Scale ops docs: KSP TensorOps surface scan + human-prose partials

[michalharakal]

Context

The Ops reference in the Antora site is near-empty: only 4 hand-annotated symbols produce generated scaffolds. The pipeline (OperatorDocProcessor KSP → operators.json → GenerateDocumentationTask → AsciiDoc) works end-to-end, but coverage is bottlenecked on @DslOp / @InProgress annotations, the backend matrix drifts from reality, and there's nowhere for human-authored math/intuition/examples to live alongside the generated reference.

Goal

One URL per op that fuses auto-discovered signature + backend matrix with hand-authored prose + LaTeX math, scaling across the full TensorOps surface and any number of compute backends.

Approach

Three cooperating layers, reusing the existing pipeline.

Layer 1 — KSP discovery: scan the interface, not just annotations

• New @Backend(id, displayName) class annotation in skainet-lang-ksp-annotations.
• OperatorDocProcessor walks TensorOps and every @Backend-tagged implementor, deriving statusByBackend from override presence.
• Annotations remain authoritative overrides for owner/issue notes.

Layer 2 — Generator fusion: human partials per function

• GenerateDocumentationTask emits include::partial\$ops/<op>/<fn>.adoc[tag=...,optional] for math / intuition / examples / references between auto sections.
• Missing partials render cleanly as empty — graceful scale-up.

Layer 3 — Backend matrix stays auto

• Existing ops-status-matrix.adoc already aggregates statusByBackend; once Layer 1 populates it from real impls the matrix becomes ground truth.

Status

Layers 1–3 landed for skainet-lang-core (commit referenced this issue). Coverage went from 4 → ~50 methods; void column is real. matmul migrated as the partials reference example.

Follow-up (Layer 4)

Wire KSP into skainet-backend-cpu and merge multiple operators.json files in generateOperatorDocs so the cpu and apple columns appear alongside void.