feat(rust-sdk): ergonomic cargo options, nextest, ci(), cargo-hack (Py+TS parity)

[markovejnovic]

What

Overhauls the hm.rust pipeline-authoring helper in both DSLs so a Rust CI pipeline is one call:

proj = hm.rust.project()
hm.pipeline(list(proj.ci(nextest=True)))   # nextest + doctest + clippy + fmt

Every action method now takes research-backed cargo options

packages=, exclude=, features=, all_features=, no_default_features=, target=, release=/profile=, locked= (defaults True for reproducible CI), workspace=, plus a generic flags= escape hatch. The surface is symmetric across build/test/clippy/fmt/doc on both RustToolchain and RustProject.

New capabilities

• test(nextest=True) → cargo nextest run; companion doctest() step (nextest can't run doctests).
• RustProject.build() and .doc() (previously missing).
• RustProject.ci(nextest=, doc=) → the zero-config DAG (test + clippy + fmt, +doctest under nextest).
• feature_powerset() → cargo-hack --feature-powerset --depth 2 --no-dev-deps (the tokio/reqwest/tracing idiom), with a path-independent, forever-cached install step.

Hardening

• Shell-injection fix: every user-supplied value (packages, exclude, features, target, profile, cargo-hack skip/include) is shell-quoted (shlex.quote / shQuote). Raw flags= stay verbatim (escape hatch). Regression tests added.
• Conflicting options raise precise errors per the repo's error-message doctrine: all_features + features/no_default_features; release + profile; exclude without workspace (cargo requires it); exclude with explicit packages.

Parity

Both DSLs route through a shared cargo-argument assembler (_cargo.py / cargo.ts) and shQuote replicates Python's shlex.quote byte-for-byte. Golden-string parity tests pin identical cargo commands across Python and TypeScript; the regenerated rust-release e2e fixtures are byte-identical across languages and pass the Rust IR-deserialization tests.

Why

Flag choices are evidence-based: surveyed CI of ~30 top Rust repos (tokio, ripgrep, rust-analyzer, ruff, axum, reqwest, rustls, sqlx, wgpu, …). --locked is the community standard for reproducible CI; nextest is dominant in large workspaces and always splits doctests into a separate step; cargo clippy --all-targets -- -D warnings and cargo fmt --all --check are the canonical lint/fmt invocations; cargo-hack --feature-powerset --depth 2 is the standard combinatorial-feature sweep.

Backwards compatibility

Additive except: RustToolchain.build/test now emit --locked by default (pass locked=False to opt out); RustProject.clippy uses --all-targets instead of --tests; fmt defaults to --all --check. Examples and e2e fixtures updated accordingly.

Tests

• Python: 543 passed, ruff + mypy clean.
• TypeScript: 405 passed, tsc + build clean.
• Rust: 14 e2e fixture tests pass (deserialization + cross-language parity).
• New: per-method unit tests on both classes/langs, shell-injection regression tests, and 6 cross-language golden-parity assertions.