When a storage slot is zeroed (SSTORE→0), the DB stores an empty-value
entry. If snapshot files later advance past that entry's step,
getLatestFromDb discarded it (returned found=false) because of the
lastTxNumOfStep < files.EndTxNum() guard. The caller then fell through
to getLatestFromFiles, which scans files newest-to-oldest. Since
snapshot files represent deletions as absent keys (no tombstone), the
scan reached an older file containing the pre-deletion value and
returned stale data.

This caused EIP-2200 SSTORE gas mispricing: GetCommittedState returned a
non-zero "original" value for a deleted slot, taking the SSTORE_RESET
path (2,900 gas) instead of SSTORE_SET (20,000 gas), producing gasUsed
deficits that fail block validation.

Fix: treat DB deletion entries (len(v)==0) as authoritative regardless
of step age. The DB returns the most-recent entry per key (inverted-step
ordering), so an empty value means the key was deleted and no newer
write exists.

Fixes #20169

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Return v ([]byte{}) instead of nil to preserve the empty-vs-nil
distinction. DeleteWithPrev and PutWithPrev(key, []byte{}, ...) are
indistinguishable at the DB level — both store empty values. Returning
nil broke TestDomain_GetAfterAggregation which writes legitimate empty
values and expects to read them back as []byte{}.

The critical part of the fix remains: found=true prevents the
fallthrough to getLatestFromFiles. For storage gas pricing, []byte{} is
correctly treated as zero by ReadAccountStorage (checks len(enc) > 0).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

…otstrapping

Two additional fixes for the gas mismatch issue (#20169):

1. **Preserve deletion entries during domain pruning** (db/state/domain.go):
   The background file collation (`buildFilesInBackground`) runs concurrently
   with block execution. When it opens its MDBX read-transaction before the
   execution's CommitCycle flushes a deletion entry, the resulting domain file
   silently misses that deletion. Pruning then removes the DB entry because the
   file "covers" that step — leaving no record of the deletion in either the DB
   or the files. Subsequent reads fall through to an older file and return a
   stale pre-deletion value, causing EIP-2200 SSTORE gas mispricing.

   Fix: route Prune() through oldPrune() and skip deletion entries (empty
   values) during pruning. Combined with the getLatestFromDb fix from the
   previous commit, this ensures deletion evidence is never lost regardless
   of collation/execution timing.

2. **Tolerate domain-ahead-of-blocks during InsertBlocks** (inserters.go):
   When pre-verified state snapshots extend past the block snapshot boundary,
   NewSharedDomains returns ErrBehindCommitment. InsertBlocks only needs to
   accumulate block headers/bodies/TDs in the overlay — it does not need a
   valid commitment state. Allow it to proceed so the ForkChoice stage can
   run execution and close the TxNums gap.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

The background file builder (buildFilesInBackground) runs concurrently
with block execution. Its collation opens a read-transaction that can
snapshot the DB before the execution's CommitCycle flushes all writes for
a step. The resulting file silently misses those entries. Pruning then
removes them from the DB (the file "covers" the step), and the values
are lost — causing both stale reads for deleted keys and missing reads
for existing keys. See #20169.

Fix: the execution layer publishes its last committed txNum via
Aggregator.SetLastCommittedTxNum after each CommitCycle flush+commit.
The file builder checks this before collating each step: if
lastCommittedTxNum < firstTxNum(step+1), the step may have uncommitted
data and collation is deferred.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Reverts the InsertBlocks change from e1475aa that allowed
ErrBehindCommitment to be silently ignored. The bootstrapping issue
it addressed is better handled by the initialCycle switch fix in
5d257a0.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

OldPrune() is never called in production code. The deletion-entry
preservation for the active prune path is handled via the
IsDeletionEntry callback in TableScanningPrune.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Mirror the serial executor fix: when the parallel executor is within
MaxReorgDepth of the batch end, switch initialCycle to false so
changesets are generated. Without this the parallel executor can reach
the tip with no changesets, making reorgs impossible.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Address review feedback (AskAlexSharov): remove `lastCommittedTxNum`
state from the Aggregator. Instead, the background file builder reads
the latest committed txNum directly from the commitment domain state
in the DB via `GetLatest(kv.CommitmentDomain, KeyCommitmentState)`.

This avoids adding statefulness to the Aggregator and uses the existing
SharedDomains.ComputeCommitment write as the synchronization point.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

This reverts commit c6c5d4a.

Commit 835dd9b in ethpandaops/ethereum-package introduced GpuConfig,
a Starlark built-in that requires Kurtosis CLI >=1.18.1. CI installs
an older version so Starlark eval fails. Pin to e07503d (the commit
just before the regression) until CI is upgraded to Kurtosis 1.18.1.

…tion-entries