v0.18.0

Run GPT-2 on host buffers and harden GPU and list lowering (#253)

## TL;DR
Run GPT-2 generation and one SGD training step on Osprey’s host GPU
backend, with typed tensor loading and regression coverage in both
syntax flavors.

## Details
- Add row-major `gpuMatMul`, raw and safetensors loaders, and `sqrt`,
`exp`, `ln`, and `tanh` kernels; report shape and file errors through
`Result`.
- Add the shared GPT-2 model, BPE tokenizer, cached generation, training
example, and independent Python reference.
- Harden scalar GPU lowering, ARC cleanup, list fold seeds, string-list
bindings and `join`; share the graphics ABI across native hosts.
- Run release credential preflight in the protected release environment,
and enforce this in the required release guard so environment-scoped
Azure credentials remain visible.
- Require at least 90% line coverage for the new GPU host runtime
(96.03% measured locally).
- Extend the transformer corpus in both flavors, retain exact LLVM
equivalence, and update its 15-case golden. Scope homepage syntax
controls and their browser checks to the greeting example.
- Synchronize specs, delivery plans and website examples. Update CI
preparation guidance to permit correcting demonstrably incorrect tests
while preserving assertions and gates.

## How Do The Automated Tests Prove It Works?
- `ml_and_default_twins_emit_identical_ir` requires byte-identical LLVM
for every paired corpus/benchmark program.
- `both_flavors_read_the_safetensors_fixture_exactly` checks loaded
F32/F16 values, named errors, raw loading and matrix products against
exact output.
- The transformer suite checks reference LayerNorm, GELU, softmax,
attention, embeddings and readout; its cached step must exactly match
the corresponding full-pass row. The differential harness compares all
goldens across native allocators, both kernel lowerings, WebAssembly and
Android.
- `join_accepts_every_string_list_representation` checks literals,
runtime lists and bound word lists without crashes.
- The website’s 139 browser tests exercise runnable examples,
navigation, responsive layouts and diagrams. `xvfb-run -a make ci`
passed, including all Rust/C/extension coverage floors, 219 native
goldens per allocator, 17 bank browser cases, installed-VSIX checks and
HTML documentation acceptance. Additional checks passed: 153 WebAssembly
goldens, 168 Android goldens plus app tests/lint, book
structure/examples, benchmark publication contracts and the Docker
compiler API. Hosted CI verifies Windows and iOS.

---------

Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>