oxmera 0.5.1 — ten defects found by deep-testing the published 0.5.0

[vyncint]

oxmera 0.5.1 — the defects deep-testing the published 0.5.0 found

0.5.0 was tested as a repository. This release is what happened when it was
tested as a published artifact — a downstream project depending on
oxmera = "0.5.0" from crates.io. Ten defects, each reproduced by running
code, each now carrying a regression test. No public API changed, so
semver-checks passes as a patch.

Three of these are defects in 0.5.0's own new work, which is the honest
argument for testing the artifact rather than the tree.

Wrong numbers

• maximum/minimum gave a tie's whole gradient to one operand (maximum/minimum give a tie's whole gradient to one operand, so BCEWithLogitsLoss is wrong at a logit of 0 #84).
  BCEWithLogitsLoss is max(x, 0) - x·t + …, so at a logit of exactly
  0.0 it returned -t instead of sigmoid(0) - t. On a balanced batch the
  true bias gradient is 0, oxmera reported -0.5, and one SGD step raised
  the loss from 0.6931472 to 0.724077. A central finite difference of
  oxmera's own forward disagreed with its own autograd. Ties now split
  evenly — the convention docs/LIMITATIONS.md already documented.
  relu'(0) = 0 is untouched: relu has its own VJP.
• mean on an f64 tensor was scaled by an f32 reciprocal (mean on an f64 tensor is scaled by an f32 reciprocal, so f64 means carry f32 precision #87), so
  every f64 mean carried f32 precision: mean([1,2,3]) = 2.0000000596046448.
• Adam/AdamW bias correction used one global step counter (Adam/AdamW bias correction uses one global step counter, mis-scaling a late-arriving parameter's first update #88), so a
  parameter whose gradient arrived late had its first update mis-scaled —
  identical first gradients produced updates of 0.100 and 0.074. 0.5.0 made
  skipping grad-less parameters normal, which makes this routine.

Panics where a typed error or a value was promised

• Tensor::from_storage's bounds check overflowed isize (Tensor::from_storage's bounds check overflows isize: panics in debug, accepts out-of-bounds layouts in release #85) — a
  regression in the check 0.5.0 added for Tensor::from_storage accepts a layout with negative strides and the first read panics #50. It panicked inside a
  Result-returning constructor in debug and accepted out-of-bounds
  layouts in release, where the first read indexed past the buffer. Now
  checked arithmetic throughout.
• Batched matmul panicked on a zero-size output dimension (Batched matmul panics with "chunk_size must not be zero" on a zero-size dimension #86) while
  the rank-2 path returned the empty tensor.
• Optimizer state never grew, so adding a parameter through the public
  groups_mut() panicked on the next step (Optimizer state is keyed by position in ParamGroup::params, so mutating a group through groups_mut() panics #89).
• serialize::save panicked inside safetensors on duplicate parameter
  names (serialize::save panics inside safetensors when a Module reports duplicate parameter names #91).

Data integrity and packaging

• serialize::load was not atomic (serialize::load is not atomic: a failed load leaves the module half-overwritten #90): a load that failed partway left
  the module holding checkpoint values for the parameters already visited and
  original values for the rest — a model from neither source, behind a typed
  error.
• No published crate carried a licence text (No licence file is included in any published 0.5.0 crate tarball #92). LICENSE-* lived only
  at the workspace root, so all twelve 0.5.0 tarballs shipped the licence
  claim without the licence. Both files are now in every crate and the
  release workflow refuses to publish one that omits them.
• Documentation described a release that no longer exists (Docs still describe the pre-main ctor removed in 0.5.0; SECURITY.md says nothing is released; README names SGD/RMSprop #93): the
  ctor pre-main constructor 0.5.0 deleted, 0.4.0's "seven crates, 47 → 40"
  (measured: three crates, 43 → 40), SECURITY.md's "Nothing is released
  yet", and a README optim row naming SGD/RMSprop instead of Sgd/RmsProp.

Closes #84
Closes #85
Closes #86
Closes #87
Closes #88
Closes #89
Closes #90
Closes #91
Closes #92
Closes #93

Not in this release

The deep-test run hit the org's monthly agent budget partway through, so the
autograd dimension never ran and 46 reported findings were left unverified
rather than disproven. Promising leads still to check: eigh's absolute
convergence threshold, its f32-computed symmetry guard on f64 input, whether
sum/mean are the compensated summation the CHANGELOG claims, and
serialize::load moving parameters back to the CPU after Module::to_device.