[sablin39]

What

Four CUDA TIR emission gaps that fire on scheduled programs finalized
against the composed cta×thread mesh (the shape every multi-CTA
schedule finalize produces). One commit, src/tilefoundry/codegen/cuda/tir
only, +45/−10.

1. arith.py — effect-form Cast emitter. Cast in effect form had
   no registered EMIT handler, so any scheduled program whose TIR carries
   one raised "no codegen handler". The store into the destination
   converts, so the elementwise functor is the identity — the same shape
   as the existing UnaryKind.CAST path, dynamic-extent clip included.
2. fill.py — Fill on a kernel parameter. Fill named the raw
   pointer for a kernel-param destination while tensor_view establishes
   the <name>_tensor view for params; the emitted kernel referenced an
   undeclared name and did not compile. Now routes through
   ctx.is_kernel_param, matching tensor_view's naming.
3. tensor_view.py — slice pointer offsets through _coord_ref.
   PtrOf(Slice(...)) offsets were rendered with render_scalar_expr,
   which cannot resolve mesh-relative Dim arithmetic; slice coordinates on
   Dim expressions failed or mis-rendered. _coord_ref resolves them
   against the mesh binding. Also teaches _COORD_OPERATORS
DimFloorDiv/DimMod and renders MeshCoord call targets as
   cute::get<axis>(cute::flatten(mesh_coords(...))).
4. tensor_view.py — named refusal for unknown coordinate calls.
   An unsupported Call target in a coordinate previously fell through to
   a generic leaf path; it now raises a NotImplementedError naming the
   target type.

May close part of #230 (sliced pointers / nested layouts) — the slice
coordinate half overlaps; happy to rebase if that lands first.

Validation

• Upstream test suite at aaa8054 with this commit: 1802 passed,
  11 skipped (pytest tests -n 4). The single failure in that run,
  tests/cli/test_cli.py::test_target_list_expressions_execute_and_show_accepts_their_identities,
  is an artifact of the development environment (a locally registered
  custom target in the CLI registry); it passes with a pristine registry
  and is unrelated to this change.
• The four gaps were hit and the fixes first exercised end-to-end while
  driving a scheduled RMSNorm through scaffold/compile() on sm_90.

Not in scope

• CopyAsyncTensor (TMA), WGMMA (TiledMma non-SM80 atoms), and Where
  emitters remain absent as documented in
  docs/spec/codegen.md
  and
  docs/spec/tir.md
  — the schedule tutorial's optimal GEMM still cannot compile. Deliberate
  raises today; unchanged here.
• A separate runtime bug: the reduce dispatch mis-counts
  warps_per_group for a composed cta×thread mesh (single is_thread
  per mesh in
  reduce_common_impl.h,
  so thread axes inside a cta-scoped composed mesh count their full
  extent as warps → OOB smem reads on an RMSNorm repro). Not touched by
  this PR; will file separately with the repro.

[zhen8838]

I’m sorry that I haven‘t had time to do much work on CodeGen and Runtime recently. I really want to further improve the runtime of tilefoundry to the extent that it can be used as a distributed runtime lib. So thank you very much for your contribution to the tilefoundry project. Help us take a step closer to that direction.