evmonly/evmonlyapp: revert the execution optimizations (#4254, #4258, #4260, #4261) ahead of a reviewable stack

[bdchatham]

giga-1 went from ~72k to ~96k tx/s through four execution changes (#4254, #4258, #4260, #4261) that were merged quickly during the testnet-2 tuning. To get them genuinely reviewed, this PR takes all four back out, and stack #4274 re-opens each one on top of it in dependency order so that the top of the stack is byte-identical to today's giga-1.

The branch is four git revert -m 1 commits and nothing else, newest first: #4261 (parallel OCC validation and merge; occ_shards.go goes, occ.go is the serial walk again), #4260 (PrepareBlock(n+1) while n executes; the router is back to fetch → execute, PrepareBlock leaves evmOnlyApplication and the proxy, and the router test's CheckBlocks helper goes with it), #4258 (Executor.ReadLatestAccount and the pending overlay; the mempool's first-seen lookup is back under the store lock), and #4254 (FinalizeBlock calls executor.ExecuteBlock synchronously again, so settler/AwaitCommits go). Conflicts were resolved so that only those four change: #4259's evmonly_finalize timer stays as take_senders → execute → tx_results (the prepare phase returns with #4254), and #4263's storage-tail phases and #4265's enable_receipt_store with its test are untouched. One thing is not preservable: #4267's evmonly_prepare timer wraps evmOnlyApplication.PrepareBlock, so evmonly_prepare_phase_duration_seconds_total goes dark until stack 3/4 re-lands #4260.

No config keys or wire formats change. The image at this commit should perform like the pre-#4254 build (~72k tx/s) and is not meant to be rolled out on its own. Validated with make lint/fmtcheck, scripts/ramtest.sh -race ./giga/evmonly/..., and go test -race on sei-tendermint/internal/{evmonlyapp,mempool,p2p,proxy} and sei-tendermint/node.

Link to Devin session: https://app.devin.ai/sessions/ff612badcded4aa5914ea408dbb41888
Open in Devin Desktop: https://app.devin.ai/desktop/session/ff612badcded4aa5914ea408dbb41888?variant=devin
Requested by: @bdchatham

[devin-ai-integration]

I'll fix CI failures and address comments from users with write access. I'll skip comments containing "(aside)".

• Disable automatic comment, CI, and merge conflict monitoring

[cursor]

PR Summary

High Risk
Changes core block execution, optimistic concurrency validation, async state commits, and RPC/mempool account reads—areas where subtle ordering bugs would affect correctness and chain halt behavior.

Overview
Reverts four performance-oriented changes (commits #4254, #4258, #4260, #4261) so they can land again as a reviewable stack. Expected throughput drops from the ~96k tx/s path back toward the pre-optimization ~72k tx/s baseline; this branch is not meant for production rollout.

Block-STM / OCC (giga/evmonly): Removes sharded parallel validation, parallel prefix acceptance, and parallel changeset merge (occ_shards.go deleted). Validation again walks transactions in order on one goroutine, with a simpler write index (conflictsWithAfter vs span-based conflictsWithin). Accepted prefix state is a single blockSTMState instead of address shards. Merge still uses optional account-row prefetch via the worker pool for large blocks before serial ChangeSetInto.

Commit pipeline: Pending block state is a cloned StateChangeSet plus newPendingOverlay again—no ReadLatestAccount, pipelineGeneration, or executor API to read balances/nonces ahead of landed commits. Merge no longer falls back to sequential execution when the OCC worker pool is closed during parallel merge (that path is gone).

EVM-only app & node: FinalizeBlock calls executor.ExecuteBlock directly—no PrepareBlock, prepared-block cache, executeBlockPipelined, or settler/AwaitCommits. EvmNonce, EvmBalance, and EvmCall read the committed store view (with the usual pending-block cursor rules for calls), not in-flight commit overlays. Shutdown no longer waits for background commits before closing Giga storage.

Mempool: Drops lock-free prefetch of first-seen EVM accounts on insert (and related epoch invalidation on Update).

Giga router / proxy: Block loop is fetch → FinalizeBlock → execute again; PrepareBlock and Proxy.AwaitCommits are removed.

Tests and docs are trimmed to match the serial behavior; load-test bench drops the ERC20 scenario in favor of transfer-only workloads.

^{Reviewed by Cursor Bugbot for commit 8789d48. Bugbot is set up for automated code reviews on this repo. Configure here.}

[github-actions]

The latest Buf updates on your PR. Results from workflow Buf / buf (pull_request).

Build	Format	Lint	Breaking	Updated (UTC)
✅ passed	✅ passed	✅ passed	✅ passed	Sep 20, 2026, 10:51 PM

[codecov]

Codecov Report

❌ Patch coverage is 88.94009% with 24 lines in your changes missing coverage. Please review.
✅ Project coverage is 83.01%. Comparing base (7e3ca24) to head (8789d48).
⚠️ Report is 49 commits behind head on giga-1.

Files with missing lines	Patch %	Lines
sei-tendermint/internal/p2p/giga_router_common.go	72.72%	8 Missing and 1 partial ⚠️
sei-tendermint/internal/evmonlyapp/app.go	86.44%	8 Missing ⚠️
giga/evmonly/occ.go	94.16%	7 Missing ⚠️

Additional details and impacted files

@@             Coverage Diff             @@
##           giga-1    #4269       +/-   ##
===========================================
+ Coverage   65.55%   83.01%   +17.45%     
===========================================
  Files        2081       65     -2016     
  Lines      157460     8725   -148735     
===========================================
- Hits       103222     7243    -95979     
+ Misses      54097     1480    -52617     
+ Partials      141        2      -139     

Flag	Coverage Δ
sei-chain	?
sei-chain-pr	83.01% <88.94%> (?)
sei-db	?

Flags with carried forward coverage won't be shown. Click here to find out more.

Files with missing lines	Coverage Δ
giga/evmonly/executor.go	91.25% <ø> (+0.04%)	⬆️
giga/evmonly/giga_store.go	91.48% <100.00%> (-1.62%)	⬇️
giga/evmonly/pipeline_overlay.go	96.29% <100.00%> (ø)
sei-tendermint/internal/mempool/tx.go	95.03% <ø> (ø)
sei-tendermint/internal/proxy/proxy.go	88.88% <ø> (-2.23%)	⬇️
sei-tendermint/node/node.go	73.44% <ø> (-1.29%)	⬇️
giga/evmonly/occ.go	90.73% <94.16%> (+1.54%)	⬆️
sei-tendermint/internal/evmonlyapp/app.go	83.98% <86.44%> (+2.21%)	⬆️
sei-tendermint/internal/p2p/giga_router_common.go	78.44% <72.72%> (+0.24%)	⬆️

... and 2029 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

[seidroid]

A faithful, self-consistent revert of #4254/#4258/#4260/#4261: no dangling references to the removed shard/pipeline/prepare machinery remain, and the synchronous Executor.ExecuteBlock (which still calls AwaitCommits) is what makes the app-level settle removals safe. Two non-blocking notes: the PR description overstates what is retained, and one standalone router test assertion was dropped that did not need to be.

Findings: 0 blocking | 3 non-blocking | 1 posted inline

Blockers

• None at the file/PR level.

Non-blocking

• [suggestion] The description says "#4267's evmonly_prepare timer [is] unchanged in the serial loop", but it is removed outright: #4267 only re-added that timer around evmOnlyApplication.PrepareBlock, which this PR deletes with #4260. There is no synchronous preparation step left in the app to wrap (executor-internal preparation is inside ExecuteBlock), so nothing is preservable here — but the claim should be corrected, and operators should know evmonly_prepare_phase_duration_seconds_total goes dark until #4260 is re-landed.
• 1 suggestion(s)/nit(s) flagged inline on specific lines.
• 1 non-blocking pre-existing issue(s) listed below under pre-existing issues.

Pre-existing issues

• [suggestion] evmOnlyApplication.EvmCall (sei-tendermint/internal/evmonlyapp/app.go) releases the cursor lock before executor.Call opens its store view, so a full FinalizeBlock+Commit can land between the block context being captured and the snapshot being opened, mixing height N's NUMBER/TIMESTAMP with N+1's state. The pending.IsPresent() guard closes the wide window (pending is staged before the state commit starts), but this narrow gap is present on the base branch too — the base's context recheck loop does not eliminate it either.