Behavior is defined by the execution contract and market-data guide. Runnable commands and focused regression groups are in the test guide.
Build against the fork revision pinned in installation. It combines official develop with inventory-unit commissions, exact gross/order scale alignment, opt-in unresolved cancel/modify retention and terminal quantity normalization. Native owns Cache, reports, order/position application and events; the adapter owns protocol and raw venue evidence. See quantity commissions.
| Changed boundary | Required evidence |
|---|---|
| Public trades/depth | Atomic conversion, direction/match types, reconnect/ownership, Strategy callbacks and catalog rollover readback |
| Signing and command parameters | SDK vectors, exact signed amounts and local request assertions |
| Fills, fees and inventory | Recorded settlement replay, gross/net accounting, duplicate suppression and subscriber-visible state |
| Recovery | Pending commands, late fills, terminal normalization and replay controls |
| Session/transport | Loopback renewal, ACK, reconnect and cancellation tests |
| Documentation/fixtures | Offline manifest verification; no funded venue run |
Ordinary tests use local services and saved inputs. Named ignored probes require their documented scope. An offline replay does not establish live venue latency, partial-tail behavior, long outages or general FOK enforcement. SDK acceptance alone does not establish venue matching semantics.
The rolling-selector tests exercise both folds of the New York fall clock change, nonexistent spring times, year rollover, and retry when a successor has not been published. The loopback test checks subscription handover and registry cleanup. See rolling windows for the maintained timing rules.
Sanitized regression inputs are indexed by
evidence_manifest.json.
Keep original primary fixtures immutable. Synthetic schedules and locally rebound
identity/time are derived replay, even when economics come from real fills.
Historical captures apply to their recorded revision and do not validate a newer
binary. Local operational captures belong in ignored data/ and can be removed
without deleting versioned regression fixtures.
Record source revision, features, lockfile, logging and harness settings. Compare venue timestamps, ingestion times and strategy callback times separately. Source cadence is not transport latency; relays add forwarding overhead. Exclude compilation, fixture setup and concurrent benchmarks from timed samples. Native event logging is part of callback latency and must be held constant between runs.