Repository navigation
Protocol v2: upgrade both endpoints together. Default QUIC Reno and compression=false remain unchanged. The separate Linux BBR package is experimental and opt-in.
Global stream/codec admission limits, adaptive compression, bounded overload, credential rotation/expiry, HTTP/3 CONNECT, and hardened UDP/connection lifecycle. The optional pinned BBR adapter includes idle/rate-step and MTU/pacing corrections.
Exact main source: ce3c7ce3207a21c918e0c50209b7693710a51006. Main CI39311899877 and release build39313236128 passed. Original archive/binary/source/patch/toolchain hashes remain unchanged in manifest.json and SHA256SUMS.
Additional portable checks passed all six jobs in CI39441638174 on 3021458c191f003c432e22626f73b2e9dd09e1c1. Only validation tools/workflows/docs differ from the measured release source. Includes Linux/Windows, real binaries, synthetic impaired links, shutdown/failure cleanup, single-CPU scheduling, long random transfers, race/vet/vulnerability/fuzz and pinned BBR tests. Exact downloaded release archives also have separate native Linux/Windows evidence in the acceptance archive.
pr4/pr5 comparison used these same archives in both directions: four profiles, text/random, three repeats, direct kernel TCP (host BBR), QUIC Reno and experimental BBR, compression off/adaptive. Capacities50Mbit/s native/100ms-no-loss/100ms-0.1%-loss and20Mbit/s at100ms/1%-loss. Generation16s on50Mbit/s profiles and8s on severe; goodput includes complete delivery/drain.240 fresh matrix transfers,12 mixed cycles, three overload/recovery groups.32 completed functional cases from the preceding100Mbit/s cohort use the identical files and are explicitly attributed, not presented as new executions.421 named real subcases have evidence, including those functional cases and cleanup/provenance observations.
Positive/runtime paths retained strict zero critical namespace UDP errors and no lab memory-limit/OOM events. Physical allowance/drop/error guards stayed unchanged during the completed cohort. Earlier uncapped/100Mbit/s attempts stopped conservatively on shared-host inbound allowance growth and are retained separately. Original service identities and host network settings were preserved; all owned lab resources were removed and independently audited. A separate deliberately overflowing Actions-only queue probe verifies error reporting; expected probe errors do not enter positive-suite counters.
Known limitation: an earlier100Mbit/s moderate-profile experimental BBR/adaptive random GET did not complete within120s. Targeted fresh/persistent repeats, expanded CI and the final50Mbit/s comparison passed, but its precise cause remains unresolved. This release does not claim a runtime fix for that event. Keep BBR experimental; default Reno is unchanged. The earlier four-client128MiB lab OOM was a stand capacity failure; the full comparison uses one active mode within the same budget, with negative evidence retained.
Compression can accelerate compressible content. Random/encrypted, clean or CPU-limited traffic can be slower than direct TCP. No universal throughput/fairness guarantee. Physical mobile migration and multi-day stability remain unproven; this task did not deploy production services.
Complete comparison | Test plan and Actions mapping | Real-channel register
Attached acceptance archive includes raw measurements, resource observations, failed/interrupted cohorts, complete case ledger and checksums. ACCEPTANCE_ASSET_SHA256SUMS covers evidence attachments; binary SHA256SUMS is unchanged.
Final acceptance documentation was merged in PR #6. All six jobs passed for the final PR commit in CI39469631858 and for merged main in CI39471542187. The merged tree equals the tested PR tree; the measured release source and pinned BBR base remain ancestors. These checks do not replace the documented experimental BBR limitation.