[michalharakal]

Finishes SKEEP-008 (precision and encoding conversion). B1 (#1470) and A1 (#1469) shipped in 0.63.0; this PR carries the three remaining phases and flips the proposal to Status: Implemented.

Tracking issue: #1433 — it can be closed when this merges.

A2 — export

• MixedPrecisionExportTest traces an FP32 network with one BF16 dense and one FP16 stage and pins the result: one convert pair per boundary in the graph, inner layers typed in the narrow dtype, and stablehlo.convert f32 -> bf16, bf16 -> f32, f32 -> f16, f16 -> f32 in the module. DtypeForwardPropagationPass leaves the boundaries alone.
• Fix: a same-dtype boundary (stage("body", dtype = FP32) in an FP32 network) still recorded convert — the CPU backend returns its input, but the tracing wrapper records the call — so the export carried two stablehlo.convert f32 -> f32. PrecisionBoundary now passes a tensor of the target dtype through without calling convert.

A3 — autograd

• TensorOps.convert is @Diff. convertBackward treats the cast as the identity and casts the upstream gradient back to the source dtype, so a gradient entering an FP16 / BF16 region is rounded to that format, as the forward values are.
• A cast from or to an integer dtype has no gradient. A recorded cast whose input and output are one tensor returns none, otherwise its gradient would be added to itself.
• MixedPrecisionBackwardTest (7 tests): the rule itself, central finite differences through FP16 boundaries, parity with the all-FP32 network for a BF16 dense + FP16 stage, and an SGD run that lowers the loss and updates every weight behind a boundary.
• The inference-only notes are removed from the KDoc and mixed-precision.adoc, which gains "What an export looks like" and "Training through a boundary". Ops matrix Backward column: 60 / 60.
• The SKEEP asked for a phase-1 test pinning "gradients stop at a boundary". It was never written and is not added here, since this PR removes the restriction.

A4 — cleanup

• Removed the generic dense<TLayer>() ×2 and stage<TStage>() from NeuralNetworkDsl and both implementations (deprecated at ERROR in 0.63.0).
• Removed MixedPrecisionModule, InternalMixedPrecisionModule, MixedPrecisionTensorOps (deprecated in 0.63.0, never implemented).
• Also removed MixedPrecisionException, which lived in the same file and which only the scaffold could throw. It was not marked deprecated in 0.63.0.
• The DType compatibility / promotion tests from MixedPrecisionTensorOpsTest move to types/DTypeCompatibilityTest.

Acceptance criteria

• repackTo(Dense(4)) is now checked bit-exact for Q5_0, Q5_1, Q5_K, Q6_K and BITNET_B1_58 as well (synthetic blocks; no quantizer produces them).
• One FP32 network with both a BF16 dense and an FP16 stage runs against the FP32 reference.
• SKEEP-008: Status: Implemented, an implementation-status table, and Open Questions 1 and 3 marked resolved (2 and 4 stay open).

Notes for review

• API: convertBackward added to DifferentiableTensorOps / DefaultGradientTape; the removals above. .api files are regenerated.
• ops-status-matrix.adoc was edited by hand (the convert row and the total), not regenerated with generateDocs.
• README is untouched: its "inference only" and "SKEEP-008 draft" lines are in "What's New in 0.63.0" and describe that release.

Test plan

• ./gradlew apiCheck checkUnmarkedStubs
• jvmTest for skainet-lang-core, skainet-lang-models, skainet-compile-dag, skainet-compile-hlo, skainet-compile-opt, skainet-backend-cpu
• linuxX64 compile of skainet-lang-core, skainet-compile-dag and the skainet-compile-hlo tests
• Native and wasm test targets (CI; not run locally)

🤖 Generated with Claude Code

[github-actions]

📖 Documentation Preview

The documentation has been built successfully for this PR.

Generated Files:

• Operator documentation: docs/modules/operators/_generated_/
• JSON schema output: operators.json

Artifacts:

• Download the documentation-preview-1492 artifact to view the complete documentation locally.

This comment will be updated automatically when the PR is updated.

[github-actions]

📖 Documentation Preview

The documentation has been built successfully for this PR.

Generated Files:

• Operator documentation: docs/modules/operators/_generated_/
• JSON schema output: operators.json

Artifacts:

• Download the documentation-preview-1492 artifact to view the complete documentation locally.

This comment will be updated automatically when the PR is updated.