Carry layout/placement decisions to StableHLO as untyped metadata

[michalharakal]

Follow-up of #1134. The #1134 resolution is decide-and-carry: core resolvers (WeightFormResolver, AllocationResolver) decide; the recorded path's job is to carry those decisions to the MLIR so the downstream compiler can honor them. This issue is the carry slice — deliberately deferred until an in-repo IREE consumer exists (today it would be write-only metadata).

Scope (when unblocked)

Follow the TensorSpecEncoding precedent — untyped TensorSpec.metadata + typed accessor extension; sk.ainet.lang.memory types are never imported into skainet-compile (rule recorded in the SKEEP resolution doc):

• Tape→graph: extend TensorRef (id/encoding, optionally layout) in TraceSession.refOf; stop dropping metadata in TraceToGraphBuilder.buildInputSpecs/buildOutputSpecs.
• Graph: TensorSpecLayout.kt mirroring TensorSpecEncoding.kt; a decision pass in skainet-compile-opt (after TransposeEliminationPass, before LLMFusionPass), registered via TargetOptimizers; populate the pre-built dead seams ResolvedComputeGraph.resolvedLayout(edgeId) / backendAssignment(nodeId).
• Emission: branch at the designated point in dag2hlo ; module attribute skainet.tensor_layouts serialized structurally (strides/blockOrder/blockSize — not name-strings, which lose blockSize today); later per-type encoding attrs via TypeMapper (requires auditing the ~102 raw "tensor<…>" literals in converters).
• Boundary: ExternalParameterRef gains layout/blockOrder so kernel-feed-order weights can be blitted verbatim into .irpa with their order declared in the MLIR reference.
• Thread target for real (ConversionContext.target is currently never set and never read).

Gate

Blocked on an in-repo IREE consumer (see the .vmfb parity issue). Do not start before that exists.

[michalharakal]

Order change (michalharakal's call): carriage is implemented before the #1148 parity harness, not after. Until the harness exists, the readers of the carried metadata are the tests themselves — trace-roundtrip assertions and MLIR-text assertions — and the external SKaiNET-IREE-tools consumer once the schema stabilizes. The dependency rule stands unchanged: sk.ainet.lang.memory is never imported into skainet-compile; carriage uses the TensorSpecEncoding untyped-metadata precedent throughout.

Sliced per the standing working rules (one sub-issue = one branch = one PR, develop green after each):

1. Tape→graph plumbing — TensorRef gains tensor identity + encoding; TraceToGraphBuilder stops dropping them; TensorSpecLayout-style typed accessors. Pure capture, no behavior change.
2. Export boundary — structural skainet.tensor_layouts module attribute (strides/block-order/block-size, not name strings) + ExternalParameterRef layout/block-order, so a feed-ordered weight declares its order to the .irpa consumer.
3. First decision pass — LayoutAssignmentPass for one target wired through TargetOptimizers into the production tape→HLO path, populating the pre-built ResolvedComputeGraph.resolvedLayout/backendAssignment seams.
4. (later) Per-type MLIR encoding attributes + the raw tensor<…>-literal audit + real target threading.

[michalharakal]

Done — all three carriage slices are in develop (#1178 → PR #1182; #1179 + #1180 → PRs #1183/#1184).

What the compile lane now carries, end to end, all under the standing rule that sk.ainet.lang.memory is never imported into skainet-compile (untyped TensorSpec.metadata + typed accessors, block order always a stable string):

1. Tape → graph: TensorRef carries the live tensor's module-path identity (TraceSession.identify, registered by HloGenerator from trainableParameters()), its encoding as an object (block size intact), and its packed block order. Four spec-construction sites propagate them.
2. Graph → MLIR / .irpa: the skainet.tensor_layouts module attribute declares machine-readable structure per tensor — {kind, block_elems, block_bytes, bits, block_order} — off the encoding objects' own constants. The redundant skainet.tensor_encodings names dictionary was removed outright (org-wide search: no reader outside this repo). ExternalParameterRef declares block order to the .irpa consumer, so feed-ordered weights are blitted verbatim and say so.
3. The decision pass on the production path: HloGenerator.generate(target = …) runs dagPipelineFor(target) with LayoutAssignmentPass as a core pass (default null is byte-identical); rank-2 GGML-block weights get kernel-feed order, the tape's carried facts are never overridden, and the pre-built ResolvedComputeGraph seams surface exactly the decisions made.

Consumer: the conformity pipeline on the SKaiNET-transformers side, after the core release — per the strategy on #1148, which stays open as the placeholder for the IREE-validation SKEEP. Until then the carriage tests are the metadata's reader.