[TheTom]

A clean public-wheel runtime reached its first audio normalization call and failed because FlashInfer's CUDA JIT could not find Ninja. Pin Ninja 1.13.0 (the working container's version), expose the venv binaries during setup, document activation during inference, and check Ninja execution alongside CPU imports. The unchanged Torch 2.13.0+cu130 / TorchAudio 2.11.0+cu130 environment passed that expanded aarch64 CPU check; the corrected clean environment then generated EOS-complete speech on Spark with verified source assets and with a separate fresh compact --runtime-manifest load. The compact 85-frame WAV was byte-identical to the earlier full-source C0 WAV. This is dependency/runtime evidence, not a quality or parity pass; a separate native SM120 / WSL2 RTX 5090 cold/warm hello smoke also completed with EOS (47 frames, 2.183 seconds audio). Warm request-to-WAV was 2.142 seconds; total device memory sampled at 100 ms peaked at 31787 of 32607 MiB. This is short-smoke fit only, with 820 MiB sampled headroom. The guide also preserves the unconditioned and reference-conditioned C0 4096-frame long-form failures.

When generation reaches its frame limit without EOS, the failure receipt also retains the prefill breakdown, generation wall time and prompt row count. It still refuses a complete WAV and reports no audio duration or RTF. Current failed profiling attempts retain their original traces and frame codes.

Validation: 17 Drama-3 CPU tests passed; executed the actual frame-limit branch with synthetic frames to confirm nested timing preservation, failed-frame output and absence of a complete WAV; shell syntax checks passed. No generation arithmetic, sampling or frame limits changed. Related #398; follows merged #399.

[TheTom]

Release/runtime progress (2026-10-14): C0 and C24B are now public together in the existing model repository. The C24B finalizer verified all 414 checksum entries and unchanged C0/shared-runtime identities. Independent anonymous metadata access then confirmed public visibility and the exact 24,274,450,464-byte GGUF/SHA256. The card includes fork setup/build and exact run commands.

C24B produced four EOS-complete WAVs with exact independently decoded ASR content. Its tested 24GiB-budget route requires --gpu-layers 32 (31 backbone blocks plus head on GPU, nine backbone blocks on CPU). This was a constrained 5090 simulation, not physical 24GB-card validation. The emphasis request failed at 1024 without EOS and produced no complete WAV; that evidence is retained. No broad quality, real-time, or source-parity claim.

A separate all-GPU/context512 cold+two-warm comparison succeeded after exposing the existing venv binaries in PATH; the earlier missing-Ninja failure was retained. This supports the documented runtime PATH requirement without changing this PR's generation arithmetic or limits. Butter's later native implementation and timing work are separate; see Butter #1069.

This is a validation/status update; the PR head is unchanged and no merge is requested or performed. AI-assisted testing/reporting.