fix(runner): verify the datadir head for a bundle inside the fixtures artifact - #306

Fixes #305.

verifyPreRunBundleHead exists to guard the replay's by-number skip — a datadir at the bundle's height but on a different chain would otherwise skip every line and benchmark the wrong state. It never ran for artifact-resolved bundles:

if src == nil || src.PreRuns == nil || src.PreRuns.LocalFixturesDir == "" {
    return false, nil
}

Since #296 a pre_runs source can resolve its bundle out of the extracted fixtures artifact using only fixtures_subdir, leaving local_fixtures_dir empty — so the guard returned early for exactly the configs that depend on that path. That gap came in with #296, which taught loadPreRunBundleSteps the fallback and left this second call site behind.

Change

Ask the source where its bundle is instead of re-deriving the path from config. The bundle may sit at a configured local directory or inside the extracted artifact, and only the source knows which — a second config-derived derivation is what silently missed the artifact case.

// optional, alongside the existing GenesisProvider idiom
type PreRunBundleLocator interface {
    PreRunBundleDir() string
}

• EESTSource.PreRunBundleDir() resolves it once and is now the single source of truth; loadPreRunBundleSteps uses it too, so the replay and the check can no longer disagree about which bundle is in play.
• Implemented as an optional interface checked by type assertion, so LocalSource, GitSource and ArchiveSource are untouched.
• ReadPreRunBundleInfoAt(bundleDir) reads a directory named outright, which is what a runner-side pre_runs source has. ReadPreRunBundleInfo(parent) keeps its parent + PreRunBundleSubdir form and simply delegates.
• The config-derived path stays as a fallback for callers without a live source.
• An unlocatable bundle now warns rather than passing quietly — the skip it guards runs either way, so silence there is the dangerous default.

No behaviour change for a local_fixtures_dir config: it resolves to the same directory by either route.

Why it matters

Two pre-runs of the same config produce the same end block number with different hashes:

compute  build 31183512437 : end_block=24410463  end_hash=0xf8115f05…
stateful build 31133523115 : end_block=24410463  end_hash=0x4525339a…

Running the compute suite against stateful-promoted datadirs, the replay reported success while skipping everything, and then every benchmark payload was an orphan — reth stuck on forkchoice updated message when syncing, nethermind 0 passed of 244 attempted. Hours of wall-clock per client, and it reads as a client bug rather than a wrong datadir. With this change the run stops up front with the datadir is on a different chain than the bundle was recorded against.

Testing

• TestVerifyPreRunBundleHeadFromFixturesArtifact — the regression: pre_runs with no local_fixtures_dir, bundle located via the source. Covers wrong-hash-now-rejected, matching-head-still-applied, and the no-locator/no-local-dir case. On master the first of these passes silently, because the early return means no verification happens at all.
• TestPreRunBundleDir — table-driven over local-wins, artifact fallback, default subdir, and both nothing-to-resolve cases, plus a compile-time PreRunBundleLocator assertion.
• TestReadPreRunBundleInfoAt — reads a bundle dir directly and asserts it agrees with the parent-form call.
• All pre-existing TestVerifyPreRunBundleHead subtests still pass; the config-derived fallback is what keeps them green, and dropping it broke them, which is how I found it was load-bearing.

gofmt, go vet and go test -race are clean for pkg/executor, pkg/builder and pkg/runner. golangci-lint run --new-from-rev=origin/master reports 0 issues for pkg/executor and pkg/builder; it cannot load pkg/runner in my environment (pkg/podman needs cgo gpgme, which also fails on pristine master here), so pkg/runner was built, vetted and tested with -tags containers_image_openpgp instead — CI will cover it properly.