v0.6.2

Fixed

• waste info and waste run crashed on K3 on every x86 build
  (#10). The tensors the
  loader skips — the vision tower, and anything outside tensor_prefix —
  kept group at 0, and the row-scratch sizing divided by it. The
  architecture decided what that meant: arm64's sdiv answers 0 and the
  run continues, x86's idiv raises #DE. waste plan was unaffected
  because it does not load. §37.
• WASTE_Q8=0 could not load a 4-bit trunk
  (#6) — that is, any
  container a default tools/convert.py run produces. The dequantizer
  read one byte per weight, true of Q8G alone, while catching every
  quantized format. It now decodes through waste_deq_row, the one place
  that knows all three widths. The same lines also predated waste_f16's
  subnormal fix and flushed group scales below 6.1e-05 to zero.
• embed_tokens stays on disk under WASTE_Q8=0, as it does
  otherwise: 7.93 → 6.52 GiB of peak RSS on Kimi-Linear, identical logits.
  The f32-equivalence check now differs from the default path in the
  storage width alone, which is what it claims to compare.

Added

• Router lookahead in the decode path. At the end of a MoE layer, once
  its reads are consumed and the disk is about to idle through the next
  layer's attention, layer L+1's router runs on layer L's hidden state and
  issues speculative reads for its top 6. Demand hit rate 14–19% → 38–40%
  with total bytes read unchanged (254.2 → 254.5 GB): the records were
  going to be read anyway, and only when changes. Nine paired runs,
  median 1.17x. WASTE_LOOKAHEAD=0 disables. §34, §35.
• WASTE_MLOCK wires the trunk and the expert cache; WASTE_MLOCK=cache
  wires the cache alone. Off by default — Linux's RLIMIT_MEMLOCK is
  commonly 8 MB. Wiring the trunk is worth 3x in the transition zone around
  52 GiB and nothing below it; it does not move the knee. §30, §31, §32.

Changed

• tests/run.sh generates its own PyTorch oracle from the container
  under test (16.9 s) instead of diffing against a shipped fixture, which
  can only ever be valid for the container that produced it: expert
  codebooks are k-means, and the same seed on a different --device
  trains different books — one layer of 26 moves the logits by 1.24
  against a 1e-3 threshold. The fixture remains as the fallback where uv
  is absent, with its provenance recorded beside it.
  (#7), §33.
• tools/make_test_container.py emits what a real conversion does: a
  Q4G/Q8G/F32 trunk rather than Q8G throughout, and --prefix for a
  container whose tensors are not all under its tensor_prefix. Both are
  shapes the suite could not previously reach, and both had a live bug
  behind them.
• Checks that cannot run now say why instead of reporting a refusal as a
  divergence: WASTE_Q8=0 on K3 wants 211 GB of f32 trunk on a 64 GB
  machine, the oracle prompt is Kimi-Linear's, and a cold hotlist run that
  already missed nothing demonstrates neither outcome
  (#5).

Measured and not adopted

• Cross-layer prefetch from next_layer_top — gated before building
  and refused: 29.0% recall against a 60% break-even. §29, revisited and
  superseded by the lookahead above in §34.
• The lookahead in the prefill path. Built, measured, removed: a chunk
  layer's disk is busy continuously, so a prefetch there does not move a
  read into idle time, it moves it in front of another read and pays an
  eviction for it — 7% more bytes. Decode keeps it. §36.

Section numbers refer to docs/LEARNED.md, which carries the reasoning and the full measurements behind each entry.

Full Changelog: v0.6.1...v0.6.2