Red Lite dev19 — user-facing native chat command

Dev19 connects the native end-to-end engine from dev18 to the installed Red Lite command line. Users no longer need to know the internal .deps/redmetal/redlite-generate build path to open a persistent conversation.

Command#

redlite chat --stats

redlite chat invokes the native generator with --interactive. The model is opened once; the DeltaNet recurrent state and full-attention KV cache remain live between turns. With no positional argument it uses the reference GGUF from the repository's models/ directory; a different path can still be supplied as redlite chat MODEL.gguf. The default profile is intended for the field-validated M4 Pro with 24 GiB unified memory:

• 4096 context positions;
• 4096 MiB routed-expert cache;
• 256 generated tokens per answer;
• temperature 0.7, top-k 40 and top-p 0.95.

Pass --temperature 0 for the greedy path that was compared token-for-token with the pinned llama.cpp oracle. --prompt supplies an optional first user message; --system supplies a system instruction. The native interactive commands remain /reset, /help and /quit.

Failure behaviour#

The launcher checks Apple Silicon and the model path before starting. If the native executable is absent, it reports make native as the recovery action. redlite doctor now reports native_redlite_generate independently from the pinned llama.cpp and oversized-runtime binaries.

Validation#

• Python parser tests cover defaults and all exposed options.
• A runner test verifies the exact native command without launching inference.
• A real-GGUF one-token smoke test opened the native Metal engine through redlite chat, completed the first turn and exited cleanly with /quit.
• The existing dev18 native build, engine parity and real-model generation tests remain the inference correctness gates; this milestone does not change engine arithmetic.

Scope boundary#

This command is a terminal chat client. It does not add a native HTTP server, batched prompt ingestion or new model/quantization support. redlite run and redlite serve retain their v0.2 behavior and continue to select the pinned upstream engines through the memory planner.