[PRIHLOP]

Preserve Qwen3.8-27B precision and validate full-depth decode(closes #148)

Preserve FP32 activation boundaries, integrate compensated DeltaNet readout,
and use consistent quantized activation sums in Q4 projections. The production
64-layer decode now passes the previously failing DeltaNet input gate, with
unchanged acceptance thresholds and weight packing.

Changes

• Preserve FP32 normalization, DeltaNet convolution history and split-FFN
  boundaries; use paired-BF16 KV with FP32 attention output-to-projection.
• Use triple-BF16 state updates and compensated factored DeltaNet readout.
  Share arithmetic helpers to fit full-layer program memory.
• Use the same INT16-reconstructed activation in both affine Q4 terms of direct
  FP32 preparation. Select it in full lx/ax only for the validated wide dense
  Q4_1 geometry; other geometries and Q8/mixed-role recipes retain existing arithmetic.
• Declare precision and arithmetic in the manifest, including
  q4_activation_sum: quantized_fp32; reject incompatible decode bundles.
  Preserve weight packing, buffer ABI and scheduling.
• Add independent causal references, controlled replays, capture/hash checks,
  compiled-header/resource audits and reproducible build runbooks.
• Add a production-arithmetic FFN probe and all-layer sweep with separate up/gate
  captures, byte-exact production norm/activation identity and streamed weights.
  Pin input/build/source provenance and reject partial fixtures or stale results.
• Extend the independent prefill oracle with FP32 conv history and paired-BF16 KV.
  Add a native host-stage boundary checker with compiler/source provenance,
  byte-exact partition parity, independent numerical gates and padding checks.
• Preserve FP32 history in rollback tapes, validate history format before state
  mutation and clear stale history when reusing tapes. Adapt manifest tests to
  move-only ownership by loading independent fixtures.

Validation

• TDD: 1380 passed, 47 skipped; native CTest 4/4. The native manifest
  test also accepts the actual new 27B export.

• Fresh export: 41 kernel sets / 82 verified binary hashes. lx/ax maximum
  text 15936 / 13760 bytes, L1 buffer ends 62336 / 59920, stacks 6144.
  All 66 model/pool hashes and 195 regenerated CPU references match provenance.

• NPU 64 layers × 3 positions, 198 dispatches: 192/192 residuals pass,
  worst error 0.0030399006 <0.005. Logit correlations
  0.9999989597 / 0.9999982780 / 0.9999986060; argmax 8678 / 198 / 2 matches CPU.

• DeltaNet input and composed state/output: 144/144 pass. Worst normalized
  K error is 0.0009785607 <0.005 at L34/T0. All 6912 strict head cases pass
  at maxrel <1e-4 and cosine >0.9999999; worst state/output errors
  1.2624e-6 / 2.3722e-6. History is byte-exact and padding zero.

• Attention: 1152/1152 heads, 48/48 projection/sublayer cases; LM-head 3/3.
  FFN front and split-down: 192/192 each; worst front error 0.0002516594,
  down partials 0.0000958907 / 0.0000693562. All 7 boundary checkers pass.

• Full-model resident runtime: all three logits are byte-identical to the
  harness after a repeated request. Full-buffer reset: 0/3 differing replays.

• Separate FFN probe: 64 layers × 3 positions, 192/192 pass against unrounded
  FP64 norm/up/gate/activation references. Worst up/gate errors:
  0.0000919118 / 0.0001016790 <0.005. Norm and activation match production
  bytewise in 192/192 cases each. One pool BO uses 0.217 GiB logical NPU
  buffers; the production export and all 82 binary hashes remain unchanged.

• Host prefill: 5/5 wide cases (1, 2, 255, 256, 257 tokens plus one continuation)
  and 5/5 small cases pass. Outputs, conv/state and KV are byte-identical
  between blocked and sequential native calls. Independent FP64 worst errors:
  DeltaNet 3.9750e-7, attention 8.8449e-7, state 2.5863e-7,
  reconstructed cache 6.1067e-6, all below 1e-3. Conv history is exact;
  cache prefix/tail and poisoned-padding checks pass.

Scope and remaining acceptance

This validates the tested three-position sequential decode and resident
lifecycle, supplemented by isolated per-layer FFN projection checks and synthetic
host prefill boundaries. Host checks condition on synthetic projections; actual
GEMM boundaries and complete model prefill remain unvalidated. Remaining
source/packing checks, full block-prefill parity, serving isolation and final resource/performance
acceptance remain open. Paired-KV block attention uses host attention; this PR
does not establish fully NPU prefill or long-context accuracy. No throughput
claim is made.

Existing model weights require no repacking. Weights, build/capture artifacts
and this description remain outside Git.