v0.6.9

Almost none of this is the engine. It is the space no CI job could reach —
Windows, real containers, and the four feasibility gates that were still open —
and most of it was found by people running hardware this project does not own.
GATES.md now has no open gate for the first time, and two of the three that
closed this cycle closed against the change they were proposing.

No recompile needed. src/waste.h changed only in a comment this cycle, so
unlike 0.6.8 there is no ABI move. waste_kernels gained a slot, but that lives
in src/waste_backend.h and is not public.

Added

• VQ4P's apply is a dispatch slot, and x86 has a kernel
  (#38,
  #41). vq_rows_p6 was an
  #ifdef in model.c with an ARM-only body, so §47's tuning did not travel.
  It is now waste_k.vq_rows_p6, with the NEON body moved behind it unchanged
  and an AVX-512 VBMI implementation registered beside it. The refactor is
  bit-identical on ARM against a real index_bits: 6 container, checked
  here on both a VQ4P and a same-geometry 8-bit control. The VBMI kernel itself
  is compiled and dispatched, never executed — no machine here has AVX-512,
  the same standing caveat docs/BACKENDS.md records. It carries
  __attribute__((target("avx512vbmi"))) so a VBMI instruction cannot be
  placed in the F+BW kernels that share its translation unit and SIGILL on
  Skylake-SP.

• WASTE_DUMP_SCORES=path (#45),
  the whole router vector per (token, layer), not only the winners. It records
  score[e] + bias[e] — what the selection loop ranks on — rather than w[j],
  which takes score[best] without the bias and is applied after the choice is
  made. A residency prior, a dynamic-k rule and a pruning criterion are all
  questions about the experts that lost, and a trace of the chosen cannot
  answer them.

• WASTE_CCR_LAMBDA (#46), a
  cache-residency bonus applied to ranking only, off unless set
  (arXiv:2412.00099). w[j] still takes the
  untouched score[best]; only which K run moves. Reproduced here on
  Kimi-Linear at --budget 4G, -n 120: hit 76.1% → 89.5% at lambda 0.10
  and bytes per token 0.231 → 0.167, a 27.7% saving. See not adopted below
  for why the default did not move.

• waste plan prints one token's working set. It quoted the quantity in
  the line below it and never showed it, so the only way to see the number was
  to compute it — and computing it from FORMAT.md's record size gives a
  different answer, because the engine counts every layer including the dense
  ones. 17.19 against 17.01 GiB on K3, deliberately, in the direction that
  recommends slightly more. src/waste.h had described the field as top_k
  records per MoE layer, which is the sentence that produces the mismatch for
  anyone who believes it.

• The fp8 block-scale mapping has a test
  (#39,
  #40). The only stub naming it
  replaced it with the identity, so nothing exercised the tile arithmetic
  0.6.8 shipped.

Fixed

The five gaps of #36 — a real
Kimi-Linear container on Windows, whole-model oracle PASS on AVX2 — and the two
that survived the first attempt:

• WASTE_MLOCK wired nothing on Windows. VirtualLock is bounded by the
  process minimum working set; the first fix raised the maximum and preserved
  the minimum, so it wired exactly as much as the bug did: 5928 of 5928 slots
  refused, twice. Measured on the reporter's host rather than reasoned about —
  the maximum went 1.3 MB → 2049 MB and the lock still failed. Raising both
  bounds wires the cache. SE_INC_WORKING_SET_NAME turns out not to be
  required.

• diskbench truncated its offset through a 32-bit off_t
  (#43), at its own default 8 GB
  file size, under LLP64. The loud half wrapped negative and was refused; the
  quiet half wrapped positive and read the wrong place successfully, keeping
  the working set inside the first 2 GiB — inside an SSD's SLC cache, which is
  the flattery the tool exists to avoid. Random rows went 0.36 → 1.48 GB/s on
  the reporter's box and stopped inverting with thread count. docs/GATES.md
  Gate H and the README's 12.78 GB/s are unaffected: macOS off_t is 64-bit.

• A container's JSON went through Python text mode, so the same conversion
  produced byte-different containers on Windows and the provenance gate added
  below reported a good fixture as stale. convert.py and requant_vision.py
  had it too, which means a K3 container converted on Windows was not
  byte-comparable with one converted anywhere else.

• A missing tool said the engine was broken
  (#42). Without diffutils the
  suite reported expert cache changes results — a bit-identity violation in
  the expert cache — on a clean checkout, because cmp was absent. Eight
  failures became eight skips naming the tool, verified here with cmp hidden;
  python3 joins curl in guarding the download checks. UCRT64 is the
  documented Windows path and ships neither.

• check_budget.sh said FAIL where it meant "could not measure", so the
  budget ceiling was untested on Windows rather than merely misreported. It now
  exits 77 and reads the real number through PeakWorkingSetSize.

• The oracle fixture gated on trunk and not on provenance. A container
  converted with a different --device trains different codebooks; the sidecar
  now carries codebooks_sha256, which covers every conversion knob at once.
  This is #7 recurring on a second
  axis and it is not Windows-specific.

• CRLF in .shards made every promise that file's header makes about
  surviving a long haul inert on Windows, first run included; the API listing
  had it too.

Also: get_small() dropped API-listed files silently and called the run a
success (#35); the cpu-list bind
check raced on a wall-clock assumption
(#30); diskbench is now built by
CI natively and cross, under -Werror, which caught a mingw driver flavour on
its first real run; and the CLI help said --threads 0 is one per core when
it is one per logical CPU (#44).

Documentation

docs/GATES.md has no open gate. Gate 7 closed refuted — see below —
and docs/LEARNED.md gains §61 through §64.

docs/BACKENDS.md carries a correction it owed: two sentences took one
measured mechanism — synchronous launch and a round trip per call over several
hundred small dependent matvecs — and restated it as a claim about shape.
Coherent memory removes the round trip without removing the dependency, so
neither claim follows. The superseded paragraphs are marked and kept.
docs/ENGINE.md gains thread placement on a single-CCD part, written as a
bound with the noise floor stated rather than as a result.

Measured, and not adopted

• The budget resolver's quantum stands. Gate 7, run on a 128 GB Strix Halo
  (#37, §63): at 200 tokens
  rather than four, a 3–4 GB cache measures 64.6%–77% of floor + 1x,
  where the kill criterion asked for 90%. §39's window is a four-token window,
  and cross-token reuse is what the extra size buys.

• Cache-conditional routing stays a knob. README prices top-8 against two
  bars — KL 0.037 and it reproduces top-16's greedy continuation. Lambda 0.10
  clears the first at 0.0232 and not the second; confirmed here on a second
  machine, where contiguous became consecutive.

• The thread default did not move, though 16 threads on an 8-core SMT part
  measure ~1.6x below the 6–8 plateau. §47 already has the setting inverting
  between models and #37 added
  that it inverts between expert formats on one machine. Three axes is a
  decision, not a table.

• No budget policy fixes two K3 opens on a 64 GB machine
  (#31, §64). Gate 1 ran: neither
  process reached decode, and the machine needed a power cycle. Two floors are
  58.38 GB against a 51.54 GB ceiling, so a reservation ledger and a trimmed
  multiplier meet the same wall — they divide the cache above the floor, and
  the floor is what busts the budget. The narrower case survives in
  #49.

• VQ4P is not a throughput upgrade over VQ3R on a GB10, despite complete
  and exact CUDA coverage, because it read 8.11% more expert bytes (§62). And
  src/cuda.cu still does not exist: the build guard stays, now for a
  maintenance reason rather than a performance one.