0 moveForward=150 no reset — pins the regression the fix addresses; 150-block forward gap routes through Reorg fast path, never through handleReorg's reset gate.
• real reorg moveBack=1 — existing reorg path unchanged.
• large reorg triggers reset — existing handleReorg reset gate still fires for genuine reorgs at depth ≥ CoinbaseMaturity.
• catch-up mid-error stops — stubs Reorg to return an error; asserts CurrentBlock() returns the pre-catchup tip (because processNewBlockAnnouncement returns early on error, leaving BA's setBestBlockHeader unreached, and reorgBlocks rolls back the subtree processor's state in the same scenario).

make lint-new clean, go test -race ./services/blockassembly/... passes.