test(m1): synthetic decode harness asserting M1's memory acceptance criteria (SKEEP-003 M1, S1.11a)

[michalharakal]

Summary

The last M1 slice (#1032), built to the decision recorded on that issue — option (c), both homes. This is half (1): a synthetic decode loop over the real memory machinery, in the repository where that machinery lives, so the criteria that are about memory behaviour are asserted on every commit. Half (2) — the real GGUF model with tok/s, TTFT and the release-notes table — belongs to skainet-decode in SKaiNET-transformers, which owns the model.

• DecodeHarness (backend-cpu commonTest): a Llama-shaped stack of packed Q8_0 matmuls whose weights live in a ModelScope, activations come from a recycled ForwardScope, the KV ring is preallocated in the model scope (feat(memory): KV cache preallocates its ring in Scope.Model (SKEEP-003 P2, S1.10) #1076), dispatch goes through KernelDispatch (feat(backend-api): KernelKey dispatch — declared formats/layouts, rank normalised once, visible adapters, reference matmul (SKEEP-003 P3, S1.7a) #1070/feat(backend-cpu): route the generic matmul path through the kernel registry (SKEEP-003 P3, S1.7b) #1071) and every event lands in a RecordingTraceSink. No tokenizer, no sampling, no checkpoint — it is a memory-behaviour fixture, not a model. Deliberately tiny (2 layers × 32 hidden, 12 steps) because the reference kernel decodes every element and these tests also run in a browser under Karma's 2 s per-test budget.
• DecodeAcceptanceTest asserts:
  • M1-A1 memory is flat — the forward scope is empty between steps and every post-warm-up reset reports the same live-bytes-before;
  • M1-A3 zero forward-scope allocations per step after warm-up (and none in the model scope during decoding);
  • M1-A8 the plan matches the run within 10 %, with no adapters inserted;
  • M1-A7 the Perfetto trace has one track per scope, kernel spans labelled by TensorId, and a live-bytes counter that returns to zero.

Two real gaps the assertions surfaced (both fixed here)

1. Adopted weights were invisible to plan-vs-actual. Borrowing is not an allocation, so mapped/borrowed weights emitted no event — yet decision BPETokenizer #11 counts weights as resident, because decode touches every weight every token. ModelScope.adopt now emits the allocation (site "adopted") and close() balances it with a Free, so a plan can be checked against a run holding mapped weights.
2. The planner assumed bf16 KV while DefaultKvCacheStore stores FP32, understating a dense ring by 2×. KvCacheMode.FP32 added and documented as what the dense store actually does. This is exactly the drift feat(memory): plan-vs-actual — reconstruct a run's memory from the event stream and fail on drift (SKEEP-003 P2, S1.9) #1074 exists to catch — and it motivated KvCacheStore should declare its Format (dtype + encoding), not just an encoding #1077, which replaces the guess with the store declaring its own Format.

Test plan

Full local gate (scripts/pr-gate.sh, JDK 25) — all legs passed (JVM, apiCheck, JS/Wasm incl. the browser targets, linuxX64, assemble, Java consumer API tests). The first run failed only on Karma's per-test timeout, which is why the harness shapes are small; the assertions are unchanged in substance.

Closes #1032

🤖 Generated with Claude Code

[michalharakal]

Local gate scripts/pr-gate.sh (JDK 25) on 60ed0c1: all legs passed — jvmTest · apiCheck · JS/Wasm (browser targets included) · linuxX64Test · assemble · Java consumer API tests.

Targeted: sk.ainet.exec.harness.* 5/5 on JVM, JS, wasmJs and linuxX64.

[github-actions]

📖 Documentation Preview

The documentation has been built successfully for this PR.

Generated Files:

• Operator documentation: docs/modules/operators/_generated_/
• JSON schema output: operators.json

Artifacts:

• Download the documentation-preview-1078 artifact to view the complete documentation locally.

This comment will be updated automatically when the PR is updated.