docs(classify): tract dies on the adaptive pool, not on attention

[jayhesselberth]

Docs only, no code changes. Targets worktree-issue-306-waveform-chunks, not
main, because that is where waveform_net.rs lives.

waveform_net.rs attributes the unloadable charging_tcn_rna004@v0.1.0 graph
to dynamo's lowering of nn.MultiheadAttention's mask handling, and explicitly
rules out adaptive_avg_pool1d ("those layers are in the model config and are
not what tract dies on"). It is the other way round. nn.MultiheadAttention
exports as plain MatMul/Softmax/MatMul; the culprit is
adaptive_avg_pool1d with an output size that does not divide the input.

The evidence is already in the file

Every measurement here was right — the five failure modes, the node names, and
the two constant initializers. The constants settle it:

charging_tcn_rna004 pools 390 down to 11. PyTorch's bin rule is
[floor(j*L/K), ceil((j+1)*L/K)), which for 390 -> 11 gives eleven bins of
width 36 or 37. A gather that evaluates every bin at once therefore needs
an (11, 37) index grid and an (11, 37) mask marking the slot the 36-wide
bins do not use — which is exactly the (1,1,11,37,2) index tensor and
(11,37) bool mask this file records. An attention mask is shaped by sequence
length and head count and would never be 11 x 37.

Why it was easy to get wrong

A non-dividing adaptive pool lowers to no ONNX pooling op at all. Grepping
the graph for one finds nothing, so the ragged-bin gather looks like it must
have come from somewhere else. The layers named in escapepod-rs#306 were the
right suspects — they just do not appear under a pooling name.

The second cause, which the table shows but the prose never named

The value_info dynamo writes for every intermediate carries the batch axis as
the symbol batch. That is the node_conv1d row: a consumer that pins the
batch cannot unify against a symbol, so it fails at the first convolution,
before ever reaching the gather. Two independent causes, not one.

Upstream status

Diagnosed and fixed in rnabioco/leech#233, released in leech 0.10.0: the
pool becomes one matmul against a constant segment-mean matrix, and every
export strips value_info. Measured there on the shipped TCNDwellResidualLN
weights, no retrain:

nodes                  479 -> 319, GatherND 2 -> 0, Gather 76 -> 0
tract, batch 1 and 32  loads, optimizes and runs (was: five failures)
tract vs torch         max |dlogit| 5.72e-06 over 256 real chunks, 0 flips

Neither cause is visible to a round-trip check against onnxruntime, which loads
the old graph happily — which is why this surfaced at integration here rather
than at build time in leech.

So the ort bridge can be removed once charging_tcn_rna004 is re-exported
and shipped (rnabioco/escapepod-models#96). Still unverified: a released
escpod running such a bundle end to end, since 0.18.1 has no waveform bundle
variant.

Scope

Corrected in all five places the claim had spread to — waveform_net.rs,
CHANGELOG.md, CLAUDE.md, docs/cli/signal-classify.md, and
examples/tract_dynamo_probe.rs — since leaving it standing in the changelog
and the agent instructions would be no fix at all.

cargo fmt --all --check and cargo clippy --workspace --all-targets --features escapepod-classify/waveform-onnx both clean.