nx_torrent_capbench.nx

nx_torrent_capbench.nx -- MULTIVARIATE COMPETITIVE CAPABILITY BENCHMARK (S-class feature, B0-B2). module: nishi-core.torrent.capbench depends: nx_swarm_sim.nx (deterministic swarm model -> nx_ss_run; transitively nx_stream_picker, nx_piece_manager/nx_pm_rarest, nx_syscalls -- so this sits HARDWARE-UP on the proven sovereign substrate, like every other organ). capability: CAPABILITY_MEASUREMENT WHAT THIS IS (and why it's a first-class feature, not a one-off test): The way DeepMind measured agent capability with a SUITE of game environments (ALE/Atari, each probing a distinct skill) and a score normalized against a baseline, this is a suite of swarm SCENARIOS -- each probing a distinct piece-selection capability -- run against the system-under- test, scored on an ORGAN-COMPUTED maturity ladder vs the fastest known clients. It is built rung by rung, every rung gate-proved, so the ecosystem can SEE its own capabilities honestly. B0 deterministic measurement core : same scenario+strategy => bit-identical decision trace, so every measurement is REPLAYABLE (the property NO mainstream torrent client has -- they are clock/GC/scheduler driven). B1 multivariate matrix : SCENARIO x STRATEGY -> measured metric (rounds / DNF). B2 organ-computed maturity grade : ABSENT<PRESENT<PARITY<EXCEEDS<PHYS-OPT, COMPUTED from the measured numbers vs the baseline -- never self-asserted. (B3 incumbent-census wiring + B4 live wire-throughput = next rungs, honestly PENDING below.) HONESTY: the fastest clients (qBittorrent/libtorrent, rqbit, Transmission, aria2) all use rarest- first + endgame -- so NISHI matching that algorithm is PARITY by construction; the MEASURED win is endgame completing adverse/snub swarms where a no-endgame client DNFs; the EXCEEDS-all axis is determinism. Wire throughput is link-bound and NOT claimed here (PENDING-live).