[krunkosaurus]

What this changes

Part 3 of the legacy CUDA stack; merge chunking #5, then sampling #6. Incremental source diff.

Accept canonical speaker tags 0..15 in requests, references, catalogs and the native writer, and carry the active role across long-form continuations, history, paragraph shifts and redraws. Original segment slices remain intact; added role text is counted at prompt admission. Existing malformed-tag, body/file/prompt and codebook limits remain enforced. The legacy formats have no new fixed reference-count cap or MLX-style three-speaker chunk cap.

The final cleanup removes the unused Redraws.salt helper requested in recipe #27. The recipe appends this as patch 0004, preserving the already merged sampling patch.

The public MLX prompt tests are a behavior reference. Modern source companion #1278 extends the corresponding current route. Native changes are independently implemented.

Validation

• Final source e06a0cda: zig build test-server-cpu -j2 --summary all passes 379 tests with four skipped on macOS arm64, Zig 0.17.0.
• Linux CUDA build and test on GB10: 106/106 steps succeeded; 870 tests passed, nine skipped. The first source archive lacked Git metadata needed by a version-generation step. Restoring metadata for the exact published commit made the unchanged code pass; the initial failure receipt is retained.
• GB10 NVFP4 stopping diagnostic: two synthetic references, five tagged catalog requests and five untagged controls all stopped normally. Mono 44.1 kHz, signed 16-bit PCM; diagnostic seeds 1–5, temperature 1.0, top-p 0.9, top-k 30, default CFG gamma 1.5 batched. Tagged output was 3.02–4.78 seconds of audio and rendered in 2.30–3.43 seconds. Settings and individual results.
• The composed HTTP regression exercises protected hard cuts, carried speaker/CFG text, explicit sampling and an exact seed above 2^53 for whole and streamed replies. Reference, continuation, history and malformed-tag cases retain their coverage.
• All four exported recipe patches reproduce this source exactly across 3,503 files; checksums and manifest coverage pass.

Speaker-role preservation does not establish distinct or faithfully cloned acoustic voices. No listening or performance qualification is claimed. The separate full-model stopping limitation is documented and tracked in recipe issue #32.

Pinned-engine packaging and user documentation: recipe PR #28.