[samcm]

Follow-up to #946. Reth sentries write about 80% of mainnet mempool_transaction_removed rows, almost all queued-pool dropped rows at 3–4 rows per hash, because queued transactions leave reth's txpool_content and come back. Mainnet over 3 hours of v1.32.0: about 527k rows/hour, of which 515k are drops and 441k are from reth.

Change

• A transaction missing from one txpool_content snapshot is held. It is reported only if the next snapshot is missing it too and it was not included in between. event_date_time stays the first snapshot without it.
• A reported hash is remembered for 2 hours, so a transaction that keeps leaving and returning is reported once per sentry. In production, reth re-admits a dropped queued transaction a median 8 minutes and a p99 46 minutes after the drop.
• The wiped-pool and over-half-vanished guards also clear held transactions, so a restart's losses aren't reported on the next poll.

Verification

• Unit tests:
  • Reworked for the two-poll rule.
  • New flapping test: a transaction missing from one snapshot then back is not reported; one missing then included is not reported; one reported, re-admitted and gone again within 2h is not reported again; after the TTL it can be.
  • Removing either rule (the reported-hash memory, or the "came back" check) fails it.
• Against real geth --dev and reth --dev pools with --txpool.lifetime 20s --txpool.nolocals (temporary test, not committed), with a fresh filler transaction each poll so expiry never empties the pool:
  • A queued nonce-gap transaction was evicted and reported once as dropped. It was then re-sent, evicted again, and not reported again.
  • A fee-bump replacement was reported as replaced with the right replacement hash.
  • The included transaction was never reported.
  • No hash was reported twice, on either client.

[redpandabot]

This follow-up delays removal reporting until a transaction is missing from two consecutive txpool snapshots and remembers reported hashes for two hours to suppress flapping re-reports. The held/confirmed bookkeeping, inclusion check, replacement lookup, and guard resets are consistent, and the reworked tests cover the two-poll rule, flapping, inclusion, and TTL expiry. I found no blocking correctness issues.

---

_{Reviewed 2 changed file(s) @ 6c4bfba5 — no blocking issues found.
"Don't drink the bear's skin before you've shot it." — Hungarian}

[redpandabot]

CI failed @ 6c4bfba5

One failed job: govulncheck / govulncheck. It fails on three vulnerability advisories (GO-2026-6888 brotli v1.2.0, GO-2026-6865 klauspost/compress v1.19.1, GO-2026-6859 net/http go1.26.10) that are not in the workflow's ALLOWED regex. The PR only changes pkg/sentry/execution/mempool_removals.go and its test, and go.mod/go.sum are identical to origin/master, so this is pre-existing; the same three advisories also failed govulncheck on unrelated branches (runs 42974332955 and 42970596458).

• govulncheck / govulncheck — pre-existing — govulncheck ./... reported three advisories not matched by the ALLOWED regex at .github/workflows/govulncheck.yaml:65: GO-2026-6888 in github.com/andybalholm/brotli@v1.2.0 (fixed in v1.2.7), GO-2026-6865 in github.com/klauspost/compress@v1.19.1 (fixed in v1.20.2), and GO-2026-6859 in net/http@go1.26.10 (fixed in go1.26.11). The PR diff touches only pkg/sentry/execution/mempool_removals.go and pkg/sentry/execution/mempool_removals_test.go; go.mod still pins klauspost/compress v1.19.1 (line 35), brotli v1.2.0 (line 81) and go 1.26.10 (line 3), identical to origin/master. The same three advisories also failed govulncheck on unrelated branches feat/mempool-lifecycle-events (run 42974332955) and feat/libp2p-goodbye (run 42970596458). (.github/workflows/govulncheck.yaml:65) — Bump github.com/klauspost/compress to v1.20.2 and github.com/andybalholm/brotli to v1.2.7 (go get + go mod tidy), and move go.mod:3, .tool-versions:1 and Dockerfile:1 to Go 1.26.11 for GO-2026-6859; if the build image cannot move yet, add GO-2026-6859 (and the other two IDs if the module bumps are deferred) to the ALLOWED list at .github/workflows/govulncheck.yaml:65. Re-running this PR's job alone will fail again.

---

_{Diagnosed govulncheck @ 6c4bfba5
"Don't drink the bear's skin before you've shot it." — Hungarian}