[ashhart]

Wire the native Qwen4 dense pocket and learner on CUDA and Metal through the shared Living Weights family,
fitting policy, shard writer and save guard. Both native adapters feed a bounded native trainer from the serving
forward's recorded HC/GDN/QSA nodes and selected IDs. Empty and untaught pockets encode no additional GPU work;
appended checkpoint neurons are loaded separately while original packed rows and columns remain exact.

Experimental draft for #1804. Public --slide remains refused by name on qwen4 and qwen4_text
on CUDA and Metal until all six raw BF16 gradient cells pass and device qualification succeeds. Depends on the host selected-attention backward in #2021. The internal experimental writable layout is
MLX affine4/group64 MLPs, BF16 or affine4/group64 other training projections, 2 through 64 layers, and no PLE.
PLE backward and unsupported quantization/layouts refuse explicitly. Other family adapters retain their refusals.

Host validation includes actual native launch wrappers with recording devices, nonzero native forwards,
untaught-stream isolation, failed-completion retention, atomic pocket replacement, shared learner refusal,
shared save failure/retry/reopen, independent GDN/QSA and full-graph frozen-store finite differences, accumulator
witnesses and seven real source mutations. The final focused host/contract graph passes 293 tests; admission
checks pass 16, and the separate host/shared/golden/fuzz graph passes 166 with zero unexpected golden differences.
The plain Linux cross-build passes. Lean adds zero problems over the pinned 424 inherited findings.

The raw BF16 whole-graph rule-v2 panel failed five of six tiny synthetic cells, with unchanged selected IDs.
Those failures are retained and reproducible with zig build diagnose-qwen4-raw-gradient-host -j1. The passing frozen-store derivative checks do not resolve or replace them.
Successful real-model keep/recall/removal, raw gradient qualification, cross-backend saves, off-path tokens/speed,
CUDA compiler/device limits and complete Metal numeric/resource qualification are still owed. The full
device-bearing standard test graph is also owed on this tree. No support or performance qualification
is inferred from compilation, recording devices, or the ideal host attention formula.

The native four-kernel selected-attention chain now agrees with the independent host backward from #2021
at 17 causal rows, 24 query heads, four KV heads and width 256, including mixed sparse/dense selections
and early key/value gradients from answer-only loss. This comparison uses continuous FP32 softmax probabilities;
it does not qualify CUDA probability rounding. The off-path dense input store is gated on an active learner,
and the primitive contract callback table has static storage. The follow-up host graph passes 175 tests in 47 steps; three planted source controls fail and restore,
and lean adds zero findings. No device-bearing full-graph rerun applies to this follow-up tree.