# Consumer qualification harness
`scripts/run_runtime_qualification.py` is a bounded, deterministic smoke
harness for the opt-in Preview runtime. It exercises a real child process,
handshake, decision, atomic state persistence, restart, delayed feedback and
duplicate-feedback rejection. Its JSON output is an audit input, not a claim
of real PM, Xray, Agent, or other downstream integration.
The eventual 1.5 qualification matrix has three distinct evidence classes:
1. repository-owned simulated consumer;
2. PM-style and network/Xray-style simulated consumers;
3. real external consumer repositories, when those repositories and their
credentials are available.
The harness keeps those classes separate in its result schema. Full synthetic
load, extended soak and fault-injection runs are deliberately post-push
activities and must not replace normal Actions or public-asset verification.
The final-stage runners are `scripts/run_runtime_final_qualification.py` for
the three simulated consumer shapes and `scripts/run_runtime_fault_smoke.py`
for deterministic bounded fault coverage. Every runner reports
`evidenceType: simulated`; none of this output is native or real-consumer
proof. The final qualification runner reports cold start, restart/restore,
snapshot and inspect timings, state-file growth, child RSS, decision and
feedback P50/P95/P99 across bounded batches, and a pending-cap rejection probe.
Its optional bounded soak repeats fresh state cycles so the completed-decision
ledger cannot grow without limit; this is still simulated evidence. The runner
uses batches below the default pending cap rather than weakening that cap.
`conformance/run.py --mode candidate` adds a candidate-artifact layer:
it records the binary SHA-256, checks the reported version, binds the supplied
commit identity, and then runs the bounded v3 qualification. A local path is
still candidate evidence, not public-release evidence; public-asset
qualification must be repeated after downloading the actual GitHub Release.