Wire text/image/audio embeddings on the v2 transpiler; add /v1/embeddings

[ncylich]

Summary

Brings the three embedding entry points in Model up on the v2 transpiler. The refactor had left them as stubs: get_embeddings returned {} silently (so RAG/tool-RAG callers in rag.cpp fell through to dimension-mismatch warnings instead of failing), and get_image_embeddings / get_audio_embeddings threw "not wired up yet" even though the vision_encoder / audio_encoder components were already built and exercised by completion and transcription.

This PR:

• Ports nomic-embed-text-v2-moe (the same model main shipped) onto the v2 transpiled-bundle architecture and turns get_embeddings into a real encoder run, verified against the HuggingFace reference.
• Adds an OpenAI-compatible POST /v1/embeddings endpoint.
• Implements image/audio embeddings and fixes a latent vision-encoder family bug.

Text embeddings (nomic-embed-text-v2-moe)

Transpile

• New text_embedding task + NomicTextEmbeddingAdapter, an export-friendly reimplementation of the nomic-bert encoder. It reuses the HF submodule weights but reimplements the forward to survive torch.export: rotary is recomputed from seq_len (no lazy cos/sin cache), the additive attention mask uses scalar sub/mul (no rsub), and the MoE is dense — two fused matmuls over the packed expert weights gated by a top-k softmax built from iterative amax/masking (torch.topk, index_add_, and slicing a quantized weight do not lower). The graph emits last_hidden_state; pooling/normalization happen in the engine.
• Routing: _family_key detection, canonicalize_model_interface + build_component_module_specs dispatch, the text_embedding component in the bundle manifest, task auto-inference (component_plan.py, hf_model.py), a fixed-length input builder, and a loader that forces trust_remote_code (the transformers-native nomic_bert is a different architecture and won't match the converted weights).
• canonicalize/cleanup.py: stop fp16-legalizing the embedding op's weight input — the kernel dequantizes CQ/FP16 weights directly, and the inserted precision_cast tried CQ4 → FP16 at runtime and failed.

Convert

• NomicAdapter now emits one fused tensor per HF parameter (Wqkv and experts.mlp.w1/w2 are no longer split into q/k/v or per-expert files) so the transpiled graph — which binds weights by HF name — gets a 1:1 match. experts.mlp.w2 is stored transposed so the second expert matmul consumes it as a direct linear weight. The tiny MoE router stays FP16 (4-bit quantizing [num_experts, hidden] corrupts routing).
• Tokenizer conversion handles Unigram (XLM-RoBERTa, used by nomic): classify it as SentencePiece, emit per-token Viterbi scores into vocab.txt, and write the unigram runtime config (sp_model_type, sp_add_dummy_prefix, metaspace). It was previously misdetected as BPE and mis-tokenized ("Paris is the capital of France." → 31 tokens vs the reference 9).

Engine

• Implement Model::get_embeddings for transpiled bundles: load the text_embedding component, wrap tokens with BOS/EOS (matching add_special_tokens), run the graph, mean-pool over real tokens, optionally L2-normalize. Add an embedding-only init path (no decode route) and map model_type bert/nomic → ModelType::NOMIC so the config loader stops requiring Gemma4 fields.
• sp.cpp: the SentencePiece (non-BPE) path honors sp_add_dummy_prefix, prepending the metaspace marker so unigram segmentation matches the reference.

Server

• POST /v1/embeddings (string or list input) backed by cactus_embed, plus EMBED_MODEL_TYPES. create_app can now serve a non-LLM bundle as its default model, so an embedding-only server boots.

HF parity

cactus vs HF cosine 0.92–0.94. The gap is purely CQ4 4-bit weight quantization: FP parity of the adapter is ~1.0, and cactus-vs-HF is identical whether HF runs fp16 or fp32. Retrieval ranking is preserved (query·relevant 0.72 ≫ query·unrelated 0.29). Convert/transpile produces a text_embedding component with all 146 weights bound.

Image / audio embeddings + vision-encoder fix

• Model::get_embeddings now throws (mirrors the sibling methods) instead of returning {}; surfaces the failure to cactus_embed and cactus_rag_query (both already wrap in try/catch).
• rag.cpp: wrap the four unprotected get_embeddings callsites in retrieve_rag_context / select_relevant_tools with try/catch so chat completion degrades gracefully (empty RAG context / unfiltered tool set) instead of throwing out.
• Implement Model::get_image_embeddings / Model::get_audio_embeddings: run the existing encoder component, dequantize FP32/FP16/INT8 → FP32, mean-pool over leading dims into hidden_dim, L2-normalize. Both call load_component_graph after extracting the output so transcribe_whisper_seq2seq / transcribe_parakeet_tdt keep working when an embed precedes a transcribe.
• Make Model::run_vision_encoder family-aware. Was Gemma4-only — unconditionally wrote pixel_values + pixel_position_ids, producing NaN outputs on LFM2-VL (expects pixel_attention_mask) and Qwen3-VL. Dispatch now mirrors run_chunk_prefill_path (LFM2-VL / Qwen3-VL / default), with precision-aware writes instead of raw write_bytes_input. Also fixes a latent bug on the legacy lm_encoder_media_step fallback path for non-Gemma4 multimodal models.
• python/cactus/bindings/cactus.py: cactus_embed / cactus_image_embed / cactus_audio_embed passed the element count (4096) as buffer_size, but the C side treats it as bytes. Masked when hidden_dim ≤ 1024; broke Gemma4 (1536) and Qwen3-VL (2048). Fixed by passing ctypes.sizeof(buf).

Tests

• python/tests/test_nomic_text_embedding.py: task/family routing.
• python/tests/test_model.py: TestNomicEmbedding (determinism, retrieval discrimination, HF-parity cosine assertion); image/audio embedding tests un-skipped.
• python/tests/test_server_live.py: live /v1/embeddings (string, list, rejects non-embedding models).
• Redistribute the convert NomicAdapter tests out of the catch-all test_nomic_adapter.py into their topical homes (test_policy.py, test_naming_qdq.py), move the unrelated LFM2 tests into a new test_lfm2_adapter.py, and delete test_nomic_adapter.py.

Verification

Model	Embed dim	Norm	NaNs	Completion/Transcribe after
nomic-embed-text-v2-moe (text)	768	1.0	0	HF cosine 0.92–0.94
LFM2-VL-450M (image)	1024	1.0	0	✅
Gemma-4-E2B-it (image)	1536	1.0	0	✅
Qwen3-VL-2B-Instruct (image)	2048	1.0	0	✅
Whisper-small (audio)	768	1.0	0	✅

Test plan

• pytest python/cactus/convert/tests/ python/tests/test_nomic_text_embedding.py — green
• pytest python/tests/test_model.py::TestNomicEmbedding -s — determinism, retrieval, HF parity
• pytest python/tests/test_server_live.py — 19 pass (chat, transcription, embeddings)
• cactus convert nomic-ai/nomic-embed-text-v2-moe → text_embedding component, 146 weights bound
• Manual: image embed + completion on LFM2-VL, Gemma-4, Qwen3-VL; audio embed + transcription on Whisper-small