[leehack]

Adds opt-in LFM3-Audio and LFM3-Omni speech output to the TTS wrapper, for core llamadart 0.18. Refs #130.

Draft. Additive only: the 89 existing required exports keep their behaviour, and llama_dart_tts_get_info / llama_dart_tts_init still refuse LFM3-Audio. No submodule, tag, release or workflow-dispatch change. On main, it ships in the next scheduled native publish.

What upstream v0.8.0 already supports

Variant	Upstream	Wrapper change
Text (LFM3-1.7B)	LLM_ARCH_LFM3 (#31252)	None
VL (LFM3-VL-2B)	lfm3vl projector	None (generic mtmd exports)
Audio input / ASR (Audio, Omni)	lfm3a encoder (#31257)	None
Speech output (Audio, Omni)	MTMD_GEN_AUDIO_TYPE_LFM3A	This PR
d1 (d1-2B, d1-Omni-2B)	non-causal LLM_ARCH_LFM3, lfm3-d1(-omni) decision types (#31275)	None (scoring is the consumer's; no KV memory, so the whole prompt goes in one batch)
Video	mtmd_helper_video_* (#31283), needs the ffmpeg binary	Not wrapped

No piece needs an upstream patch.

Changes

• New exports, in their own block: llama_dart_tts_supported_model_types (build probe), llama_dart_tts_get_info_ext and llama_dart_tts_init_ext (a mask of the model types the caller handles), and llama_dart_tts_start_with_voice. New enum values: LLAMA_DART_TTS_MODEL_TYPE_LFM3_AUDIO = 2 and LLAMA_DART_TTS_CAPABILITY_VOICE.
• The new generator needs more than an allowlist entry. request.text is rendered through the model's tts chat template (common_chat_get_tts_prompt, with the language and voice). The text head is not sampled, so a text EOG cannot end the speech. Speaker references are refused, because upstream would ignore them. Output is complete PCM only.
• CMake probe LLAMADART_MTMD_HAS_LFM3A_GEN. On v0.7.0 the probe is off, so the build reports Qwen3 only.
• Tests: tts_types_test (new), extended tts_api_test, and a local llamadart_tts_lfm3_smoke that every wrapper-contract lane compiles. validate_exports.py now requires 93 exports.
• Docs: docs/lfm3_support.md, README, docs/v080_qualification.md, AGENTS.md.

Local evidence (macOS arm64)

Check	Result
Wrapper contract build (Debug, CI flags) + ctest, v0.8.0	58/58
Same on pinned v0.7.0 (LFM3A probe off)	58/58; LFM3 smoke exits with "this build cannot drive LFM3-Audio"
validate_exports.py v0.8.0 / v0.7.0	93 / 93
Failing-first	tts_api_test failed to link on the four missing exports before the implementation
llamadart_tts_lfm3_smoke, LFM3-Audio-2B-GGUF 02355c1a: Q4_K_M (9a773289…) + mmproj Q8_0 (2b6648b1…), 1.4 GB	PASS on Metal and on CPU. The ASR round trip returns "Hello, how are you today?". Seed repeats after cancel and reuse; voice, language and seed change the PCM; cancel lands at frame 3; max_frames truncates
Qwen3-TTS 1.7B Q4_K_M tts_smoke (Metal), regression control	PASS
Mutation (16 mutants, unit tests + smoke)	14 killed, 2 survived (below)
Python suites / actionlint	279 OK + 22 OK; only the known windows-11-vs2026-arm label

Notable kill: restoring Qwen3's text-head sampling and EOG stop for LFM3 (the allowlist-only change) cuts the speech, and the transcript drops to "Hello,".

Survivors:

• The init-time tts template check: no local model lacks the template. Without the check, start fails at render with UPSTREAM_ERROR.
• The build-mask term in the accepted-type check: equivalent and redundant, so it was removed.

Not covered

• LFM3-VL, LFM3-1.7B, Omni and d1 were not run. The smoke uses the LFM3-Audio backbone (arch lfm3) for text decoding and audio input.
• Not exposed upstream features: streaming PCM, Omni mixed text/audio and sequence continuation, live ASR (mtmd_audio_stream_*), and video.
• The tts prompt is tokenized with special-token parsing, as in upstream llama-tts, so control tokens in request.text are parsed as tokens.
• Not qualified: Linux, Windows or Android devices, and the consumer bindings.
• High-risk independent audit not run.

[leehack]

Independent audit at exact head a378dce: no P1 or P2 findings, and 64/64 checks pass. merge-tree against current main 52303ee (#140) is clean, and the merged tree passes CTest 58/58, the LFM3 smoke and the Qwen3 smoke.

• Additive only: exactly 4 new symbols, no struct changes, existing enum values unchanged, and TTS API version 1.
• Legacy calls unchanged: get_info and init still refuse LFM3 with type 255 and capabilities 0. Core 0.17 bindings compare model_type as an int, so companion 0.17.0 and core 0.17.1 are safe.
• Qwen3-TTS unchanged: its PCM is byte-identical between base and head on v0.8.0 and v0.7.0.
• Prompt injection: passing control tokens in request.text only changes the speech. The text goes in as a Jinja variable, so it cannot inject into the template.

Non-blocking P3s are tracked on #130:

• capability-assert gaps (mutants M08, M12, M14);
• an unknown LFM3 language is not rejected;
• get_info_ext reports OK when the model has no tts template;
• LFM3-Omni is claimed but untested;
• docs/v080_qualification.md still says 89 exports.

Evidence: llamadart-evidence/native-pr-139/.