Kernel SPI: make weight addressing order and byte source declared contract dimensions #1193

Follow-up to #1189/#1190 (siblings: #1191 FFM tier, #1192 remaining formats). This one is about the kernel contract, not any single kernel.

The two hidden dimensions

#1190 exposed that every packed matmul kernel actually lives in a matrix that the SPI never named:

1. Addressing order — where block (o, b) sits in the weight bytes:
   • feed order (b · outputDim + o) · bpb — the prepacked fast path (skainet_q4k_matmul)
   • row-major file order (o · blocksPerRow + b) · bpb — what an mmap'd GGUF is (skainet_q4k_matmul_rm)
2. Byte source — where the weight bytes live when the bridge hands them to C:
   • heap ByteArray (JNI GetPrimitiveArrayCritical, FFM staging copy)
   • off-heap direct address (JNI GetDirectBufferAddress, FFM MemorySegment zero-copy)

The C functions are pointer-based and never cared — the cost is in the bridges: today each occupied cell of formats × order × source × (JNI|FFM) is its own interface/class/shim (Q4KMatmulKernel, Q4KMemSegMatmulKernel, JniBufferPackedMatmulKernel, …), most cells are simply empty, and an empty cell fails soft: PackedViewMatmulKernel/JniBufferPackedMatmulKernel silently fall back to the decoding reference — correct, hours-scale, invisible.

Proposed

1. Order stays two symbols in C (the loop structures genuinely differ; keep hot loops branch-free) — but a pair becomes the definition of done for a new packed format: feed-order + _rm, parity-tested against each other on permuted bytes (RowMajorMatmulParityTest is the template). The inner block-dot helpers are shared, so the second variant is ~60 lines.
2. Byte source collapses in the bridge: one view kernel per (encoding, order) key that unwraps whatever Storage arrives — Heap → array shim, OffHeap/Mapped → direct shim — instead of parallel kernel classes registered under the same key. Order is already in the dispatch key (BLOCKED_ROW_MAJOR/BLOCKED_INPUT_MAJOR, Packed storage that knows its own block order, so KERNEL_FEED can reach a kernel #1120); source is a runtime property of the operand and belongs inside run(), not in the key.
3. Fallback becomes visible: when a packed view kernel punts to the decoding reference (unsupported storage, strided activation), emit a trace event — SKEEP-003 §5.1 already demands adapters be visible in the trace; a 1000× silent slowdown deserves the same.
4. The kernel support matrix (generateKernelMatrix) grows the two columns, so coverage gaps are documentation, not archaeology — and StorageCapabilities.mappedServableEncodings can eventually be derived from registrations instead of hand-maintained (the plan/load/dispatch truth-coupling from feat(#1189): packed-tensor mapped staging — Q4_K/Q6_K served straight from the mmap on Android #1190's planner fix).

Why now

Android is the next primary target, and the mapped tier is its load-bearing feature: every future format pays these two dimensions. Naming them in the contract makes each new kernel a checklist instead of a rediscovery.