WASTE Files

Six pull requests and two issues, and one theme running through all of them: a test that compares a thing to itself is not a weak oracle, it is not an oracle. Three of the defects this release fixes were sitting under green boards, and each one was found by diffing against something outside the code rather than by a check inside it.

No ABI move — src/waste.h is untouched — and no measured number in the docs changes. The engine gains one optional capability, bank striping, and two refusals.

Added

• Bank striping across N devices (#53). WASTE_BANK_SHARDS=/mnt/a,/mnt/b reopens each layer's bank as N shard files; expert e lives on shard e % N at offset (e / N) * rec_bytes. Round-robin rather than per-layer because the per-token demand is k experts of one layer, so layer-granularity placement would leave every read of a token on one drive. tools/split_banks.py produces and byte-verifies the shard sets. Unset is the N=1 case of the same code path, and no container format changes. Striped and unstriped logits are byte-identical, under ASan/UBSan too.

No bandwidth number, deliberately: that needs a multi-drive rig with a real container, and this is the mechanism. The 1.8-2x it targets is a prediction, not a measurement, and is recorded as one.

• Kimi K2 native tool calling for containers served from chat.json (#50), in serve/kimitools.py — its own module beside xtml.py, because it is neither of the two formats serve/ renders: it is carried in a container's tokenizer while that container's chat.json says nothing about it. Kimi-Linear carries the five markers; GLM does not and is still refused by name. The split is by subject: whether a container can do tools is a fact about its chat.json, so chatfmt.py decides it; how a call is spelled is a fact about the protocol, so kimitools.py owns it and raises its own KimiToolError, the same shape xtml.py has.

• tests/serve/test_chatfmt_upstream.py, the oracle that made the above trustworthy (docs/LEARNED.md §73). The 438 lines of tests that came with [#50] all passed and the rendering was still wrong: they fed the renderer's output to the parser, which agrees with itself whatever the format is, and one of them pinned the defect. Diffed against K2's published chat_template.jinja, everything matched byte for byte except the turn a tool result opens. K2_DIR names the release; the check needs no weights and no container, which makes it the cheapest oracle here.

• A container the suite can reach at index_bits 6 (#48). vq_rows_p6 had none — the synthetic container was always index_bits 8 and a real conversion costs hours — so every green run covered VQ3R and nothing covered VQ4P, on any platform, which CLAUDE.md warns about by name. Four checks, and the arm is not decorative: swapping T1/T2 in vq_rows_p6_neon fails it, naming the specific verdict.

• tests/test_vq4p_packing.py (#56), which runs make_test_container.py's VQ4P packing against convert.py's. The arm above cannot: it compares the engine against itself, so a generator that packed differently would produce a container both backends decode the same wrong way and all four checks would pass. Verified not to be decoration by breaking the generator two ways.

Fixed

• A python3 that exists and is not an interpreter said the engine was broken (#51, [#52]). On Windows the name on PATH is usually the Microsoft Store App Execution Alias: a zero-byte reparse point that command -v finds, that exits 49, and that prints an advert instead of running anything. Every guard in the suite was written for an absent interpreter. Measured on a stock Windows box: 7 passed / 7 failed / 16 skipped before, 44 / 0 / 13 after, with the Store's advertisement text appearing verbatim inside failure messages. The suite shims PATH; the Makefile and the two shell tools resolve a name.

• The other half of that (#54): seven checks still reported the engine, the budget, the converter and the release's own preprocessing config broken when no interpreter answered. 14 failures before [#51], 7 after, 0 now, with every skip naming the interpreter. 77 rather than a new number — tests/check_budget.sh already meant "could not measure, skip loudly" by it.

• The turn a tool result opens is named for the tool, name or "tool", never system. K2's own template renders <|im_system|>{{ message.get('name') or message['role'] }}<|im_middle|>.

• Two claims about shards that were not true. [#53]'s comment promised that a short or long shard fails the load; bank_open's rec_bytes parameter is (void)-ed and nothing measured anything. A shard short by one record loaded, ran, and produced correct logits — correct because the missing expert simply was not routed to in an 8-token prompt, which is the hazard rather than the reprieve. [#55] then added the size check at open, which is what should have been written in the first place: waste_file_size has been in platform.h all along.

Its own justification is corrected too. A misplaced record cannot be served as another expert's — every record carries the expert it belongs to and record_check reads it, and two shards padded to the right sizes but carrying another split's records refuse on the first read. What the check buys is when: a long shard loaded happily and generated correct output, and a short one waited for the router to reach the missing expert.

• WASTE_BANK_SHARDS truncated at 1024 bytes was ignored, silently dropping shards. A shard count that disagrees with the split that produced them is a different layout, not a smaller one. Refused.

• The O_DIRECT probe used a stack buffer, which broke every container load on linux-arm64 — 25 passed, 29 failed. O_DIRECT rejects a transfer whose buffer is unaligned, not only its offset and length. It passed on macOS, whose F_NOCACHE has no such requirement, and on the x86_64 runner, whose filesystem declines O_DIRECT and falls back to buffered: 56 passed / 0 failed on the same code and the same OS as the job that failed. Two platforms agreeing is not coverage when only the third exercises the path. main was red for one commit.

• A refusal check had two outcomes where it needed three. rope_refused called anything that was not the expected message "loaded instead of being refused", so a test_forward that was killed — which says nothing — was reported as a container loading that should have been rejected.

Also

CI's SPDX glob gained tests/*.py, tests/*.sh, tests/serve/*.py and serve/*.py. Every file in those directories already carried the header; the glob had simply never looked, three separate times this release.

route_verdict takes its reference as a parameter, so the three arms that ask it — VQ3R, rotated, VQ4P — share one copy instead of open-coding it.