feat(backend): bridge the packed SPI kernels through the ordered key (#973.2)

[michalharakal]

Closes #1095 · #973.2 · unblocks what #1029 deferred

What was in the way

The packed matmul kernels read their weight input-block-major; a packed TensorView loaded from a file is canonical. Until #1094 nothing in the key said which was which, so #1029 deliberately left them unbridged rather than risk the silent-wrong-numbers class of #968/#971. Now the order is part of the key, and a kernel can declare what it reads.

• PackedViewMatmulKernel — a ViewKernel over any of the packed SPI kernels, keyed on BLOCKED_INPUT_MAJOR and checking it at run time rather than assuming. A strided activation or non-heap storage falls back to the reference kernel instead of being mis-indexed.
• KernelPacks.installPacked(provider) registers Q4_0, Q5_0, Q5_1, Q8_0, Q4_K, Q5_K and Q6_K when the provider offers them; anything it does not offer keeps being served by the decoding reference.

The default that a failing test chose

KernelDispatch.matmul(..., prepackWeights = false) can relayout a canonical weight to reach those kernels — off by default, on purpose.

I first wired it on by default. That broke M1-A3 (zero forward-scope allocations per decode step): the dispatcher was relayouting the whole weight on every matmul of every token — precisely the per-forward O(bytes) copy #973 objects to in ops.transpose, merely relocated. The regression is the argument: a weight is prepacked once, at load, and a canonical weight handed straight to the dispatcher still gets the correct, slower reference path. prepackWeights = true exists for one-shot calls where the copy beats decoding.

This is also why #1097 (engine-owned prepackForMatmul at the load boundary) is the natural next slice: it is what makes the fast path the default without a per-call copy.

Acceptance

PackedDispatchBridgeGoldenTest, for all seven encodings:

• the registry path is bit-identical to calling the SPI kernel by hand with block-major bytes;
• the recorded digests' leading values are the existing scalar-matmul/* goldens — the bridge changes nothing about the numbers, which is the whole claim;
• the relayout is visible as one prepack-input_block_major adapter;
• a weight already in kernel order costs no adapter at all;
• every weight is three blocks wide — at one block per row the two orders coincide and the test would pass vacuously.

Gate

scripts/pr-gate.sh — all legs passed. --golden — passed: 46 golden tests, the 39 pre-existing digests unchanged.

Keeps develop green by

Additive registration plus one defaulted parameter whose default is today's behaviour. The is-ladder in DefaultCpuOps/DefaultCpuOpsJvm is untouched — it can start shrinking once #1097 makes loaded weights arrive prepacked, and that is a separate PR with its own benchmark evidence.

🤖 Generated with Claude Code