test(recc): align seam tests with runner rollback

[castrojo]

Summary

• Classify the missing Dakota RECC seam as a deliberate rollback, not a regression.
• Update the seam tests to cover the documented outer-remote-execution path: production lanes retain the shared helper mount but do not invoke nested RECC or probe the unavailable runner socket.

Evidence

• 6e932b442 (feat(recc): wire the fail-closed runner seam) added the mandatory overlay invocation and buildbarn-config admission probe.
• 1004110a0 (fix(bst): roll back unsatisfiable nested-RECC admission gate) explicitly removed both from Dakota, Cosmic, Bluefin Server, BST QA, and cache warmup because the pinned bb_runner cannot consume remoteApisSocketPath.
• Live verification shows BuildBarn is running (frontend, scheduler, worker, and bb-remote-asset pods) and buildbarn-config exists, but the deployed runner config has no remoteApisSocketPath field.

Dakota still uses outer BuildStream remote execution and BuildBarn artifact/cache endpoints. It is currently missing only the nested RECC compiler-cache overlay by design. scripts/apply_recc_overlay.py is not dead code: it remains embedded in the shared ConfigMap and is invoked by the operator-only recc-baseline-pipeline; it is simply not invoked by production lanes until a capable runner is deployed.

Validation

• just lint — passed
• argo lint --offline argo/workflow-templates/ — passed
• pytest -q tests/unit/test_recc_runner_seam.py — 13 passed
• pytest -q tests/unit — 385 passed, 4 pre-existing unrelated failures (no new failures)

[castrojo]

Superseded by #654, which landed these exact test changes as part of unblocking main.

main had become un-mergeable: every open PR failed the required test-validation check on the same 7 pre-existing baseline failures, and each PR fixed only a subset — so none could go green on its own. #654 collected every test-side fix into a single change to break that deadlock.

Verified: every file in this PR is now byte-identical to main, so there is nothing left to merge. The analysis here was correct and is preserved in #654's history — closing only to avoid a no-op merge commit. Thanks!