[michalharakal]

Summary

A Laya checkpoint is an encoder plus a small decision head, and the head is what gets fine-tuned for a domain: layaMOE (Apache-2.0) freezes the laya-typed-decisions encoder, copies the head, trains one per domain group (safety, customer_ops) and lets the base head route each request by answering "What kind of text is this?". sk.ainet.models.laya could load one head per checkpoint only. This PR lets any number of heads share one encoder, loads a head on its own, routes, and compares — all in commonMain, every signature additive.

• Head-only checkpoints: LayaWeightPreparation.prepareHead (the head/scorer/type-embedding branch of prepare, now shared) and LayaDecisionModel.buildHead bind an expert.safetensors (type_emb.*, head.*, scorer.*; act_head.* ignored) into a LayaDecisionHead; the checkpoint's own head is LayaDecisionModel.defaultHead. A head prepared on its own reproduces the base model bit for bit (test).
• Heads on encoded states: LayaDecisionModel.logitsFrom / logitsBatchFrom / encodeBatch(sequences, padId) and LayaAgent.encodeFor / logitsFrom / decodeAll run the encoder once per question and any head on the result; the existing hidden / logits / logitsBatch / logitsFor delegate to them, numerically unchanged.
• Experts: LayaExpertMeta (expert.json: name, domains, calibration temperatures), LayaExpertLoader.load(metaJson, weights, base, ctx, weightFormFor) → LayaExpert with a LayaDecoder carrying the expert's temperatures over the base agent's budgets (LayaWeightForms.q8_0For packs the head's projections like the base's).
• Routing: LayaRouterSpec(question, options, experts, threshold) — the prompted router (kinds with descriptions, kind → expert, inclusive threshold, general fallback when the kind is unmapped, the expert is not loaded or the top probability is too low); LayaMixtureAgent.decide(state, questions, expert, compare, route) routes (one extra encoder pass) and answers with the chosen head, reports a forced head as overridden, adds the base head's answers on the same encoder output with compare, and carries encoder / head milliseconds and the token count. Sequence-level routing, one head per request — the docs say it is not a sparse mixture of experts.
• Typed contract: LayaMixtureDecisionAdapter puts a mixture behind DecisionAdapter (routed once per request, answered with the chosen head's calibration, onRouted reports the head since the contract has no field for it). LayaDecisionAdapter keeps its API; the packing and outcome code moved to an internal LayaAdapterSupport both adapters use.
• ModernBERT-large loads: LayaEncoderConfig.requireSupported() accepts position_embedding_type: "absolute" beside sans_pos — the transformers 5 export of convaiinnovations/laya-typed-decisions writes absolute while the positions are rotary either way. The english section of the Laya sequences.json fixture (ModernBERT-large ids) now has a gated test.
• Docs: Expert heads and routing in typed-decisions.adoc, README row, CHANGELOG.

Verification

• jvmTest (no checkpoint): LayaMixtureSpecTest (routing rules, inclusive threshold, override, expert metadata, per-expert temperatures), LayaExpertSyntheticTest (head-only preparation reproduces the base model exactly on the single and the batched path; rejects encoder tensors, missing parts, wrong width; mixture routes / overrides / compares; the mixture adapter answers like the plain adapter when the routed head is the base head's twin), plus the existing synthetic and builder suites. compileKotlinWasmJs, compileKotlinJs, compileKotlinLinuxX64 of the module.
• Gated on the checkpoints (LAYA_TYPED_DECISIONS_PATH, LAYA_MOE_EXPERTS_PATH, LAYA_MULTILINGUAL_PATH): LayaMoeEvalTest replays two cases per domain (18 cases, 52 answers, compare on) — routing identical to the published run on 18/18, 612 values (router probabilities, expert and general answers, scores, noul, confidences) within 5e-5 of the PyTorch FP32 numbers (their 4-decimal rounding), base + two experts loaded in 13.5 s; sequences.json english 26/26 and multilingual 26/26; tokenizer corpus 40/40; predict.json 55 values max 4.99e-5; LayaDecisionAdapterTest unchanged. Cost on a 12-core laptop, FP32 eager, engine 0.61.0: about 39 s per case (four to five ModernBERT-large encoder passes), the heads 2–5 s of it.
• apiDump (additive), jvmApiCheck.