1.0 -- Advanced compiler tiers and schedule-quality follow-through

[lukstafi]

See CHANGES.md for details and ROADMAP.md for a more thorough overview than here. From the roadmap:

v1.0 — August 13, 2026 (released)

Theme: Advanced compiler tiers and schedule-quality follow-through

Advanced compiler tiers:

• Branch-and-bound schedule inference (Schedule inference as branch-and-bound: search in constraint space, legality queries as fathoming #514) — Ir.Schedule_space as a refinement tree over partial schedules, Cost_model.completion_floor is an admissible downward bound.
• Inlining as a first-class, searchable schedule decision (Retire the concrete-index tracer: derive virtualization metrics from affine access relations #554, Inlining as a first-class, searchable schedule decision #555, Reuse one Low_level analysis across sibling candidate compiles (gh-555 follow-up) #560).
• CUDA tensor-core profile completeness (CUDA tensor-core profile completeness: fp8 transposed operands, e4m3, ldmatrix/swizzled staging, Stage ~pad_stride (the T4 ceiling) #481), software-pipelined double-buffered staging (Software pipelining: double-buffered cooperative Stage (cp.async / Metal async copies) #487), CUDA/HIP graph capture of the fissioned step (CUDA/HIP graph capture of the fissioned step (generalize Metal's fused segment encoding) #488), and budget-driven rematerialization on top of the liveness planner (Rematerialization: budget-driven recompute-vs-store on top of the liveness planner (gh-489 follow-on) #498).
• CPU reduced precision: 16-bit storage with f32 compute (16-bit storage with f32 compute on CPU: convert-on-load/store for bandwidth-bound kernels #517) and native fp16 arithmetic (Native fp16 arithmetic on the CPU backends: admit _Float16 to the vectorized and tensorized renderings #516).

Schedule quality — the gpt2_mini arc (done):

• The v0.9 sweep left gpt2_mini 72x off torch CUDA; Companion-coverage rule pins five FFN-class kernels to 1.3% of fp32 peak — 70% of the gpt2_mini step, ~2.4x measured floor #569 took the tuned step 107.4 → 52.4 ms on CUDA and 45.6 → 25.4 ms on HIP. Autotune sketch seeding: rank-3 (batched/attention) matmul sites are never proposed, so gpt2_mini cannot reach tensor cores even after #521 #528 made tensor cores reachable on transformer workloads at all.
• CUDA bf16 mma timed scalar-fallback code under an mma label.
• Search-plumbing consolidation.

Frontend, configuration and diagnostics (done):

• Shape inference: "close down when known" narrowed to the leaf-tensor rule it always was, with stretch requesting use-site resolution by name (Shape inference: 'close down when known' for operations, empty batch for parameters — use-site row resolution is a leaf-tensor rule applied too widely #544).
• Config profiles reproducible / performance with picker-inherited precedence (Config profiles: reproducible vs performance preset bundles, with picker-inherited precedence #559); config startup chatter moved off stdout (Config startup chatter prints to stdout, polluting every ocannl-linked executable's redirectable output #581).
• Site-targeted materialization: the capability exists twice over and ships nowhere measured (Site-targeted materialization: reach the tensorized candidate family without the materialize-all arm #558).

Infrastructure:

• CI moved to OCaml 5.5, Windows off the per-PR path onto a twice-weekly schedule, and the GPU backends onto a daily cross-machine sweep (tools/sweep.sh).

---

This discussion was created from the release 1.0 -- Advanced compiler tiers and schedule-quality follow-through.