Excess Worker 0.2.2

Release sequence 30. Anonymous Minisign release signing. Windows executables have no trusted Authenticode publisher signature.

Changes since 0.2.1: streamed text is metered correctly when the pinned llama.cpp runtime withholds token IDs inside a multibyte character (seen with gpt-oss), so those jobs no longer fail.

Linux NVIDIA CUDA: the signed 0.2.2 candidate, with the same worker payload as this release, ran actual buyer jobs for all 17 pinned local catalogue models on one recorded NVIDIA B300 configuration (driver 580.178.04, CUDA 12.9). Drain, restart, revocation and cleanup passed there. GPU memory budgets are sampled whole-device limits, not hard partitions. Other GPUs, drivers and distributions are unverified. See docs/verification/linux-b300-cuda-sequence29.md.

Windows CPU/CUDA: the eight release-gate routes passed, plus Phi-4-mini and Qwen3-8B on CPU and CUDA. Linux CPU (WSL2): Qwen3-4B, Qwen3-8B and Qwen3-14B passed. See docs/verification/candidate29-other-models.md.

Updating: Windows and Linux 0.2.1 installations update with excess-worker update after draining. Linux 0.2.0 installations still upgrade by running the verified 0.2.2 installer in place. See docs/INSTALLATION.md ("Stop, revoke and update").