Gate CubeCL parity and retire the OpenCL backend

Objective

Remove the deprecated OpenCL backend only after CubeCL satisfies documented functionality, performance, native/browser, fusion, and streaming gates.

Procedure

1. Freeze a parity matrix against the current OpenCL behavior and supported public API.
2. Run correctness, memory, cold/warm compilation, throughput, and block-stream benchmarks on documented hardware.
3. Publish migration and release notes, including feature-name changes and fallback behavior.
4. Deprecate opencl for at least one release unless the maintainer explicitly approves a same-release removal.
5. Remove ocl, the opencl feature, OpenCL build detection, kernels, tests, and docs.
6. Verify default/no-default/CubeCL/stream/browser build combinations.

Acceptance criteria

• #35, #36, #37, #38, #39 and benchmark migration #22 are accepted.
• The parity matrix has no unexplained functional gaps.
• Performance regressions beyond an agreed threshold are documented and explicitly approved.
• rg -n "ocl|opencl|OpenCL" Cargo.toml src tests README.md has no runtime/backend references except migration history.
• All supported feature combinations and downstream TinyChain integration tests pass.
• Rayon remains available unless a separate measured decision authorizes its retirement.
• Release notes give users an actionable migration path and rollback version.

Parent: #33

Protected retirement decision

Tembo may implement removal only after a maintainer records the accepted parity matrix, supported runtime matrix, benchmark hardware, regression thresholds, deprecation/release policy, and rollback release. Tembo may generate evidence and a draft removal PR; it may not choose thresholds, waive unexplained gaps, approve performance regressions, or make the final backend-retirement decision.

Required evidence includes host/CubeCL differential results, designated accelerator results, actual-browser WebGPU results, cold compilation versus warm execution, transfer time, peak host/device memory, fused/unfused comparisons, streamed-block behavior, downstream TinyChain integration, and exact environment identities.

automation:
  eligible: false
  executor_profile: tembo
  readiness: validation/review
  primary_repository: TinyChain-Inc/ha-ndarray
  base_branch: main
  writable_scope: []
  prerequisites: ["#41", "#42", "#43", "#44", "#45", "#46", "#47", "#48", "#49", "#39", "#50", "#22"]
  concurrency_group: opencl-retirement-decision
  required_commands: []
  external_actions: none
  draft_pr_only: true
  completion_authority: human-maintainer-release-review

Start by reviewing the parity and benchmark prerequisites in #35, #36, #37, #38, #39, #22 and the referenced Cargo.toml, src, tests, and README.md areas. Establish the required runtime and benchmark evidence, then verify supported build combinations and TinyChain integration; completion also requires maintainer approval of thresholds, migration notes, deprecation policy, and rollback release before removal.