wave2-polish: tests, CI, and formatting

[Peterc3-dev]

Polish-to-completion pass on inference-bench.

Summary

• Formatting: ran cargo fmt across all source files. No logic changes, purely whitespace/line-wrapping normalization.
• Tests (new): added 18 unit tests for the pure, network-independent core logic:
  • Endpoint::generate_url per endpoint kind (Ollama / FastFlowLM / llama.cpp)
  • Endpoint::build_body request-shape correctness for both API families
  • parse_ollama_stream_line — token, done flag, invalid-JSON -> None
  • parse_llamacpp_stream_line — SSE token, [DONE] marker, stop flag, non-data: lines -> None
  • parse_ollama_timings / parse_llamacpp_timings — field extraction, ms->ns conversion, invalid/missing input falls back to defaults
  • metrics::mean_std — empty, single-sample, identical-values, and a known-value case checked against the textbook sample std (n-1 denominator)
• CI (new): .github/workflows/ci.yml — runs cargo fmt --check, cargo clippy --all-targets -D warnings, cargo build, and cargo test on push to master and on every PR. Uses dtolnay/rust-toolchain@stable + cargo cache.
• Docs: added a Development section to the README documenting the test/lint/fmt commands and CI.

What was finished

Nothing was a broken/incomplete stub — the existing code already built cleanly with no clippy warnings and no todo!()/unimplemented!()/FIXME markers. The gaps were the complete absence of tests and CI, plus unformatted code. Those are now filled.

Verified on this machine (cargo 1.95.0, rustc 1.95.0)

• cargo build and cargo build --release — PASS
• cargo test — PASS (18/18)
• cargo clippy --all-targets -- -D warnings — clean
• cargo fmt --check — clean

Not verified

• The GitHub Actions workflow itself has not been executed; it is validated only by inspection. It will run on this PR.
• Networked behavior (live benchmarking against Ollama/FLM/llama.cpp) was not exercised — no servers were started. Only the pure parsing/stat logic is unit-tested.

TODOs left

• None introduced. No ambiguous stubs were guessed at.

🤖 Generated with Claude Code

[Peterc3-dev]

Independent verification: VERDICT = solid

Re-cloned wave2-polish, reviewed the full diff vs master, and reproduced every claim on a clean machine (cargo/rustc 1.95.0).

Build & checks (independently reproduced):

• cargo build — PASS
• cargo build --release — PASS
• cargo test — PASS, 18/18 (matches claim; 14 in endpoints, 4 in metrics)
• cargo clippy --all-targets -- -D warnings — clean
• cargo fmt --all --check — clean

Diff review:

• output.rs, bench.rs, detect.rs, main.rs, metrics.rs, endpoints.rs non-test hunks are pure rustfmt churn (line-wrapping, import reorder, struct-literal expansion). No logic changes, no behavior changes. The cyan-green palette (0,255,200) is identical on both sides — already in master, not touched here.
• New tests are additive, correct, and assert the right invariants (e.g. llama.cpp body carries no model field; ms→ns conversion 10ms → 10_000_000ns; sample-std n-1 denominator verified against textbook value 2.13809).
• No regressions, no incorrect stub completions, no todo!()/unimplemented!() introduced. No security concerns.

No overclaims. The PR body's 'Not verified' section (CI not executed, networked benchmarking not exercised) is accurate and appropriately scoped.

Nit (non-blocking): PR prose writes the clippy invocation as cargo clippy --all-targets -D warnings; the actual workflow file correctly uses --all-targets -- -D warnings. Prose only — the YAML is right and CI is syntactically sound.

Ready to merge.