docs(runbooks): validate gas-vs-execution-time methodology end to end

.agent/runbooks/README.md

@@ -32,7 +32,7 @@ Agents fetch this `README.md` to discover what's available, then `WebFetch` the
 | [`operating-archive-node-byov.md`](operating-archive-node-byov.md) | Required volume contents, PV/PVC spec, SeiNode/SeiNetwork spec, controller validation surface, and EBS-swap cutover sequence for archive nodes using the bring-your-own-volume (`dataVolume.import`) path. | Bringing up an archive node from a pre-populated EBS; swapping the underlying volume of an existing archive PV; debugging `ImportPVCReady=False`; receipt-store pruning concerns. |
 | [`migrating-validator-to-byo-secrets.md`](migrating-validator-to-byo-secrets.md) | Cutting a live validator from a legacy host onto the platform carrying its consensus identity via Secrets (`signingKey`/`nodeKey`): what migrates, SeiNetwork spec, controller validation surface, the stop-before-start double-sign discipline + layered equivocation defenses, cutover/rollback sequence, and dry-run gotchas. | Migrating an existing validator (e.g. arctic-1 node-19) off EC2 onto K8s; any cutover where a consensus key changes hosts; understanding the `replicas:1` CEL guard or the double-sign alerts. |
 | [`validating-flatkv-memiavl-parity-via-sharded-replay.md`](validating-flatkv-memiavl-parity-via-sharded-replay.md) | flatKV↔memIAVL storage-engine parity validation by differential historical replay: the flatKV+memIAVL replay-pair topology (same binary, same snapshot, blocks from a shared archive), the correctness gates (compare pair-not-archive; verify migration complete so flatKV reads are genuine; `historical_replay` build for pre-v6.5 txs), the seictl shadow comparator, result aggregation + Notion report, and the fan-out to 50+ shards. | Standing up a flatKV-vs-memIAVL correctness validation on harbor; driving a sharded replay campaign; debugging why a replay node is stuck or a comparison reads all-indeterminate/vacuous; understanding `migrate_evm` vs `memiavl_only`/`evm_migrated`/`flatkv_only` read routing. |
-| [`benchmarking-evm-gas-vs-execution-time.md`](benchmarking-evm-gas-vs-execution-time.md) | Testing whether Sei's EVM gas schedule tracks real compute, for both (1) replaying a real historical block against restored state and (2) executing arbitrary bytecode against a specific historical state: why hardware parity with prod validators is automatic (`NodepoolForMode`, no engineer config needed), the replayer-node mechanism for functionality 1 and its real gap (gas is captured, execution time is not), the unverified `eth_call`-state-override dependency for functionality 2 and its private-writable-fork fallback, the confounds that invalidate a naive comparison (governance-mutable `SSTORE` gas, precompiles, OCC re-execution/determinism), the differential-loop bytecode-crafting technique for per-opcode isolation, the privileged-attach gate for eBPF on harbor's `eng-<alias>` namespaces (PSS posture, scoped `kubectl debug --target=`), and concrete probe design against real `seid`/go-ethereum symbols (why per-opcode uprobing doesn't work, why `offcputime`/`profile` beat raw uprobe timing). | Designing a gas-vs-compute benchmark; replaying a historical block for timing rather than storage-parity; deciding whether an opcode's gas cost is a fair compute proxy before trusting a result; attaching any eBPF probe to a harbor pod; resolving `seid`'s real symbol names before writing a probe. |
+| [`benchmarking-evm-gas-vs-execution-time.md`](benchmarking-evm-gas-vs-execution-time.md) | Testing whether Sei's EVM gas schedule tracks real compute, for both (1) replaying a real historical block against restored state and (2) executing arbitrary bytecode against a specific historical state: the validated RPC-native measurement primitive (`debug_traceTransactionProfile` gives clean per-tx gas + execution time in one call, no eBPF required — replacing an earlier eBPF-first draft), the version-pinning discipline a replayer needs to avoid an app-hash-mismatch halt (including the `mock_chain_validation` build mode for testing unreleased code against real transaction load), why hardware parity with prod validators is automatic (`NodepoolForMode`, no engineer config needed), the scope limits confirmed by reading the actual code (only `Y=X-1` is supported for functionality 1; functionality 2's `debug_traceCall` path has a real execution-time gap), the confounds that invalidate a naive comparison (governance-mutable `SSTORE` gas, precompiles, whole-block-latency being the wrong timing proxy), the differential-loop bytecode-crafting technique for per-opcode isolation, subtracting intrinsic gas for a cleaner correlation, reading whole-block execution time from the existing log line with no new instrumentation, and eBPF (§9) demoted to a fallback tier for needs beyond what's specified. An appendix documents OpenTelemetry tracing as an investigated-and-invalidated avenue for this same measurement, with the real data that ruled it out. | Designing a gas-vs-compute benchmark; replaying a historical block for timing; choosing between a release-pinned replayer and `mock_chain_validation`; deciding whether an opcode's gas cost is a fair compute proxy before trusting a result; understanding why `Y=X-1` is the supported case, not the general one; attaching any eBPF probe to a harbor pod only if the RPC primitive genuinely isn't enough. |
 
 ## Adding a new runbook