[zero] perf(zebra-consensus): cache successful transparent script verification

[aphelionz]

Reintroduces transparent script verification reuse: a transaction verified at mempool admission is not re-verified in the block that mines it (zebra lost this in upstream #10494). This is the caching half of #34, redesigned per its review; the overlapping-UTXO-lookups half is #46.

• Key: the transaction's WtxId alone. The ZIP-244 txid commits to the consensus branch id and to the outpoints, which pin the immutable prevout scripts and values; the authorizing-data digest commits to the scriptSigs, so a CVE-2026-34377 same-txid twin misses. v1-v4 ids don't commit to the branch id and are never cached.
• Replacement: random victim slot via seeded SipHash-1-3 (siphasher, already in the dependency tree and vet-exempted); unguessable seed in production, fixed seed in tests. One deviation from the review's sketch: the victim is the hash of an insert counter, not of the incoming key, so the choice is uniform over the slots (a per-key victim would be direct-mapped, and two hot colliding keys would evict each other forever).
• The cache key is derived from the transaction being verified, never from the caller-supplied request hash, so no field on a public request type can decide whether scripts run.
• Cache check, script verification, and success-only insert are co-located in make_transparent_input_and_output_checks (renamed: it builds deferred checks, it doesn't verify).
• Transactions spending unmined mempool outputs bypass the cache entirely (zebra#10346 pairing history).

Benchmarks (cargo bench -p zebra-consensus --bench script; a 1001-input P2SH consolidation with a real ECDSA signature on every input, so a miss pays the interpreter, ZIP-244 sighash, and signature verification per input):

bench	time
the per-input script work a hit skips (serial)	165 ms
block-path verification, cache miss	18.0 ms
block-path verification, cache hit	2.1 ms

The miss/hit gap is smaller than the serial number because the verifier spreads input checks across cores; a 4-vCPU host sits correspondingly closer to the serial figure.

145/145 cargo nextest run -p zebra-consensus; fmt, clippy -D warnings, and cargo vet --locked clean.

🤖 Generated with Claude Code

[Copilot]

Pull request overview

Reintroduces a process-global cache in zebra-consensus so successful transparent script verification performed during mempool admission can be reused during subsequent block verification (v5+ only, keyed by WtxId), reducing repeated script-interpreter/FFI work.

Changes:

• Add a bounded, randomly-evicting transparent script verification cache (transaction/script_cache.rs) keyed by WtxId, and consult/insert it from the transparent script-check pipeline.
• Wire the cache into both block and mempool transaction verification paths, including bypass for spends of unmined mempool outputs.
• Add extensive tests and a Criterion benchmark to validate and measure cache hits/misses.

Reviewed changes

Copilot reviewed 6 out of 7 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
zebra/zebra-consensus/src/transaction.rs	Integrates script-cache keying and hit/miss behavior into transparent script verification checks for block/mempool verification.
zebra/zebra-consensus/src/transaction/script_cache.rs	New bounded, process-global cache for successful transparent script verifications with keyed random replacement + metrics/test hooks.
zebra/zebra-consensus/src/transaction/tests.rs	Adds cache-focused integration tests covering key soundness (twins, v4 exclusion, poisoning resistance, cross-upgrade behavior, mempool→block reuse).
zebra/zebra-consensus/Cargo.toml	Adds siphasher dependency and registers the new script benchmark target.
zebra/zebra-consensus/benches/script.rs	New benchmark measuring per-input script verification cost and end-to-end block verification miss vs hit.
zebra/Cargo.toml	Adds siphasher to workspace dependencies.
zebra/Cargo.lock	Locks siphasher into the dependency graph for zebra-consensus.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Copilot]

Pull request overview

Copilot reviewed 6 out of 7 changed files in this pull request and generated no new comments.

Suppressed comments (1)

Previously missed (1) — in code that hasn't changed since the last review.

zebra/zebra-consensus/src/transaction/script_cache.rs:168

• victim_index calls self.siphasher.hash(&key.as_bytes()), but SipHasher13 is typically used via the std::hash::Hasher API (stateful), not via a pure hash() helper. If this is using the Hasher implementation, reusing a single SipHasher13 instance would make victim selection depend on prior calls and risks thread-safety issues; if hash() is not available, this won’t compile. Prefer storing the SipHash key/seed and constructing a fresh SipHasher13 per victim_index call (or cloning/resetting per call) to get a deterministic PRF output for just key.

    /// The slot a full cache replaces when inserting `key`.
    fn victim_index(&self, key: &WtxId) -> usize {
        // Casts are lossless: `capacity` is a usize, and the modulus keeps
        // the result below it.
        (self.siphasher.hash(&key.as_bytes()) % self.capacity as u64) as usize
    }

[Copilot]

Pull request overview

Copilot reviewed 6 out of 7 changed files in this pull request and generated no new comments.

[Copilot]

Pull request overview

Copilot reviewed 6 out of 7 changed files in this pull request and generated 1 comment.

Suppressed comments (1)

zebra/zebra-consensus/src/transaction/tests.rs:4221

• The panic message is inverted: this branch runs when the mock input is not a PrevOut, but the message says it is.

    let (input, known_utxos) = uniquely_sourced(0xA4, input, known_utxos);
    let transparent::Input::PrevOut {
        outpoint, sequence, ..
    } = input
    else {
        panic!("mock input is a PrevOut");

[Copilot]

Pull request overview

Copilot reviewed 6 out of 7 changed files in this pull request and generated no new comments.