perf(frequency): Back Frequencies with hashbrown 0.17, use entry_ref

[jqnatividad]

Summary

Swaps the Frequencies backing map from std HashMap (via foldhash::HashMap) to hashbrown::HashMap, and switches the byte-slice insert hot path to hashbrown's entry_ref.

The enabling detail from hashbrown's changelog: hashbrown 0.16+ uses foldhash 0.2 as its default hasher — the exact hash qsv-stats already used. So hashing is unchanged; only the table implementation differs.

Why it's a dependency swap, not an addition

foldhash was used in exactly one place (frequency.rs, the crate's only HashMap user). After the swap:

• Direct production deps stay at 4: foldhash → hashbrown (num-traits, rayon, serde unchanged).
• foldhash 0.2 remains in the tree transitively under hashbrown — a single copy, no duplication.
• Net-new transitive crates: allocator-api2, equivalent (both tiny).

The hot-path change

add_borrowed / increment_by_borrowed previously did get_mut(&[u8]) then insert(v.to_vec()) on a miss. std's HashMap has no stable raw-entry API, so a new key was hashed + probed twice. hashbrown's entry_ref probes once with the borrowed key and only allocates ([u8]::to_owned()) on the vacant branch. Result is identical (verified by the unchanged test suite + a cardinality assertion in the bench).

Benchmarks (1M rows, identical hashing — benches/freqhb.rs)

build (add_borrowed):

shape	std	hashbrown	delta
low_card_50	3.82 ms	3.05 ms	−20%
low_card_5000	6.12 ms	5.28 ms	−14%
high_card_short	54.3 ms	50.0 ms	−8%
high_card_long	93.3 ms	87.0 ms	−7%

merge (Commute::merge pattern):

shape	std	hashbrown	delta
low_card_5000	1.18 ms	1.00 ms	−15%
high_card_short	40.2 ms	38.8 ms	−3.5% (noisy)

The high-cardinality build is the dominant wall-clock cost in qsv's frequency command, and it improves ~7–8%.

foldhash is kept as a dev-dependency solely to retain the std baseline in the comparison bench.

Test plan

• cargo test --lib — 161 passed, 0 failed
• cargo clippy --lib — clean
• cargo bench --bench freqhb

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR optimizes the crate’s frequency-counting hot path by switching Frequencies from a foldhash-backed std HashMap to hashbrown::HashMap, and by using hashbrown’s borrowed-key entry_ref API to avoid double hashing/probing and unnecessary allocations when counting &[u8] keys.

Changes:

• Replace foldhash::HashMap usage in Frequencies<T> with hashbrown::HashMap.
• Update the byte-slice counting hot path (add_borrowed / increment_by_borrowed) to use entry_ref.
• Add a new Criterion benchmark (benches/freqhb.rs) and wire it up in Cargo.toml; update CLAUDE.md documentation accordingly.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 2 comments.

File	Description
src/frequency.rs	Switch backing map to hashbrown::HashMap and use entry_ref for borrowed byte-slice insertion/counting.
CLAUDE.md	Update documentation to reflect the new hashbrown backing and entry_ref usage.
Cargo.toml	Swap prod dependency foldhash → hashbrown, keep foldhash as a dev-dependency for benchmarking, and add the freqhb bench target.
benches/freqhb.rs	New benchmark comparing std+foldhash vs hashbrown for build/merge paths.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.