The simulation suite passes all 33 scenarios, but that does not establish the same accuracy or convergence on real transports. Add a repeatable validation layer between the fluid simulator and public-server smoke tests.
Current evidence and limitations:
internal/linksim explicitly does not model TCP. It approximates rates, queueing, send-buffer credit, and capacity changes.- The simulator creates a foreign and a self sample together at each probe step, while production interleaves probes and records actual asynchronous completion and failure.
TestLive checks positive idle latency, throughput, and RPM; it does not assert accuracy against known link conditions or default-budget convergence.- Public live runs use Linux; six-target cross-compilation is not measurement validation on every supported OS/architecture.
Relevant code: simulator, simulated probes, live smoke test.
Acceptance criteria:
- Provide a reproducible harness running the real client and reference server over controlled bandwidth, delay, queue, and loss conditions, with explicit accuracy and budget tolerances.
- Cover upload send-buffer accounting, flow ramp/convergence, bufferbloat, capacity changes, and sparse/failed probes.
- Compare client estimates with independent observations of delivered bytes and path timing; do not use the simulator itself as ground truth.
- Record sanitized observations sufficient to calibrate the simulator and reproduce relevant engine decisions.
- Match simulator probe scheduling/completion more closely where evidence shows a material difference.
- Document a staged runtime-validation matrix across supported platforms, with unsupported environments reported as blocked.
- Keep ordinary tests offline and deterministic; separate public smoke tests from controlled accuracy checks.
Related: #20, #24, INV-3, LOAD-3 through LOAD-14, LAT-10, docs/guidelines.md algorithm-change requirements. This is a validation gap, not a claim that every current real-network measurement is inaccurate.
The simulation suite passes all 33 scenarios, but that does not establish the same accuracy or convergence on real transports. Add a repeatable validation layer between the fluid simulator and public-server smoke tests.
Current evidence and limitations:
internal/linksimexplicitly does not model TCP. It approximates rates, queueing, send-buffer credit, and capacity changes.TestLivechecks positive idle latency, throughput, and RPM; it does not assert accuracy against known link conditions or default-budget convergence.Relevant code: simulator, simulated probes, live smoke test.
Acceptance criteria:
Related: #20, #24, INV-3, LOAD-3 through LOAD-14, LAT-10, docs/guidelines.md algorithm-change requirements. This is a validation gap, not a claim that every current real-network measurement is inaccurate.