fix(bench): make sure benches do not fail by fixing determinism

[gilcu3]

Closes #3516

[gilcu3]

these are broken window fixes. Not strictly needed, but better to have them in case a participant does not finish

[claude]

PR title type suggestion: This PR only changes benchmark files (test infrastructure), so the type prefix should probably be test: instead of fix:.
Suggested title: test(bench): make sure benches do not fail by fixing determinism

[claude]

Pull request overview

Fixes the simulated-replay benches in crates/threshold-signatures/benches/ so they fail loudly when the replay diverges, instead of silently timing the path to an error. The previous code passed the Result of run_simulated_protocol/run_protocol straight to criterion which dropped errors, and rebuilt the real participant's RNG via fragile order-dependent reconstructions (MockCryptoRng::seed_from_u64(42) + N next_u64() calls). The new approach records each per-participant seed in a HashMap<Participant, u64> during snapshot capture and rebuilds the RNG via a small participant_rng helper. Resolves #3516.

Changes:

• Add .expect("...") on every run_simulated_protocol/run_protocol call inside iter_batched closures across the advanced_* and naive_* benches.
• Introduce participant_rng(&seeds, participant) and a seeds: HashMap<Participant, u64> field on the various Prepared* structs (PreparedPresig, PreparedDkgPackage, PreparedCkdPackage, FrostEd25519SigV1, OTECDSAPreparedTriples, RobustECDSAPreparedPresig).
• Replace handwritten RNG-replay loops in advanced_ot_based_ecdsa.rs, advanced_robust_ecdsa.rs, and advanced_eddsa_frost_sign_v2.rs with participant_rng(&preps.seeds, real_participant).
• Add assert_eq!(results.len(), participants.len()) after run_simulation in simulate_ckd.rs, simulate_ecdsa.rs, and simulate_frost.rs to catch participants that silently dropped out.
• Convert prepare_dkg's return from Vec<(Participant, ...)> to a struct PreparedDkgPackage { protocols, seeds }; simulate_dkg.rs and advanced_dkg.rs updated to destructure.

Reviewed changes

Per-file summary

File	Description
benches/bench_utils.rs	Adds participant_rng helper and seeds field on PreparedPresig.
benches/bench_utils/ckd.rs	Records per-participant seed in PreparedCkdPackage.seeds.
benches/bench_utils/dkg.rs	Returns new PreparedDkgPackage carrying protocols + seeds.
benches/bench_utils/frost_eddsa.rs	ed25519_build_presign_protocols now returns (protocols, seeds); ed25519_prepare_sign_v1 exposes seeds.
benches/bench_utils/ot_based_ecdsa.rs	Records seeds for ot_ecdsa_prepare_triples; ot_ecdsa_prepare_presign returns an empty seeds map (no per-participant RNG).
benches/bench_utils/robust_ecdsa.rs	Records seeds for robust_ecdsa_prepare_presign.
benches/advanced_dkg.rs	Destructures PreparedDkgPackage; uses participant_rng; .expect on replay.
benches/advanced_eddsa_frost_sign_v1.rs	Replaces RNG passthrough with participant_rng; .expect on replay.
benches/advanced_eddsa_frost_sign_v2.rs	Removes handwritten replay loop; uses participant_rng; .expect on replay.
benches/advanced_ot_based_ecdsa.rs	Same refactor for triples setup; .expect on replay.
benches/advanced_robust_ecdsa.rs	Same refactor for presign setup; .expect on replay.
benches/ckd.rs	.expect on replay; uses participant_rng for coordinator RNG.
benches/naive_ot_based_ecdsa.rs / naive_robust_ecdsa.rs	.expect("protocol should complete") on every run_protocol call.
benches/simulate_ckd.rs / simulate_ecdsa.rs / simulate_frost.rs	Add assert_eq!(results.len(), participants.len()) after run_simulation.
benches/simulate_dkg.rs	Updated to use .protocols on new PreparedDkgPackage.

Findings

Non-blocking (nits, follow-ups, suggestions):

• crates/threshold-signatures/benches/bench_utils.rs:79-82 — The doc comment on PreparedPresig::seeds ("empty when the protocol is built from deterministic inputs") describes a quirk of one specific constructor (ot_ecdsa_prepare_presign, which passes HashMap::new() at bench_utils/ot_based_ecdsa.rs:96). Because the field is the same type in either case, a future caller that mistakenly calls participant_rng(&preps.seeds, ...) on an OT-based presign would hit the expect in participant_rng only at runtime. Consider modeling this with two distinct types (or an Option<HashMap<...>>) so the absence is encoded in the type. Acceptable as a follow-up given this is bench-only code.
• crates/threshold-signatures/benches/bench_utils.rs:40-48 — participant_rng is generic over BuildHasher, but every caller passes HashMap<Participant, u64> with the default hasher. The generic parameter adds noise without buying anything; could drop S and accept &HashMap<Participant, u64> directly.
• crates/threshold-signatures/benches/advanced_dkg.rs:106, advanced_eddsa_frost_sign_v1.rs:77, advanced_eddsa_frost_sign_v2.rs:104, advanced_ot_based_ecdsa.rs:161, advanced_robust_ecdsa.rs:123, ckd.rs:84 — The "rebuild the exact rng the real participant used during snapshot capture" comment is repeated verbatim across 6 sites and largely paraphrases the next line (participant_rng(&preps.seeds, ...)). The why — that the simulator's recorded messages only match if the real participant replays its exact internal randomness — would be more valuable in the participant_rng doc comment alone, and the call-site comments could be dropped. (Per engineering-standards.md "Write helpful code comments".)
• crates/threshold-signatures/benches/simulate_ecdsa.rs:381, simulate_ckd.rs:113, simulate_ckd.rs:151, simulate_frost.rs:161, simulate_frost.rs:193 — As the author already noted on the inline thread, the assert_eq!(results.len(), participants.len()) additions catch the case where a participant drops out entirely. Note that assert!(results.iter().any(|(_, sig)| sig.is_some())) remains too weak (only one valid output is required), so participant failures past assert_eq! can still slip through as None outputs — worth a follow-up to assert quorum size, not just non-emptiness. Not new to this PR.

Approved