Close v0.1: opinionated main + harness complete

[debashishc]

Summary

Close v0.1 by making main opinionated and harness-first, with Week 0–2 ops present (implemented or explicit ref-backed stubs) and docs aligned with actual support.

Scope

• Week 0: copy/transpose (minimal working kernel)
• Week 1: reduce_sum (ref-backed stub is acceptable)
• Week 2: softmax_online (ref-backed stub is acceptable)
• Harness and docs are the primary deliverable

Out of Scope for v0.1

• FA / KV-cache / FP8 / NCCL / C++ extension builds
• Aggressive perf tuning beyond correctness + benchmark scaffolding

Checklist

• Define “opinionated main” policy in README or DEVELOPMENT
• Harness commands succeed on supported GPU:
  • uv run python -m forge_cute_py.env_check
  • uv run pytest -q
  • uv run python bench/run.py --suite smoke
• Copy/transpose: minimal working kernel + correctness tests
• reduce_sum: explicit ref-backed stub (documented)
• softmax_online: explicit ref-backed stub (documented)
• Update README kernel status + ROADMAP to reflect v0.1 closure
• Update CHANGELOG for v0.1 release notes
• Resolve code organization guideline
• Fix testing command docs

Sub-Issues

Testing & Benchmarking

• Add dim=-1 tests + benchmark shapes for reduce_sum (opinionated) #29 Add dim=-1 tests + benchmark shapes for reduce_sum
• Add dim=-1 tests + benchmark shapes for softmax_online #30 Add dim=-1 tests + benchmark shapes for softmax_online
• Add explicit tests + benchmark shapes for copy_transpose #31 Add explicit tests + benchmark shapes for copy_transpose
• Add short/long benchmark suites with kernel-specific shapes #32 Add short/long benchmark suites with kernel-specific shapes
• Add a Modal benchmark runner (CLI + script) to run run.py remotely via modal run. #28 Add Modal benchmark runner (CLI + script)

Documentation & Policy

• [RFC] Code Organization Guideline #4 [RFC] Code Organization Guideline
• Testing command doesn't work #15 Testing command doesn't work

Harness & Week 1 Work

• Opinionated reduce_sum harness (no variants) #23 Wire reduce_sum variant dispatch and benchmarking scaffolding

Acceptance Criteria

• All checklist items complete
• Docs clearly state what is implemented vs ref-backed
• Harness is the gate for correctness
• No WIP-only code on main