[michalharakal]

SKEEP-007 "The browser lane" is the design record for making inference in a browser usable and quick. This PR adds the proposal page and lands its Tier 0: the common-code fixes that keep the scalar wasm lane off the decoding reference kernel and out of memory trouble. The SIMD128 sidecar kernel pack (Tier 1) and quantize-at-load for SafeTensors (Tier 2) follow as separate pull requests, as the page lays out.

What was wrong

• matmulWeightTransposed on a canonical (row-major) packed weight — every weight loaded from a GGUF — reached KernelDispatch on a target without a row-major kernel pack and ran the per-element decoding reference kernel on every call (the dispatcher's own "~1000x slower" warning). The scalar packed kernels every target carries are keyed on kernel feed order and were never reached from this path; the prepackedWeights cache declared for exactly that relayout had no common-code user.
• A dense FP32 [out, in] weight was transposed into a cached heap copy before first use on every tier, doubling resident weight memory for a dense encoder (about 0.5 GB for a 22-layer, 768-wide model).
• A 16-bit token table was widened to FP32 at load although ops.gather and the transformers Embedding already read rows from a RowDequantSource.
• The web samples each carried their own in-memory RandomAccessSource; JsBlobRandomAccessSource converted buffers one interop call per byte and logged every read.

What changes

• Dispatch order in DefaultCpuOpsBase.matmulWeightTransposed for packed weights: an exact kernel for the canonical view (JVM FFM, Android JNI, the coming browser pack) → the one-time cached relayout into feed order served by the scalar packed kernels → views and the reference kernel last. The JVM overrides this method and is unaffected; ternary formats have row-major kernels and take the first branch as before.
• Direct dense path: x · Wᵀ for heap FP32 [out, in] weights is computed from the weight's own rows with a 4 × 4 tile and one sequential accumulator per output, bit-identical to transpose-then-matmul. preferDirectWeightTransposed() is the hook; DefaultCpuOpsJvm and AccelerateCpuOps return false and keep their cached transpose because their [k, n] kernels are vectorised.
• NarrowFloatDenseTensorData : RowDequantSource (FP16 and BF16 tables serve rows on demand).
• sk.ainet.io.ByteArrayRandomAccessSource in skainet-io-core, single array or equal-sized chunks; JsBlobRandomAccessSource drops its per-read console logging (the per-byte conversion stays: Kotlin/Wasm 2.4 has no typed-array adapter; Tier 1's linear-memory bridge replaces it).
• skainet-backends/benchmarks/wasm-bench: a Kotlin/Wasm executable (Node and browser) timing the matmul shapes of Laya, LEAF, SmolLM2 and Qwen3-0.6B as dense FP32 and packed Q8_0, printing [SKAINET_PERF_WASM] lines. Baseline numbers are recorded in the SKEEP page.
• Docs: SKEEP-007 page, index and nav; eager-execution.adoc and kernel-selection.adoc describe the new wasm behaviour; CHANGELOG Unreleased.

Tests

• DenseWeightTransposedDirectTest: tiled row-dot bit-identical to the sequential definition for every tile remainder and with array offsets; the primitive agrees with transpose-then-matmul for rank-1, 2 and 3 activations.
• PackedWeightKernelOrderTest: a canonical Q8_0 weight matches the relayout path, the per-weight cache is hit on the second call, and a RecordingTraceSink on KernelDispatch.defaultSink sees no reference-kernel run.
• NarrowFloatRowDequantTest, ByteArrayRandomAccessSourceTest.

Verification

• scripts/pr-gate.sh --quick (every module's jvmTest + apiCheck): passed.
• :skainet-backends:skainet-backend-cpu:linuxX64Test (full, includes the golden packed-encoding parity): passed.
• wasmJsTest for skainet-lang-core, skainet-io-core, skainet-backend-cpu (ChromeHeadless): passed.
• wasm-bench on Node: baseline recorded in the SKEEP page (dense 0.84–1.96 GFLOPS, scalar Q8_0 0.35–0.63 GFLOPS; one Laya layer ≈ 1.34 s dense).
• API dumps regenerated: additive only (ByteArrayRandomAccessSource, RowDequantSource on NarrowFloatDenseTensorData, preferDirectWeightTransposed).
• The remaining legs (JS tests, wasmWasi, assemble, Java consumer tests) run in CI.