chore: standardize model feature flags and adopt cargo-nextest as test runner

[xiaguan]

Summary

Three commits that turn two implicit conventions into enforced, documented standards: how model lines are feature-gated, and what the test sweep contract is.

1. fix(glm52): restore cfg(test) imports of Glm52ProjBytes

#550 marked from_host/upload test-only but dropped the imports they need — cargo test -p openinfer-glm52 --lib has been broken on main since.

2. chore(features): one feature name per model line, no crate-internal gates

New standard in docs/conventions/feature-flags.md. Model selection lives in exactly two places: openinfer-server (what the binary serves) and openinfer-kernels (what nvcc compiles). Two crate tiers:

• Tier A (pure-CUDA kernels) — qwen3, glm52, kimi-k2, deepseek-v2-lite (migrated here): no self-named feature, the crate always compiles whole, openinfer-kernels = { features = ["<model>"] } non-optional. Removes 22 scattered #[cfg] gates, reject-only stubs, and dsv2-lite's optional kernels dep (the only one in the workspace).
• Tier B (build-time Python toolchains) — qwen35-4b (Triton), deepseek-v4 (TileLang + CuTe DSL): the only allowed gate is a single whole-crate #![cfg]. deepseek-v4's 10 scattered item gates collapse into one, matching qwen35-4b.

Consumers of the removed crate-level features updated: pre-commit kimi clippy hook, scripts/bench_dsv2lite_vllm_matrix.py (4 call sites), living gate docs. Server-level --features kimi-k2 / --features deepseek-v2-lite invocations are unchanged.

Known cost, priced in the convention doc: workspace-wide builds now compile kimi/dsv2-lite kernels (cached after first build), and a bare cargo test --workspace (no --lib) builds their GPU gates — same behavior qwen3/glm52 already had.

3. chore(tests): adopt cargo-nextest as the test runner

The sweep contract moves from convention to config (.config/nextest.toml):

• Bare cargo nextest run --workspace = lib unit tests, green on any single-GPU box; hardware gates are excluded by default-filter and reported as skipped — selection is explicit, never silent. Gates opt in per package with --ignore-default-filter.
• retries = 0 (flaky = bug; exact-match gates must not rot under retry).
• slow-timeout + terminate-after: a wedged NCCL collective goes red instead of hanging the session.
• gpu-gate (serial) / gpu-lib (max 4) groups: nextest is process-per-test, engines and CUDA contexts must not stack.

nextest also compiles every test target where --lib compiles none, which immediately surfaced two rotted files (fixed here): kv-offload/tests/cpu_roundtrip.rs still used pre-#522 QueryOutcome struct fields; kernels/tests/glm52_indexer_smoke.rs had an unused variable.

Requires nextest >= 0.9.138 (0.9.70 fails to list proc-macro test binaries).

Verification (RTX 5070 Ti, OPENINFER_NCCL_ROOT set)

• cargo test --release --workspace --lib: green (only pre-existing env failure: gdr_copy_flag_GPU needs GDRCopy hardware)
• cargo check --release -p openinfer-server × {default, kimi-k2, deepseek-v2-lite, glm52}; -p openinfer-kimi-k2 --features kernel-report --bins --tests; -p openinfer-deepseek-v2-lite --bins --tests
• Sweep: 725 passed / 12 skipped / ~10s, green (baseline cargo test --lib --no-fail-fast: 4.2s but red on gdr)
• Gates through documented nextest invocations: qwen3 hf_golden_gate PASS 46.7s; qwen35-4b e2e_scheduler (Tier B --features path) PASS 26.6s
• fmt, clippy (-D warnings local hooks), pre-commit chain all pass
• Not verifiable on this box: deepseek-v4 with its feature (no TileLang Python); its feature-on item set is byte-identical to before

🤖 Generated with Claude Code