Repository navigation
TEST 127
Use a pinned fake Actor MCP server with opaque runs in capturing, finalizing, stopped, and failed states. Feed records with exact run/VM identities, sequence numbers, monotonic timestamps, frame hashes, and event links. Inject wrong server/revision/run/VM, stale cursor, duplicate/reordered records, partial frame writes, altered hashes, absent hashes, cursor gaps, overflow, and path/secret canaries.
Only ordered, complete, identity-bound records from the pinned server/run enter the session. Gaps, overflow, path leakage, malformed/partial frames, or stale/out-of-order records fail closed with explicit incomplete status. capturing, finalizing, failed, and unknown states cannot produce a final assessment. The consumer never reads Actor host paths or mutates capture lifecycle. A stopped batch following an active session cannot finalize provisional metadata; terminal metadata must be refreshed and admitted frames verified. scripts/tests/test_assisted_benchmark_active_capture.py tests pin/activation/base-digest drift refusal, legacy-default isolation, explicit desktop workload binding, deadlines/cancellation, duplicate keys, hashes/PNG structure, and pathless terminal errors without retries. These offline tests are integrated through the Actor and ProjectBuddy PRs; set PROJECTBUDDY_ACTOR_FIXTURE_SERVER to the pinned candidate's compiled stdio entry to enable raw/SDK fixtures for both LabVIEW and desktop workloads. SDK admission uses the explicit candidate manifest for Actor revision b408dd706fc8f85b478e4a49f9156e47ec65b6cd, package 1.5.0. Fixtures never start real capture or inference. Permitted hosted Windows tests pass, but native live qualification is not claimed and remains blocked on issue #257.