Reuse cached batch provers and pin qp-zk-circuits to v4.4.0

[illuzen]

Summary

• Pin wormhole/zk crates to the published qp-zk-circuits v4.4.0 tag so the CLI can use the reusable PrivateBatchProver / PublicBatchProver API. crates.io still only has 4.3.0, so this is a git-tag pin rather than a crates.io bump.
• Cache those provers process-wide (OnceLock) so multiround and prepare pay the public-batch circuit build once (~tens of seconds) instead of once per batch.
• Redistribute multiround output assignments toward the partition minimum so a dust/zero exit no longer leaves the next-round address unfunded.

Test plan

• ./clippy.sh (already run locally)
• cargo test --lib compute_random_output_assignments (includes the 200-round multiround starvation regression)
• quantus wormhole multiround --num-proofs 7 --rounds 2 --public ... against a local node: first public batch still pays circuit build, later batches should be ~20s, and step 5 should capture all 7 next-round transfers
• After crates.io publishes 4.4.0, switch the git-tag pins back to version requirements

Made with Cursor

[illuzen]

@n13 please re-review — lockfile is now a surgical pin of the qp-zk-circuits crates to v4.4.0 only (no full cargo update), so the cooldown gate should stay quiet.

[n13]

Reviewer model: GPT 5.6 Sol

REQUEST_CHANGES: the output redistribution still permits the multiround starvation it is intended to prevent, and the git-only dependency declarations make this published crate unpackagable.

1. [P1] Do not submit a round while target deficits remain — src/cli/wormhole.rs:625-650

   This loop returns Ok(assignments) after break even when deficit > 0. That is reachable with inputs admitted by this flow: for five distinct targets, quantized inputs [3, 3, 3, 3, 100] at 4 bps yield per-proof outputs [2, 2, 2, 2, 99]. Passes 1-2 pair one proof with each target; only the 99-unit proof has excess, and after its sole output_2 slot tops up one target, no eligible donor remains. On this exact head, 20/20 targeted runs left three targets at 2 units. The initial partition explicitly permits 3-unit inputs, and subsequent fee rounding takes 2 to 1 and then 0, so the multiround command can still commit a round on-chain and later fail when a next-round mint disappears. Ensure every next-round target meets a pre-fee-safe minimum, or fail/reduce the proof set before submission when the two-output capacity cannot satisfy that invariant, and cover this skewed case deterministically.

2. [P2] Retain package versions alongside the temporary git pins — Cargo.toml:124-141

   All qp dependencies now specify only git/tag. cargo package --locked --no-verify --allow-dirty fails immediately with all dependencies must have a version requirement specified when packaging (first reported for qp-wormhole-aggregator). This repository's release workflow publishes to crates.io, so the manifest is no longer releasable even after qp 4.4.0 appears there. Add version = "4.4.0" alongside each git/tag pin (including the build dependency); Cargo can build from git now and retain the registry requirement when packaging.

Validation on 1075b4e69cdb2ff6ee8a8ad2665007f08dfe09f1:

• cargo +nightly-2026-08-31 fmt --all -- --check — passed
• taplo format --check --config taplo.toml — passed
• cargo test --locked with full circuit generation — passed (334 lib + 332 bin; 1 doc test ignored)
• SKIP_CIRCUIT_BUILD=1 cargo clippy --all-targets --locked -- -D warnings — passed
• Targeted starvation harness — reproduced in 20/20 runs
• cargo package --locked --no-verify --allow-dirty — failed as described above

GitHub's Fast Checks job failed during runner setup on an external Google apt index hash mismatch, before repository checks ran.

[illuzen]

@n13 both review points addressed:

1. P1 — compute_random_output_assignments now errors when every next-round address could have been funded (enough slots and pot) but some would still mint below the 3-unit floor. Deterministic coverage for the [3,3,3,3,100] case in compute_random_output_assignments_skewed_dust_proofs_fail_closed.
2. P2 — each qp crate pin is now version = "4.4.0" plus the v4.4.0 git tag (including the build-dep).

Please re-review.

[n13]

Reviewer model: GPT 5.6 Sol

APPROVE: The prior blockers are resolved, and I found no remaining blocking issues on 7f10e799b8f0052cdaa902de6b0c50c358c9adbb.

The skewed [3, 3, 3, 3, 100] path now fails before submission when limited output slots cannot fund every next-round address, with deterministic regression coverage. All qp dependencies retain version = "4.4.0" beside the temporary v4.4.0 git pins; the packaging preflight now gets past those entries and stops only at the pre-existing quantus_ur git dependency from the base branch. I also checked the pinned v4.4.0 PublicBatchProver::prove_batch path: it verifies supplied inner proofs and enforces block/asset/fee compatibility before proving.

Validation:

• cargo +nightly-2026-08-31 fmt --all -- --check — passed
• taplo format --check --config taplo.toml — passed
• SKIP_CIRCUIT_BUILD=1 cargo test --lib compute_random_output_assignments --locked — passed (6/6)
• SKIP_CIRCUIT_BUILD=1 cargo clippy --all-targets --locked -- -D warnings — passed
• Current GitHub checks — all passed, including Ubuntu/macOS build-and-test, examples, analysis, security audit, formatting, and dependency cooldown

No live-node multiround was run during this re-review; the focused regression tests and complete CI matrix cover the changed code sufficiently for approval.