test: add criterion benchmark harness (#25)

[joeldsouzax]

Closes #25 — Phase 0 (P0.1) instrument.

Why

Performance is the project's paramount goal, but there was no way to measure it — no benches/, no bench dependency. Every later perf card (P1.x) depends on being able to answer "is this faster?". This adds a criterion-based harness over the realistic hot paths.

What

• criterion 0.8.2 dev-dependency, default-features = false (drops rayon/plotters) to keep the cargo deny/cargo audit surface minimal. Three [[bench]] targets gated required-features = ["stream"].
• benches/matter.rs — encode (matter_to_qb64, fixed codes) + decode (from_qualified_base64, fixed and variable-size StrB64_L0) + qb64↔qb2 conversion.
• benches/counter.rs — counter encode (v1 + auto-big) + counter-led group parse (parse_group).
• benches/stream.rs — full multi-primitive attachment-stream parse via groups().
• README "Benchmarks" section documenting how to run them.

Design notes

• Benches are std-only and feature-gated, so the no_std/WASM builds are untouched — cesr-wasm and cesr-nostd checks stay green.
• The flake runs clippy --all-targets --all-features, so bench code obeys the crate's full deny-set (no unwrap/expect/panic/as): operations are benchmarked by black_box-ing the returned Result, and pre-built fixtures are guarded with if let Ok(..).
• Fixtures are real qb64 vectors from src/core/matter/test_vectors.rs, so they parse without contrivance.

Research (per the strategy card's standing rule)

Tool choice (criterion vs divan/iai-callgrind/tango) and the CI-gate analysis (CodSpeed deferred) are captured with primary-source citations in the #25 research comment. The deterministic CI perf-gate (CodSpeed) is a deliberate follow-up.

Verification

nix flake check green (all 16 checks: clippy, fmt, taplo, audit, deny, nextest, doctest, wasm, nostd, + hygiene). All three suites run locally with tight confidence intervals (Matter decode ~60 ns / encode ~85 ns, counter parse ~26 ns, stream multi-group parse ~74 ns).

Note: second commit

This branch also carries one unrelated CI fix — ci: skip statix's broken test suite in dev shell — because statix 0.5.8 in the current nixpkgs rev fails its own tests and breaks nix develop/direnv. Happy to split it into its own PR if preferred.

🤖 Generated with Claude Code