Shrink crates.io footprint: build precompute_hash/sample tables lazily

[GordonYuanyc]

Summary

Replaces the ~131K-line checked-in precompute tables (PRECOMPUTED_HASH, PRECOMPUTED_SAMPLE, PRECOMPUTED_SAMPLE_RATE_1PERCENT) with std::sync::LazyLock-backed tables that materialise once on first access.

Net diff vs main: +115 / -147,611 lines.

Why not just generate them at build time?

Two clean library-idiomatic options exist for "table that is really just a deterministic function of i":

1. Build-time codegen into OUT_DIR — the approach the previous commit on this branch took for the sample tables. It works, but for PRECOMPUTED_HASH specifically the generator would need to either duplicate the byte-conversion + XxHash3_128 logic in build.rs or pull twox-hash in as a build dependency (drift risk; the contract is literally "what hash128_seeded returns for these inputs"). It also writes 64K+ lines of generated Rust under target/ on every clean build.
2. LazyLock — describe the function, materialise the table on first access. No build script involvement, no extra build deps, and for the hash table the values are computed by the crate's own hash128_seeded, so by construction they can never drift.

This PR picks (2) for all three tables.

Per-table summary

• precompute_hash.rs: 16 390 lines → 31 lines. pub static PRECOMPUTED_HASH: LazyLock<Box<[u128]>> computed from hash128_seeded(0, &DataInput::U64(i)).
• precompute_sample.rs / precompute_sample2.rs: LazyLock<Box<[f64]>> from a SmallRng seeded with the same 0xA5A0_5A71_B11B_C0DE the previous commit on this branch settled on, so values remain reproducible across runs/versions. The 1%-rate table reuses a private build_ln_one_minus_u_table(scale) helper.
• build.rs: stripped of all codegen; only compiles .proto files now.
• Cargo.toml: removed rand from [build-dependencies], the three [[bin]] entries for generate_precomputed_*, and the now-unused internal-bins feature.
• src/bin/: deleted (only ever held the three generator bins; directory removed entirely).
• docs/library_map.md: updated to point at the new LazyLock design instead of the removed bins.

API impact

Names are unchanged. Indexed access (PRECOMPUTED_HASH[i]), iteration (.iter()), and length (.len()) continue to work via Deref to the underlying slice — the existing call sites in src/sketch_framework/nitro.rs and src/common/structure_utils.rs are unmodified.

The element container changes from [T; N] to Box<[T]>:

• PRECOMPUTED_HASH[i] -> u128: unchanged.
• &PRECOMPUTED_HASH as &[u128; 0x4000]: no longer compiles; use &PRECOMPUTED_HASH[..] or &[u128].
• New PRECOMPUTED_HASH_LEN (= 0x4000) and PRECOMPUTED_SAMPLE_LEN (= 0x10000) constants are exposed for compile-time length needs.

Runtime cost

One-time initialisation on first access:

• hash table: ~16K XxHash3_128 evaluations + one 256 KiB heap alloc
• each sample table: ~64K SmallRng draws + ln + one 512 KiB heap alloc

Steady state: indexed access adds one relaxed atomic load (well-cached) versus the previous static array.

Test plan

• cargo build succeeds
• cargo test --lib — 394/394 pass, including the nitro_batch_* tests that hammer PRECOMPUTED_SAMPLE_RATE_1PERCENT in the hot path
• cargo clippy --lib --tests --no-deps clean
• Reviewer to sanity-check the Box<[T]> API change is acceptable for downstream callers (the names and indexed-access pattern are preserved)

Made with Cursor