[palango]

Since the rebase onto op-node/v1.19.9 no op-acceptance-tests job runs in CI: upstream removed its op-geth variants and we disable the op-reth and kona ones, which run binaries without Celo's changes. This adds memory-all-opn-op-geth back, the job celo-rebase-18 had, and makes ci-gate wait for it. It restores less than the name suggests, because the devstack now skips every chain that schedules Karst when the EL is op-geth, and Karst is the default fork.

What that means in practice, from a local run of seven test directories on op-geth: 7 tests passed and 24 skipped, every skip from the guard in op-devstack/sysgo/deployer.go:546-555. The Ecotone FeeHandler assertion (op-acceptance-tests/tests/ecotone/fees_test.go) runs and passes. The Fjord one from #412 skips, and so do all twelve base/conductor tests. A green run of this job says the Jovian-and-older tests pass on Celo's op-geth and little more.

Full coverage needs celo-reth as the test EL, tracked in celo-org/celo-blockchain-planning#1462. I decided against making the devstack default to Jovian for op-geth: that's fork-only test plumbing for an EL we're retiring. Karl raised the gap on #541 (#541 (comment)).

The first CI run (39 minutes, 274 results: 46 passed, 217 skipped, 11 failed) failed on upstream's ZK proposer, which Celo doesn't run, so the second commit skips those tests:

• TestZKDisputeGameWithdrawal in tests/proofs/super: kona-sp1-proposer re-executes the block holding the withdrawal transaction and gets a different output root than Celo's op-geth ("Failed to validate L2 block build(deps-dev): bump semver from 5.7.1 to 5.7.2 #33 with claimed output root …"), so the game is never proven. Skipped with gt.Skip, like the fork's other fault-proof skips.
• tests/proofs/zk: on one node the package runs serially and passes the 30-minute package timeout; 15 tests passed, one was cut off and nine never started. Skipped as a whole through a TestMain in a new celo_test.go, so nothing in the upstream files changes. Its tests then don't show up as skipped in the report; the package prints one line saying why.

Everything that passed locally passed in CI, proofs/cannon_kona included, and the Ecotone FeeHandler test passed while the Fjord one skipped, as expected.

For review:

• The job requires rust-kona-binaries and rust-op-reth-binary directly, since rust-binaries-for-sysgo is disabled and the job exits early unless op-reth and op-reth-sdm-fixture exist.
• It runs on one celo-org/2xlarge node (the job definition on this branch ignores parallelism) and blocks every non-docs push. Without the ZK tests it should take well under the 39 minutes of the first run; the second run will give the number.
• Which Celo rule makes kona's output root differ isn't established. Base-fee routing to the FeeHandler is the obvious candidate, but no log line says so.

The ci-config-reviewer pass found nothing blocking; I dropped the discord context it flagged as unused. The merged config passes circleci config validate and circleci config process. .circleci/scripts/test-decision-tree.sh doesn't run under macOS bash 3.2, so that check is unrun; this change doesn't touch routing. I didn't run the Go review agent on the two-line skip commit. The base is palango/celo-rebase-19 because celo-rebase-19 isn't published yet.

[palango]

Squashed into the rebase branch as 1cb6f33 (#541), same tree as this PR's head. Closing this one.