[0xAndoroid]

What changed

• Build the G1/G2 prepared-point vectors concurrently with rayon::join; each vector uses par_iter.
• Do all preparation outside the global cache write lock, then double-check before replacing the cache.
• Keep the existing init_cache API and first-initializer-wins semantics.

PR #27 is already on main, but it parallelizes DoryRoutines vector operations in ark_group.rs; cache preparation in ark_cache.rs remains serial. This replaces closed PR #26 on top of v0.4.1.

Benchmarks

Apple M4, 10 cores, log_n = 34 (131,072 G1 + 131,072 G2 generators), release build, median of 3 runs:

	main	this PR	change
init_cache	6.451 s	0.529 s	−91.8%, 12.2×

Setup generation was excluded from timing.

Verification

• cargo nextest run --cargo-quiet --features backends,cache: 85/85
• cargo nextest run --cargo-quiet --features backends,cache,zk: 142/142
• cargo clippy --all-targets --features backends,cache,zk -q --message-format=short -- -D warnings
• cargo fmt -q --message-format=short -- --check

Companion to a16z/jolt#1737.

[markosg04]

Looks good to me—thanks!