refactor: collapse Llama/Qwen/Voxtral NetworkDef onto shared decoder body

[michalharakal]

Summary

• Each architecture's xNetwork(metadata) is now a thin (~5 line) caller of decoderTransformerNetwork from feat(llm-core): shared decoder transformer body + DecoderModelMetadata #109, instead of duplicating the transformer DAG or stub-delegating across modules.
• qwenNetwork() is no longer a stub — it now actually applies QK-norm and metadata-driven RoPE base / eps, the architectural difference from Llama that issue Wire QwenNetworkLoader into CLI for proper Qwen3 inference #46 originally called out.
• voxtralBackboneNetwork no longer imports llamaNetwork from :llm-inference:llama, removing the last cross-model import flagged by the no-model-duplication plan.
• Net diff: -33 lines, no behavior change for existing consumers.

Changes per file

• LlamaModelMetadata: now implements DecoderModelMetadata. Override-only diff; field names were already aligned.
• llamaNetwork: collapsed from ~50-line sequential{} block to a thin decoderTransformerNetwork(..., qkNorm = false) call. Same module tree as before; eps and ropeBase now flow through metadata defaults.
• qwenNetwork: was a 3-line stub delegating to llamaNetwork; now a real function exposing qkNorm: Boolean = true. Per-architecture knobs (ropeFreqBase, rmsNormEps) propagate via metadata.
• QwenNetworkLoader: auto-detects QK-norm from *.attn_q_norm.weight presence in loaded weights — same pattern Gemma uses. Real Qwen3 GGUFs always carry these tensors; synthetic test fixtures don't, so both keep working.
• voxtralBackboneNetwork: drops the cross-model import; calls decoderTransformerNetwork directly.

Intentionally NOT touched

• Gemma 4 — legitimate divergence (geGluFFN, sandwich norms, layer output scale, KV sharing, PLE hooks, custom GemmaModel wrapper).
• Apertus — legitimate divergence (xIELU activation, ungated FFN with separate dense up/down).
• Voxtral acoustic — 3-layer flow-matching transformer with no embedding/output projection, different beast.
• BERT — encoder-only, bidirectional, LayerNorm not RMSNorm, GELU FFN.

Test plan

• ./gradlew :llm-core:jvmTest :llm-inference:llama:jvmTest :llm-inference:qwen:jvmTest :llm-inference:voxtral:jvmTest
• ./gradlew :llm-inference:apertus:jvmTest :llm-inference:gemma:jvmTest :llm-inference:bert:jvmTest
• ./gradlew :llm-runtime:kllama:jvmTest :llm-runtime:kqwen:jvmTest :llm-api:jvmTest :llm-agent:jvmTest
• All pass.
• CI green on PR.

DSL-vs-LlamaRuntime numerical parity test deferred to a focused follow-up PR with careful tolerance choice. Refs the closed #46.

🤖 Generated with Claude Code