Milestones
List view
The three 0.5 enhancements that need real work rather than a fix: training resume (#72), a re-runnable benchmark harness (#71), and the end-to-end acceptance workload (#76) that builds on both.
No due date•0/3 issues closedWhat the framework claims, checked against what it does. Every item was found by exercising `main` at `9e009c1` as a downstream consumer — building the workspace, driving the CLI, and probing the public tensor surface op by op — and every one carries a reproduction that was run, not inferred from reading the source. Two of them are correctness: `eigh` answers a question about a matrix the caller did not pass, and nothing checks that the shipped PTX was built from the shipped CUDA source. Three are the project describing itself as smaller or older than it is — a `doctor` capability list frozen before 0.2.0, crates.io descriptions two releases behind, and an undocumented dtype boundary at `nn`. One is packaging: the CUDA backend is mandatory on macOS. One is a decision that was deferred rather than made: four reductions over an empty extent behave four ways. The design bar for this milestone is that each fix leaves behind something that fails next time. A refreshed capability list decays again by 0.5.0; a test that fails when a release adds a capability and does not list it does not. Where a fix is a one-line string, the issue asks for the guard as well. Deliberately out of scope: throughput work of any kind (Metal buffer pooling, cuBLAS, multi-stream, `f16`), `f64` on the GPU backends, general LU/solve/SVD, and device-side `argmax`. Those are feature work with their own measurement requirements; this milestone is about the gap between the documentation and the binary. Also out of scope: making `nn` support `f64` — issue #39 asks only that the boundary be stated and the error be legible, and records the API question as deferred on purpose. Two candidates were investigated and discarded rather than filed. `Tensor::randn` looked like a reproducibility hazard until `randn_with_seed` turned up beside it and every `nn` constructor turned out to take an explicit seed — that design is right as it stands. And `--epochs 0` exiting 0 having trained nothing is defensible: zero epochs is zero work.
No due date•8/8 issues closed