[michalharakal]

Summary

First step of bringing the wasmJs tier to feature parity with the JVM tier (compute layer). No C changes; everything here is Kotlin, the support matrix and the bench.

• BF16 / FP16 matmul on wasm. The sidecar has exported skainet_bf16_matmul and skainet_fp16_matmul since 0.61.0, but WasmKernelProvider never bound them, so 16-bit dense weights fell to the scalar Kotlin kernels in a browser. WasmBf16MatmulKernel / WasmFp16MatmulKernel bind them; a contiguous weight is uploaded once and stays resident, a strided window is staged per call like the activation.
• Sidecar diagnostics. WasmSimd.features() (mode, SIMD128, threads, cross-origin isolation, logical cores, resident weight bytes), and WasmSimd.install() prints the same one-line [SKaiNET] Using … banner the JVM factory prints.
• KernelStrictness is common code. It was JVM-only (System.getProperty). Now KernelStrictness.overrideEnabled is the programmatic switch on every target (the only one on JS, Wasm and Kotlin/Native); JVM and Android still honour -Dskainet.strict.kernels. Two new fail-fast sites make strict mode mean the same thing on every tier: KernelDispatch before the decoding reference fallback, DefaultCpuOps.matmul before the generic per-element loop. The JVM subclass keeps its earlier check.
• Kernel support matrix. Float16 row (every tier had the kernel, the matrix had no row); BF16/FP16 in the wasm column; Q6_K on the JVM corrected from panama-vector to native-ffm (NativeKernelProvider.matmulQ6K() has existed, the declared tier had not been updated — DefaultCpuOps resolves the kernel from the best available provider, so FFM was already serving it); a third table, Resident serving, for the sidecar-resident weights that stand in for mapped serving on JS/WASM. The test now also gates the declared Panama and FFM tiers against what the providers expose when they are available on the runner.
• wasm-bench runs the engine-* scenarios of jvm-cpu-publish (same ids, shapes, generators and metrics) on each tier it measures, so JVM and wasm numbers sit in one table. engine-ternary-f32-gemv is reported as skipped on wasm until the ternary kernels join the sidecar (the decoding reference path takes hours at that shape).
• Docs: eager-execution tables gain the wasm column and FP16 row, the stale "Native FFM Q6_K" gap is replaced by the real wasm gaps (ternary, non-matmul ops); dtype-model documents the common strict switch.

Verification

• :skainet-backends:skainet-backend-wasm-simd:wasmJsNodeTest — parity for BF16/FP16 (contiguous and windowed) against the scalar kernels, features, strict mode on wasm, plus the existing Q8_0/Q4_0/dense tests.
• :skainet-backends:skainet-backend-native-cpu:jvmTest --tests '*KernelSupportMatrixTest*', generateKernelMatrix (page regenerated).
• :skainet-backends:skainet-backend-cpu:jvmTest --tests '*KernelStrictnessTest*'.
• scripts/pr-gate.sh --quick.

Follow-ups in the same series: explicit SIMD128 bodies for the K-quants, Q5_0/Q5_1 and dense FP32/BF16/FP16 plus row threading for the dense kernels; ternary kernels in the sidecar and the direct x · Wᵀ kernel on the JVM; resident activations and fused elementwise/attention kernels; the browser IO gaps.