Manual reapply of #21245 (runloop refactor: split prune from commit-cycle) onto the performance branch, plus follow-up fixes for the post-frozen-blocks prune-budget regression observed on bloatnet.

Changes

• Split CommitCycleFn from PruneFn in RunLoopConfig; drop BeforeIteration/PruneTimeout/FirstCycle.
• ProcessFrozenBlocks: kicks agg.BuildFilesInBackground from the new CommitCycle callback so snapshot files advance during PFB (previously they only progressed after normal sync resumed). On the last iter (!hasMore) returns (nil, nil) to skip a wasted BeginTemporalRw. Outer defer tx.Rollback() is closure-form so it follows the closure's tx reassignments across iterations.
• updateForkChoice: PruneFn is a no-op at tip (post-RunLoop path handles flush+commit+prune); in catchup it drains pipeline prune via runForkchoicePrune(true) — initialCycle=true so PruneExecutionStage gets the catchup budget instead of the 2 s slot budget that can't keep up with 40-block bursts.
• initialCycle predicate — block-count delta against finishProgressBefore:

  const smallBlockJumpThreshold = 16
  headNum := fcuHeader.Number.Uint64()
  initialCycle := headNum > finishProgressBefore && headNum-finishProgressBefore > smallBlockJumpThreshold

  The headNum > guard avoids uint64 underflow when an FCU goes back (e.g. ePBS on Glamsterdam). Threshold 16 engages catchup mode for bloatnet's ~40-block bursts and stays out of the way for steady-state 1-block tip FCUs and test fixtures.
• runForkchoicePrune: short-chain skip-gate removed (maxTxNum < (stepSize*5)/4). The gate left ChangeSets3 un-pruned on disk for short chains, which broke MaxReorgDepth enforcement in tests (TestFcuReturnsReorgTooDeepCode38006 on the upstream main branch). Skip removal is also consistent with the exec3/storage-component direction (c380b438e7).
• FCU CommitCycle safety: defer commitRwTx.Rollback() immediately after BeginTemporalRw so the Commit-error path doesn't leak the RW tx. Rollback after a successful Commit is idempotent (per Copilot review on execution/execmodule: apply runloop refactor with furious-prune fix #21245).

Why the predicate change matters — bloatnet

bloatnet stabilises at ~38 GB on tip only when initialCycle=false is held until execution actually reaches tip:

• A 40-block batch (what --batchSize=100mb settles on) writes ~1.4 GB into MDBX (lots of commitment trie changes).
• At-tip prune budget (SecondsPerSlot/3 ≈ 2 s) removes less than one such batch.
• Flipping too early — while bursts are still arriving — lets writes outrun prune and MDBX runs away.

The earlier !isSynced and wall-clock head-age variants either flipped too eagerly (Caplin keeps headers/finish aligned during bursts) or never (test fixtures use ancient timestamps, regressing several rpc/jsonrpc + engineapi tests). Block-count delta with threshold 16 is a clean proxy and works for tests, mainnet tip, and bloatnet bursts.

Plateau holds across cycles vs. the prior failure mode where the file extended +6 GB/cycle once the predicate flipped early.

Remaining differences vs #21245 (intentional)

• Aggregator.MaxCollationTxNum() getter and the BuildFilesInBackground cap pattern from execution/execmodule: apply runloop refactor with furious-prune fix #21245 are NOT ported here — the underlying maxCollationTxNum field doesn't exist on the performance branch lineage. Followup, requires also adding the field on this branch.
• runForkchoicePrune body uses e.db.UpdateTemporal(...) directly (matches this branch's storage design — agg.CollateAndPruneIfNeeded ownership is being moved out of the FCU path on the exec3/storage-component track), whereas execution/execmodule: apply runloop refactor with furious-prune fix #21245 still calls CollateAndPruneIfNeeded via this function.

Test plan

• CI green on rerun (race-tests / tests-mac-linux all OSes; sonar).
• Bloatnet → chain-tip; db_size stays bounded.
• Once at tip, confirm block-count delta predicate flips initialCycle=false for steady-state 1-block FCUs.
• Sanity-check mainnet behaviour (FCU bursts ≤16 blocks at tip).