aprender 0.70.2

aprender 0.70.2 — release notes

How this release was gated

0.70.2 ships under a recorded operator emergency scope, crux-smoke, as 0.70.1 and 0.69.1 did.
The operator's words: "0.70.2 ships on CRUX smoke on lambda and gx10 GPU. Everything bigger is nightly."
The record is the crux-smoke entry for release "0.70.2" under emergency_scopes in
contracts/model-capability-ladder-v1.yaml. It is data only: the judge
(scripts/lib/crux_smoke_scope.py) matches a recorded scope by its whole release string, so no code changed
to admit it.

The gate. The CRUX smoke runs on the two GPUs the quote names (an RTX 4090 and a GB10), with nothing else on
either GPU. It covers every certified model (Qwen3.5-2B Q4_K_M, Qwen3.5-4B Q4_K_M and Qwen3.5-4B UD-Q4_K_XL),
every admitted thinking mode (off, on), on the control prompts, and the verbs run, chat, serve and code. The
smoke is run with the binary built at the release commit. A commit cannot hold results produced from itself,
so those receipts are published with the GitHub release for v0.70.2.

The prompt certification is carried from 0.70.1, not re-run.
evidence/crux/0.70.2/prompt-certification.json and prompt-certification-inventory.json are byte copies of
the evidence/crux/0.70.1/ files. Carrying is allowed because nothing the certification depends on changed:

• the models are the same: the three certified files have the same sha256 as for 0.70.1;
• the prompts are the same: between v0.70.1 and this release's base, evidence/crux/0.70.1, the 0.69.1 prompt
  certification, scripts/crux_inference_dogfood.sh, scripts/lib/crux_smoke_scope.py and
  scripts/crux_sweep_shards.sh have no diff;
• the oracle engines are the same: the oracle record is still the 0.69.1 manifests, pinned by sha256 inside
  the certification, which 0.70.1 carried the same way;
• the smoke does not re-run the oracle engines.

Printed as evidence, not the gate for this release. The full model matrix and the release-readiness grade
still run at publish, and their verdicts are printed. They stay red where they were red; see Known failures.

Unchanged. Clean tree, tag on HEAD, release branch, dogfood GO for this commit, and the clean-room green on
the tagged commit.

Changed since 0.70.1

• apr serve: tools in a chat request now reach the chat template, and Qwen3.5 XML tool calls are parsed
  (#4650).
• apr serve: KV caches no longer commit every byte when they are built (#4769).
• aprender-compute: the 8x6 AVX2 kernel behind gemm_blis_with_prepacked_b and gemm_profiled could read up to
  128 bytes past its input and crash (SIGSEGV); fixed (#4891). Results were never wrong. The bug is also in
  0.70.0 and 0.70.1.

Known failures (shipped, moved to 0.71.0)

The operator refused all six for 0.70.2 under the scope above. None has a per-model contract edit, and none is
marked known red or green. Each stays red in the printed evidence and in its ticket, and each is on the 0.71.0
milestone. Each was re-measured on the RTX 4090 on 2026-10-06, against 0.70.1 built from main, and still fails the same way.

model	GPU	check	failure	ticket
qwen35-0.8b-q4km	RTX 4090, GB10	golden_output thinking_on	think block unclosed within 2048 tokens	#4666
Qwen2.5-0.5B-Instruct-f16.gguf	RTX 4090	cuda serve /api/chat	gibberish output	#4661
Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf	RTX 4090, GB10	cuda serve /api/chat	gibberish output	#4662
Qwen3.5-0.8B-IQ4_XS.gguf	RTX 4090	golden_output thinking_on	no measured thinking budget (named refusal)	#4663
Qwen3.5-0.8B-UD-IQ2_XXS.gguf	RTX 4090, GB10	golden_output	GPU forward fell back to CPU (run and chat rc=14); apr qa also fails tensor_contract	#4664
Qwen3.5-35B-A3B-UD-IQ4_XS.gguf	RTX 4090	capability_match	no CUDA forward for qwen35moe (named refusal)	#4665

Moved to 0.71

• The six known failures above (#4661, #4662, #4663, #4664, #4665, #4666).

Not measured for 0.70.2

• Capability cells (model_ladder.sh --cells), as for 0.70.1: a gap, not a pass
  (#4667).
• The full CRUX sweep (beyond the smoke above) and CPU lanes: nightly, per the scope.