[SKEEP-008]: Precision and encoding conversion #1433

Proposal document: docs/modules/skeep/pages/008-precision-and-encoding-conversion.adoc (#1434)
Status: Implemented
Branch: feature/skeep-008-precision-and-encoding-conversion

Trigger

Public Kotlin API (NeuralNetworkDsl members, TensorStorage) · DSL syntax/semantics (what a layer's dtype means, where a cast happens) · compatibility/migration policy (removing members that always threw, rules for precision loss).

Summary

Two public entry points promise to change a tensor's numeric form and do not work:

• Mixed precision in the DSL. dense<TLayer>(…) ×2 and stage<TStage>(…) are TODO() in both NeuralNetworkDslImpl and StageImpl. That's the six stubs allowlisted in config/stub-allowlist.txt. They can't be implemented as declared, because a type parameter on an interface method can't be reified, so the implementation never gets the layer's KClass.
• Encoding change after load. TensorStorage.repackTo throws for every encoding change.

The proposal:

• takes the dtype as a value (dense(64, dtype = FP16), stage("head", dtype = BF16)) behind one PrecisionBoundary module, using the existing TensorOps.convert, which is traced and lowers to stablehlo.convert;
• keeps repackTo lossless and adds an explicit, lossy requantizeTo on top of the existing Requantizers registry, so it accepts the same encodings as loader quantize-at-load;
• retires the unused MixedPrecisionModule / InternalMixedPrecisionModule / MixedPrecisionTensorOps scaffold.

Related DARC features

• [Feature]: Close the remaining NotImplemented stubs outside the operator surface #1414 (remaining non-op stubs umbrella). The mixed-precision stubs were deliberately not filed there and wait on this proposal.
• [Feature]: TransposedConv2d — convTranspose2d op, CPU kernel, backward and StableHLO lowering #1371 / [Feature]: GroupNormalization forward pass (groupNorm {} DSL currently crashes) #1379 (layer gaps found by the same inventory).

Sub-issues

No sub-issues were filed; the phases shipped in three PRs:

• B1 lossless repackTo + requantizeTo — feat(storage): lossless repackTo and explicit lossy requantizeTo (SKEEP-008 phase B1) #1470
• A1 DSL PrecisionBoundary, inference only — feat(nn): mixed-precision dense/stage via PrecisionBoundary (SKEEP-008 phase A1) #1469
• A2 export test (stablehlo.convert at boundaries) — feat(nn): finish SKEEP-008 — export test, differentiable convert, scaffold removal (phases A2, A3, A4) #1492
• A3 convert becomes @Diff — feat(nn): finish SKEEP-008 — export test, differentiable convert, scaffold removal (phases A2, A3, A4) #1492
• A4 removal of the deprecated overloads and scaffold — feat(nn): finish SKEEP-008 — export test, differentiable convert, scaffold removal (phases A2, A3, A4) #1492

Status upkeep

• Proposal registered in docs/modules/skeep/nav.adoc and the "Current Proposals" table
• Maintainer moved status to Accepted (went straight to Implemented in feat(nn): finish SKEEP-008 — export test, differentiable convert, scaffold removal (phases A2, A3, A4) #1492)
• Implementation PR(s) linked here and flipped the proposal's Status: to Implemented: B1 feat(storage): lossless repackTo and explicit lossy requantizeTo (SKEEP-008 phase B1) #1470, A1 feat(nn): mixed-precision dense/stage via PrecisionBoundary (SKEEP-008 phase A1) #1469, A2 + A3 + A4 feat(nn): finish SKEEP-008 — export test, differentiable convert, scaffold removal (phases A2, A3, A4) #1492