spec: extend orca-lang with neural and sparse capabilities

[jascal]

Summary

• Spec-only openspec change at openspec/changes/extend-orca-lang-neural-and-sparse/ — proposal, design, tasks, language + verifier delta specs, and three worked examples.
• Adds three type primitives (tensor<dtype, shape>, weight_ref<path, sha256, key>, sparse_matrix<dtype, shape, format>), four sections (## weights, ## cluster_manifest, ## faithfulness, ## provenance), and six ## verification rules invariant kinds.
• Frames neural + sparse as one coherent grammar extension, not two new dialects: both share the classical data model, so the type-system work would otherwise be duplicated. q-orca stays a separate dialect on data-model grounds — that rationale is cited in design.md (decision D1).

What's in the change

• proposal.md — Why / What Changes / Capabilities / Impact / Scope
• design.md — eight decisions (D1 extension-not-dialect; D2 type primitives; D3 four sections; D4 six invariants; D5 neural module = machine with tensor states; D6 sections-only artifact; D7 binary-absent warning policy; D8 capability negotiation), risks, migration, open questions
• tasks.md — 9 sections of work items + 7 parked follow-ups
• specs/language/spec.md — 9 ADDED Requirements with WHEN/THEN scenarios
• specs/verifier/spec.md — 1 MODIFIED + 9 ADDED Requirements; 12 new error/warning codes
• examples/ — tiny-transformer.orca.md, sparse-sae-16f.orca.md (sections-only), hybrid-transformer-sparse-head.orca.md

Design calls worth checking

• One change, not two — the type primitives are shared and the hybrid example mixes ## transitions + ## cluster_manifest in one file.
• Inlining tensors in markdown is forbidden; weight_ref is the only supported mechanism.
• ## provenance is genuinely new (no front-matter convention exists in the parser).
• Missing weight binaries are a warning, not an error, so artifacts travel cleanly for review.
• Soft dependency on a future add-orca-schema-version change; this one documents the neural-and-sparse-v1 capability string but doesn't enforce a version field.

Out of scope (noted in proposal)

• Implementation. Spec only.
• Inlining weights into markdown.
• Compile back-ends / hardware targeting (parked: add-neural-compile-targets).
• Schema versioning (parked: add-orca-schema-version).
• Runtime faithfulness check (parked: add-runtime-faithfulness-harness).
• Algebraic / probabilistic / circuit extensions (a-orca / p-orca / c-orca).

Test plan

• Read proposal.md end-to-end; confirm the framing (one coherent extension, not two new dialects) is correct for this repo's trajectory.
• Walk design.md decisions D1–D8; flag any conflict with existing conventions I missed.
• Skim specs/language/spec.md and specs/verifier/spec.md for ADDED Requirements that overlap existing requirements.
• Open the three example files and confirm each one parses-on-paper against the spec'd grammar.
• Confirm the capability-string name neural-and-sparse-v1 is acceptable, or rename in coordination with the schema-version change.

🤖 Generated with Claude Code