WASTE Files

0.7.0 did not build on x86_64 Linux, on Windows, or under ASan. That is the reason this release exists; if you are on any of those, 0.7.0 is not installable and this is the fix. Apple Silicon was unaffected, which is why it shipped.

The rest is the test suite learning to tell a tie from a defect. Four checks were red or blind across the three models, and every one of them turned out to be measuring a discrete choice by the distance between logits — a question that measurement cannot answer. None was an engine bug. The engine changes here are two: the i8mm guard above, and one printf format.

Fixed

• The i8mm call site is guarded on the same predicate as its source file (docs/LEARNED.md §72). model.c declared and called waste_mvq4_rows_i8mm unguarded, but the Makefile adds src/simd_i8mm.c to SRC only for arm%|aarch64%, so on every other platform the symbol does not exist:

model.c:805: undefined reference to `waste_mvq4_rows_i8mm'

Three of five CI jobs failed at the link, and one of them is asan + fuzz — so the sanitizers had never run against anything 0.7.0 added. simd_i8mm.c was already careful, defining the symbol twice so it exists in every ARM build even where -march did not take; what was missing is that a dispatcher and the list of files that satisfies it have to be guarded on the same predicate. TK_I8MM is already unreachable off ARM, so this is dead code being compiled, not behaviour. Verified rather than assumed: compiling model.c for x86_64 leaves the symbol undefined at 0.7.0 and unreferenced now, a full x86_64 build links, and the arm64 object still calls it.

• The two K3 checks were measuring the router, not the arithmetic (docs/LEARNED.md §71). Chunked prefill and the CPU backend each differed from the default path by max-abs 0.2858 against a 1e-3 threshold, and 0.7.0 shipped them red on the explanation that this was the i8mm/SMLAL trunk kernels. It was not — trunk_kern defaults to TK_F32, so neither kernel was in the run.

It is one routing tie. Of 1472 routing decisions in a 16-token K3 prefill, the earliest the paths disagree on is token 12, layer 56, experts 889 and 712, at a relative margin of 7.311e-07 — the minimum over all 1472. The 1st percentile is 4.4e-05 and the median 7.4e-03; the 47 differences that follow average 5e-03, because they are computed on a hidden state that has already moved. Two independent paths produce byte-identical route traces and differ from the default in the same 48 places: NEON summation order against scalar, disagreeing by 1e-08 on a decision that needed 1e-07.

A top-K router makes an arbitrarily small arithmetic difference discrete, so past the first flipped expert the logit distance measures how much the model cares which of two indistinguishable experts it ran. No threshold on it works: 1e-3 fails on a tie forever, and the 0.3 that would pass could not catch a broken kernel.

tests/route_diff.py asks the question the threshold stood in for, over the route and score traces WASTE_DUMP_ROUTE / WASTE_DUMP_SCORES already wrote: identical, tie (the first disagreement is under a relative margin of --eps, default 1e-5) or diverged. The default sits in the empty decade between the tie that flipped and the tightest call that held. The three chunked/backend checks now use it, and the result is a stricter suite, not a looser one — the argmax became a hard failure on every path rather than a clause on one, a route that flips on a resolvable margin fails while naming the token and layer, and a threshold miss with the routing unchanged — the case that is a real arithmetic defect — is its own verdict instead of being pooled with the tie.

• WASTE_DUMP_SCORES prints %.9g. The dump exists to say how close a ranking decision was, and six digits cannot resolve one: the two scores above both printed as 0.112161 while selecting differently, which reads as a selection bug. Nine digits round-trip a float.

• The GLM oracle check was failing on both Linux jobs, on an exact tie (docs/LEARNED.md §72). It had been red since 0.7.0, saying only "the GLM path diverges from the oracle". linux-x86_64 and linux-arm64 report rel L2 0.00783997 and 0.00783999 — two different ISAs agreeing to five significant figures is not a platform difference — and at the worst logit macOS, both Linux runs and the shipped fixture all print -7.3257, while only the freshly generated Linux oracle prints -7.43141.

At layer 2, token 15 the four visible pools score 0, 0, 0 and 0.00164 with keep=2: pool 3 wins outright and the second slot is an exact three-way tie, margin 0.000e+00. The engine takes pool 1; torch.topk, called in kimi_ref.py with sorted=False and so free to answer in any order, takes pool 2 on Linux and pool 1 on macOS. One pool of four tokens attended differently is worth rel L2 0.0078 in the logits, and it is a defect in neither.

kimi_ref.py already wrote the trace that says this, under the same WASTE_DUMP_DSA the engine uses, and its own comment says why: so the two selections "can be diffed directly rather than inferred from a logit difference". Nothing was diffing them. tests/dsa_diff.py now does, reporting the same three answers route_diff.py does over the pool ranking instead of the expert ranking.

Recorded because the measurement is the useful part, this is what it was not: not a stale fixture (the engine matches it to 5.72e-06), not macOS skipping the real comparison, not torch's version (uv resolves the same 2.13.0, fla-core 0.5.2 and einops 0.8.2 on both), not -ffp-contract, not thread count, not UB under ASan/UBSan, and not a router tie — the tightest top-2-of-8 boundary gap in that prefill is 9.4e-03.

• A refusal check has three outcomes, not two. rope_refused grepped test_forward's output for the expected refusal and called anything else "loaded instead of being refused". A test_forward that is killed says nothing, and saying nothing is not the same as loading a container it should have rejected; the check could not tell them apart and reported the more alarming one. It now separates them and prints what came back either way.

• Checks say what they saw. Every comparison above now reports magnitude and location on failure rather than a bare verdict. Two of these were investigated for hours against a bare FAIL, and the answer each time was four numbers the check already had in hand.