Support bounded streams of Array blocks through CubeCL and fusion

Objective

Execute an ordered asynchronous stream of Array blocks through one reusable CubeCL/fusion plan with bounded host/device memory, backpressure, cancellation, and typed per-stream failure behavior.

Contract

Define a public, backend-neutral adapter for a fallible stream of Array blocks. Blocks may vary only along explicitly declared streaming dimensions; dtype/rank/shape violations fail visibly. Preserve input order unless an API explicitly opts into reordering.

Procedure

1. Specify stream shape/dtype metadata and terminal/error/cancellation semantics.
2. Plan and compile compatible operations once, then reuse the plan across blocks.
3. Bound in-flight blocks, host buffers, device allocations, and compilation work.
4. Overlap transfer/compute only when the runtime supports it and ordering remains correct.
5. Release buffers promptly on cancellation, downstream drop, or error.

Acceptance criteria

• A stream larger than device memory completes with bounded measured peak allocation.
• Backpressure prevents unbounded buffering when the consumer is slow.
• Fused plans are reused across compatible blocks without recompiling each block.
• Ordering, partial-progress, cancellation, upstream/downstream error, and cleanup tests pass.
• Incompatible block metadata returns a typed error naming the offending block.
• Results match equivalent sequential Array execution.

Parent: #33

Tembo-sized implementation issues

• #48 — ordered Array-block stream contract and bounded backpressure
• #49 — fused CubeCL block-stream execution and memory validation

Dependency order: #48 plus accepted fusion #47 → #49. Host sequential execution remains the result oracle.

Tembo program boundary

This parent is a non-executable acceptance rollup. Tembo must be assigned to the child issues, and may not close this issue from a child PR.

automation:
  eligible: false
  executor_profile: tembo
  readiness: rollup
  primary_repository: TinyChain-Inc/ha-ndarray
  base_branch: main
  writable_scope: []
  prerequisites: ["#48", "#49"]
  concurrency_group: cubecl-stream-rollup
  required_commands: []
  external_actions: none
  draft_pr_only: true
  completion_authority: human

This is a non-executable acceptance rollup; start by reading child issues #48 and #49, then review the host sequential execution used as the result oracle. The work is done when the child contracts and fused CubeCL execution satisfy the listed bounded-memory, backpressure, ordering, error, cancellation, cleanup, and equivalence tests.