Wave2 polish: clippy/fmt clean, CI, finish dead diagnostic

[Peterc3-dev]

Summary

Polish-to-completion pass on the kv-compressor crate. No changes to the
compression or attention logic — this is lint, test, and CI hygiene plus
finishing one dead diagnostic helper.

Changes

• Clippy clean (-D warnings) across lib + both binaries:
  • usize::div_ceil / usize::is_multiple_of replacing manual ceil-div and
    % == 0 checks (src/lib.rs, src/bin/realdata-bench.rs)
  • regrouped two unreadable hex RNG-seed literals in tests
  • module-level #![allow(clippy::needless_range_loop)] on the four numeric
    kernels (attention, outlier, grouped, twobit), with a comment
    explaining the rationale: the loop counter indexes packed sign-word slices
    together with magnitude/score/weight arrays in lockstep, and the
    i-outer/d-inner bit-walk; iterator rewrites would obscure the bit-packing
    layout. This is the honest call rather than mangling hot math paths.
• Finished a dead stub: removed _ensure_v2_view_compiles (a dead
  #[allow(dead_code)] helper sitting after the test module that only
  existed to force GroupedKv::flatten_to_v2 to compile) and replaced it with
  a real unit test, flatten_to_v2_averages_group_mags_and_preserves_outliers.
  Dropped the now-unused CompressedKv import in grouped.rs.
• Formatting: ran cargo fmt repo-wide (the repo was not fmt-clean, which
  is why the diff is large — most of it is pure reflow).
• CI: added .github/workflows/ci.yml (stable toolchain) running fmt
  check, clippy --all-targets -D warnings, release build, and tests.

Verification (run locally on this machine — Rust 1.95.0 stable, x86_64-linux)

• cargo build --release --all-targets — passes
• cargo test --all-targets — passes, 20 tests (was 19; +1 new test)
• cargo clippy --all-targets -- -D warnings — clean
• cargo fmt --check — clean
• cargo run --release --bin kv-bench — runs, reports 6.50x ratio on synthetic
  data as expected

The CI workflow itself has not been run on GitHub Actions yet
(UNVERIFIED until it executes on the PR); it mirrors the exact commands
verified locally above.

Not done / left alone

• No new benchmark numbers or features were invented; this is a documented
  negative-result research crate and stays that way.
• realdata-bench still requires externally dumped Qwen3 KV .npz tensors
  (per the Python scripts) to exercise its full path — not run here, no such
  dump on this machine. The synthetic kv-bench path was exercised.
• criterion remains a dev-dependency with no [[bench]] target wired up
  (pre-existing); left as-is to avoid scope creep.

🤖 Generated with Claude Code

[Peterc3-dev]

Independent verification — VERDICT: solid, ready to merge

Re-cloned wave2-polish and ran the full pipeline locally (Rust 1.95.0 stable, x86_64-linux). Every claim in the PR body checks out:

Check	Claim	Result
cargo build --release --all-targets	passes	PASS (clean)
cargo test --all-targets	20 tests pass (+1 new)	PASS — 20 passed, 0 failed
cargo clippy --all-targets -- -D warnings	clean	PASS — no warnings/errors
cargo fmt --check	clean	PASS
cargo run --release --bin kv-bench	6.50x ratio on synthetic	PASS — exactly 6.50x (1.2 MB vs 8.1 MB q8)

Diff review (452+ / 140-, 9 files)

• No behavior changes to compression or attention logic — confirmed by reading every hunk in attention.rs, outlier.rs, twobit.rs, grouped.rs. The bulk is pure cargo fmt reflow.
• The only semantic substitutions are the documented (d+63)/64 → d.div_ceil(64) and d % h == 0 → d.is_multiple_of(h) — both are exact equivalents for the non-negative usizes involved. Verified call sites in lib.rs::sign_words_for, realdata-bench, and the ratio math.
• Dead-stub replacement is legit: _ensure_v2_view_compiles (which fabricated an empty CompressedKv) is gone; the new flatten_to_v2_averages_group_mags_and_preserves_outliers test exercises the real path. Its arithmetic is correct — group mags 0.15 and 13.25 average to 6.7, and flatten_to_v2 does exactly average the per-group mags and clone signs/outliers through unchanged.
• The #![allow(clippy::needless_range_loop)] on the four numeric kernels is the honest call given the lockstep sign-word/magnitude bit-walk indexing — better than mangling hot paths.

No overclaims

PR body is honest, including the appropriate caveats: GitHub Actions CI not yet executed (correctly flagged UNVERIFIED), realdata-bench not run (needs external Qwen3 .npz dump), and criterion left unwired to avoid scope creep. The CI workflow mirrors the exact commands verified locally.

No required fixes. Approve / merge.