Post-#674: conv1d/gather/pooling output-spec + reduce_window emission still block iree-compile

[michalharakal]

Follow-up to #673 / #674. #674 fixed reshape/matmul/concat output-spec inference; verified end-to-end against an unsigned 0.28.1-SNAPSHOT published to mavenLocal and run through the skainet-iree-conformance harness:

• op micro-suite: 23/27 iree-compile to a vmfb (was 20/27 on 0.28.0; reshape/matmul/concat now pass)
• models: 6/7 compile (grayscale, tiny-mlp, whisper, yolo, leaf-embed, gemma3-260m)

Still failing — this issue: conv1d, gather, maxpool2d, avgpool2d (and the mnist-cnn model, via maxpool).

Repro test (committed, RED)

Branch test/dag-conv-gather-pool-shape-673b → DagConvGatherPoolExportTest (real dag{} DSL path). 4 tests, all RED on develop after #674.

Two root causes

1. inferDagOutputSpecs has no rule for conv1d / gather / pooling

Same gap #674 closed for matmul/concat — these still fall back to echoing operand-0's shape, so the declared result/return type contradicts the value iree-compile infers:

op_conv1d  (x:1x3x8, w:4x3x3, stride 1 pad 0)  declared 1x3x8, true 1x4x6
  error: inferred shape '[1, 4, 6]' is incompatible with return type 'tensor<1x3x8xf32>'

op_gather  (table:8x4, idx:3, axis 0)          declared 8x4, true 3x4
  error: 'stablehlo.gather' op inferred type(s) 'tensor<3x4xf32>' are incompatible with return type(s) 'tensor<8x4xf32>'

op_maxpool2d / op_avgpool2d (1x3x8x8, 2x2 s2)  declared 1x3x8x8, true 1x3x4x4

Note: conv2d already passes — only conv1d is wrong. Extend inferDagOutputSpecs with conv (out = floor((in + 2·pad − dilation·(k−1) − 1)/stride)+1, channels = weight[0]), gather (indices.shape ⊕ table.shape without the gathered axis), and pooling (same windowed formula as conv).

2. reduce_window is emitted in a form IREE rejects

op_maxpool2d:5:11: error: custom op 'stablehlo.reduce_window' has no custom assembly form

The converter emits the pretty stablehlo.reduce_window(%a, %init) applies stablehlo.maximum over window dimensions = … form. IREE's parser needs the generic region-based form ("stablehlo.reduce_window"(%a, %init) ({ ^bb0(...): stablehlo.return ... }) {window_dimensions = …, …}). This blocks both pooling ops and the mnist-cnn model. (Same class as the gather/slice emission fixes in 0.27.0.)

Acceptance

• DagConvGatherPoolExportTest (4) green.
• Harness op micro-suite reaches 27/27 compile and mnist-cnn compiles to a vmfb on llvm-cpu.

Verified via mavenLocal 0.28.1-SNAPSHOT + skainet-iree-conformance.

[michalharakal]

Housekeeping: the repro branch test/dag-conv-gather-pool-shape-673b referenced above has been deleted now that this is fixed (PR #676). The repro test lives permanently on develop/main:

skainet-compile/skainet-compile-hlo/src/jvmTest/kotlin/sk/ainet/compile/hlo/DagConvGatherPoolExportTest.kt

— and is green there, with the original 4 cases (conv1d / gather / maxpool2d / avgpool2d) plus a 5th (flatten_preserves_leading_batch_dim) added in #676.