v0.36.0

Added

• Per-position knowledge distillation (full-sequence KD) for the
  aprender-train-distill pipeline. The existing per-row path trains on ONE
  target per window (the next token after the window); per-position KD trains
  on EVERY position (position p predicts token p+1), giving up to
  seq_len× more distillation signal per forward pass. New
  kd_step_per_position, and additive trait methods logits_per_position /
  apply_kd_gradient_per_position (teacher + student) /
  next_batch_per_position (BatchSource) whose defaults wrap the per-row
  methods — so existing providers, including the CUDA backend, compile and
  behave unchanged. Opt-in via APR_DISTILL_PER_POSITION (default off → the
  production loop is byte-identical). Contract:
  contracts/distill-per-position-kd-v1.yaml (5 falsifiers + 2 kani
  harnesses, all passing on the CPU/fixture path).
  • Scope: the CPU/fixture path is fully verified. The real throughput
    benefit needs the CUDA teacher/student to emit all-position logits (a GPU
    forward change) — until then CUDA falls back to one position via the
    defaults. That GPU per-position forward is a documented follow-up.
  • While implementing, corrected a misleading comment in
    ShardBatchSource that claimed "identity-mapping semantics": the per-row
    labels were always genuine next-token (LMBatch causal-shifted target),
    not identity. Pinned by a falsifier so it can't mislead again.

Fixed

• PMAT-706 re-land: the APR_DISTILL_MAX_STEPS=N smoke-validation mode
  announced in #1888 (v0.35.2) was never actually in pipeline.rs — commit
  52650c60c squash-dropped the implementation and shipped only the
  apr-distill-smoke-validation-v1.yaml contract. The early-break, [SMOKE]
  summary, 0-steps guard, and no-export side-effect are now implemented in
  crates/aprender-train-distill/src/pipeline.rs and bound to the contract's
  four falsifiers (pipeline::tests::pmat_706_{smoke,no_regression,summary_format,no_output_in_smoke},
  all passing). scripts/dispatch-distill-stage-d.sh now forwards
  APR_DISTILL_MAX_STEPS across the ssh/env boundary so the documented
  APR_DISTILL_MAX_STEPS=10 ./scripts/dispatch-distill-stage-d.sh actually
  triggers smoke mode (previously a silent no-op).