evm-fork-cache

A forked-EVM simulation engine for EVM search, MEV, and backtesting — built on revm, alloy, and foundry-fork-db.

It exists to answer one question fast and repeatedly: "if I sent this transaction against current on-chain state, what would happen?" — for thousands of candidate transactions per block, without paying an RPC round-trip or re-deriving state on every call.

Why it exists

A search loop evaluates many hypothetical transactions against the same recent chain state. Doing that with a naive fork means re-fetching state, paying RPC latency on the hot path, and either sharing mutable EVM state across tasks (unsafe) or deep-cloning a fork per candidate (slow). evm-fork-cache is built around three capabilities that target exactly this workload:

1. Cheap parallel fan-out — freeze state once into an immutable snapshot, hand a cheap Arc clone to each task, and run many isolated simulations in parallel. No task can observe another's writes.
2. Reactive, event-driven state sync — keep hot state correct from the chain's own logs with no RPC on the hot path: a default-enabled reactive runtime decodes events into targeted writes, invalidates/resyncs what it can't derive, and recovers from reorgs — driven out of the box by a live WebSocket subscriber (or your own transport), and seeded by protocol-neutral cold-start that warms a working set into the cache in one batched pass.
3. Freshness as a first-class concept — the engine tracks what it can trust, for how long, and verifies the rest. The optimistic verify-and-rerun loop hides RPC latency: act on speculative results immediately, get a Confirmed or Corrected verdict when the background validation lands.

Maturity. This crate is pre-1.0 and under active development against a phased roadmap. All three capabilities ship today: copy-on-write snapshots + overlays (1); a default-enabled reactive runtime with a live AlloySubscriber (WebSocket subscribe_logs/subscribe_blocks/subscribe_pending_transactions, exponential-backoff reconnect, and get_logs backfill) plus protocol-neutral cold-start (2); and the optimistic verify-and-rerun loop (3). Honest remaining transport work: full block bodies and full pending-transaction hydration are follow-ups; owner-scoped log backfill is available for newly added reactive interests. The public API still changes between minor versions — see Stability.

What it provides today

• Forked EVM cache backed by foundry-fork-db with lazy RPC loading and on-disk persistence for accounts, storage, bytecode, and immutable metadata.
• Snapshots and overlays — snapshot() produces an immutable, Send + Sync point-in-time view; each EvmOverlay is a cheap clone that simulates in isolation, ideal for parallel candidate evaluation.
• Bundle simulation — simulate_bundle applies an ordered sequence of transactions over cumulative block state (each transaction sees the previous one's writes), with an Atomic / AllowReverts(indices) revert policy and coinbase/miner-payment accounting (the beneficiary balance delta — priority fee plus direct tips; the base fee is burned in-EVM per EIP-1559). This is the shape a searcher evaluates a candidate set (victim + backrun, sandwich) with.
• Freshness control plane — a four-layer model (classification, observation, policy, mechanism) plus an optimistic verify-and-rerun execution loop with deferred validation. See the freshness module. Scope note: the validation loop reconciles storage slots observed in the simulation read set; native balance, nonce, and bytecode freshness remain caller-managed through event-driven writes or out-of-band reconciliation.
• Targeted state manipulation — direct storage injection, account/slot purge, and balance overrides for hot-state refresh workflows.
• Event-to-state pipeline — decode logs into StateUpdates, apply them in order, purge touched state on reorg, and reconcile sampled event-derived slots against RPC. The crate ships the generic driver, the ERC-20 Transfer decoder, and in-memory examples; protocol-specific decoders stay with the consumer or companion crates.
• Reactive runtime — register pure handlers for logs, block notifications, and pending transaction signals. Handlers emit StateUpdates, invalidations, resync requests, speculative signals, and hook signals; the runtime routes inputs, deduplicates and orders canonical logs, validates pending semantics, applies canonical cache mutations through EvmCache::apply_updates, and can optionally execute storage resync requests through the cache's provider-neutral storage batch fetcher before dispatching reports to hooks. Canonical block effects are journaled for depth-bounded reorg recovery: removed logs, explicit reorged inputs, and parent-hash discontinuities emit ReactiveReport::Reorg, roll back reversible storage writes, fall back to targeted purges for irreversible effects, and cancel stale hash-pinned resyncs. The ReactiveRegistry exposes consolidated Alloy log filters for provider subscription setup and exact local log routing with optional route keys. The registry and runtime support incremental handler lifecycle: register_handler remains append-only and duplicate-checked, while unregister_handler(&HandlerId) removes only that handler's future decode routes and interests. It does not purge EvmCache, reset health/metrics, clear the reorg journal, or drop root/freshness tracking; cache eviction stays an explicit caller action. The provider-agnostic EventSubscriber trait and AlloySubscriber are included; the Alloy subscriber uses WebSocket/pubsub subscribe_logs, subscribe_blocks, and subscribe_pending_transactions by default for live log, block-header, and pending-transaction-hash inputs. If an established WebSocket subscription stream terminates, the subscriber recreates that source immediately, retries three times by default with exponential backoff between later attempts, and backfills log subscriptions from the last seen block through get_logs, marking recovered records as InputSource::Backfill while suppressing recent duplicate canonical inputs. HTTP polling watch_logs / watch_pending_transactions remains available behind the opt-in reactive-polling feature. For pool/feed churn (register a new AMM on a PoolCreated event; drop one that is no longer of interest), the recommended binding is ReactiveEngine, which owns a ReactiveRuntime plus an EventSubscriber and drives handler lifecycle as one operation: engine.register_handler(handler) updates runtime routing and subscriber interests together and — once ingestion has journaled a canonical block — backfills the new handler from that block automatically, so a pool discovered mid-stream misses none of its own logs between discovery and live subscription (register_handler_with_backfill for deeper history, register_handler_live_only to opt out). Growing an existing handler's filter set is continuity-safe too: the changed subscription inherits the old delivery anchor and self-heals the gap. Use stable per-pool or per-adapter HandlerId values. Dropping an adapter is engine.unregister_handler(&id) for routing/transport, plus — for a pool that will not return — runtime.untrack_account (stop root-gate eth_getProof probes) and runtime.cancel_pending_resyncs (drop its queued repairs); cache eviction stays an explicit caller action. Full block bodies and full pending transaction hydration remain explicit follow-up transport work.
• Cold-start — declaratively warm a working set of accounts and storage slots into the cache in one batched pass via EvmCache::run_cold_start and a ColdStartPlanner (discover slots via a view-call, then verify them), returning a structured ColdStartRunReport. This is how a consumer adopts a working set (pools, feeds) into the fork before going reactive. (Reactive-gated.)
• ERC20 helpers — balances, allowances, decimals, and controlled balance / allowance mutation for simulations — layout-aware across Solidity, Vyper, and Solady/assembly token storage (not just Solidity-order tokens).
• Trace-based storage-slot discovery — derive a mapping's base slot and its byte-order layout from a single instrumented eth_call, by matching the KECCAK256 preimage that feeds each hashed SLOAD against the value the call returns — no 0..n slot brute-forcing. Works across Solidity (keccak(key‖slot)), Vyper (keccak(slot‖key)), Solady packed layouts, and nested mappings such as allowances (keccak(k2 ‖ keccak(k1‖slot))). The HashStorageProbe inspector is token-agnostic — trace_hashed_slots derives any hash-keyed slot for any call — and reusable TrackedMapping descriptors recompute any key's exact slot for freshness/prefetch pinning.
• Overlay-scoped mocking — EvmCache::mock_overlay() hands you a throwaway EvmOverlay carrying mock_balance / mock_allowance / mock_call; each derives the driving slot via discovery, writes it to the overlay's dirty layer, and verifies — so mocked state lives only for that simulation and never persists to the cache. Zero-address balance/owner writes are refused.
• Transfer-inspector simulation that reports per-token balance deltas straight from the Transfer event stream, no extra pre/post balance queries.
• Call-frame tracing — CallTracer reconstructs the nested CALL/CREATE frame tree of a simulation (from/to/value/gas/status/subcalls); InspectorStack composes it with transfer capture (or any revm::Inspector) in a single pass, driven through EvmOverlay::call_raw_with_inspector.
• Access-list tooling — StorageAccessList captures the EIP-2929 warm-access touch set; helpers build an EIP-2930 access list and estimate whether attaching one is profitable on an L2.
• Multicall3 batching for running many view calls inside the fork in one pass.
• Bulk storage extraction — the default storage loader — thousands of storage slots (across many contracts) in a single eth_call via state-override code injection (Dedaub's technique), with automatic point-read fallback for providers without override support. One 26-CU call replaces up to 10,000 20-CU eth_getStorageAts on Alchemy; a full Uniswap V3 pool tick range (7,674 slots) loads in 2 calls / ~220 ms, and eth_callMany dispatch drops the whole batch to 20 CU. Also ships custom storage programs (derive what to read in-EVM — e.g. a one-shot V3 observation-ring loader with zero calldata), bulk account-field and block-context extractors, and EvmCache::prewarm_slots for declared working sets — see docs/bulk-storage-extraction.md.
• Verified code seeding — skip eth_getCode (and the lazy backend's per-account round trips) for contracts whose bytecode the adapter already knows: seed_account_code writes the claim, verify_code_seeds settles the entire pending set against on-chain EXTCODEHASH in one eth_call, and a confirmed seed is Verified durably (persisted across restarts, never re-checked). A wrong template degrades to one refetch — never to wrong sims — and a newly detected contract materializes fully verified in ~1 round trip. The cold-start driver verifies pending seeds before anything simulates.
• Deployment & etching — deploy from creation code, or etch runtime bytecode (etch_account_code for raw bytes, no source account needed) over a forked contract while preserving its storage; every locally-divergent code site is tracked and queryable via etched_accounts().
• CREATE3 address derivation utilities.
• An extensible revert decoder — the two Solidity built-ins (Error(string) and Panic(uint256)) decode natively; register your own contract-defined custom errors in one line. Duplicate custom-error selectors keep the first registration and can be rejected explicitly with try_register*.

Quick start

use std::sync::Arc;

use alloy_eips::BlockId;
use alloy_provider::{ProviderBuilder, network::AnyNetwork};
use alloy_primitives::{Address, Bytes};
use alloy_sol_types::sol;
use evm_fork_cache::cache::EvmCache;
use revm::primitives::hardfork::SpecId;

sol! {
    function balanceOf(address account) external view returns (uint256);
}

# async fn example() -> Result<(), Box<dyn std::error::Error>> {
let provider = ProviderBuilder::new()
    .network::<AnyNetwork>()
    .connect_http("https://example-rpc.invalid".parse()?);

// Build a cache pinned to the latest block. (Requires a multi-thread tokio
// runtime — see the note below.)
let mut cache = EvmCache::builder(Arc::new(provider))
    .latest_block()
    .spec(SpecId::CANCUN)
    .build()
    .await;

let token = Address::repeat_byte(0x22);
let owner = Address::repeat_byte(0x33);
let balance = cache.call_sol(token, balanceOfCall { account: owner })?;
println!("owner balance: {balance}");

let from = Address::ZERO;
let to = Address::repeat_byte(0x11);
let calldata = Bytes::new();

// Simulate, capturing the EIP-2929 touch set as we go.
let (_result, touched) = cache.call_raw_with_access_list(from, to, calldata)?;
println!(
    "touched {} accounts and {} storage slots",
    touched.account_count(),
    touched.slot_count()
);
# Ok(())
# }

Runtime requirement. EvmCache lazily fetches missing state through a synchronous façade over an async provider (tokio::task::block_in_place), so its constructors and any method that may touch RPC must run on a multi-thread tokio runtime (#[tokio::main(flavor = "multi_thread")] or #[tokio::test(flavor = "multi_thread")]). The offline examples and tests build the cache over a mocked provider and never touch the network.

Core concepts

The state stack flows bottom-to-top; reads flow up and the fork DB lazily fetches misses from RPC. The event-log path writes hot state in with no RPC (the reactive-sync control plane):

flowchart BT
    RPC["RPC provider"] -->|"lazy fetch · once"| CACHE
    LOGS["on-chain event logs"] -.->|"decode → write · 0 RPC"| CACHE
    CACHE["<b>EvmCache</b> · !Send<br/>fetch · cache · targeted writes/purge"] -->|"snapshot()"| SNAP
    SNAP["<b>EvmSnapshot</b> · Send + Sync<br/>immutable · Arc · point-in-time"] -->|"cheap Arc clone × N"| OV
    OV["<b>EvmOverlay × N</b> · Send<br/>isolated parallel simulations"]
    classDef hot fill:#102a17,stroke:#3fb950,color:#e6edf3;
    classDef cool fill:#0d1f2d,stroke:#388bfd,color:#e6edf3;
    class SNAP,OV hot;
    class RPC,CACHE,LOGS cool;

• EvmCache owns the mutable fork: it fetches, caches, persists, and applies targeted writes/purges. It is !Send (it block_on's RPC internally).
• EvmSnapshot is an immutable flattening of the cache at a point in time, shareable across threads via Arc.
• EvmOverlay wraps a snapshot with a per-simulation dirty layer; clone one per candidate transaction and simulate without RPC and without touching the live cache.

The freshness module layers a freshness controller on top: classify each address/slot (Pinned / Volatile / ValidThrough), observe how often slots change, pick what to verify each cycle with a FreshnessPolicy, and run the optimistic loop that returns speculative results immediately and a Confirmed/Corrected/Unverified verdict asynchronously. Time-to-actionable-result is gated on local simulation, not on the RPC validation that runs behind it:

sequenceDiagram
    autonumber
    participant S as Search loop
    participant C as FreshnessController
    participant V as Background validator
    participant R as RPC
    S->>C: run(candidate sims)
    C->>C: snapshot + run optimistic sims
    C-->>S: SpeculativeSim — optimistic results (~µs)
    Note over S: act on speculative results now
    C->>V: spawn (Send data only)
    V->>R: verify volatile read-set (~L ms)
    R-->>V: fresh values
    alt nothing the sims read changed
        V-->>S: validate().await? → Confirmed
    else a read slot changed
        V->>V: re-run only the affected sims
        V-->>S: Corrected { results, changed }
    end

Examples

The examples/ directory has runnable, documented examples. Run any with cargo run --example <name>.

Offline examples need no network — they build the cache over a mocked provider and inject all state directly:

RPC examples fork real mainnet state. Set RPC_URL to an Ethereum RPC endpoint (they print instructions and exit if it is unset):

cargo run --example revert_decoding
RPC_URL=https://eth.llamarpc.com cargo run --example fork_token_balance
WS_RPC_URL=wss://example-mainnet-endpoint cargo run --example reactive_alloy_amm_live_probe

Feature Flags

Default features enable the reactive runtime and WebSocket/pubsub subscriber support (reactive, reactive-ws). The HTTP polling subscriber is opt-in: consumers that disable defaults can enable reactive,reactive-polling.

Foundry artifact etching

Use etch_foundry_artifact when replacing an existing forked contract while preserving its storage, balance, and nonce. Use etch_foundry_artifact_or_create for synthetic simulation addresses. See the runnable foundry_artifact_etching example.

use alloy_primitives::Address;
use evm_fork_cache::deploy::{encode_constructor_args, etch_foundry_artifact_or_create};

# fn example(cache: &mut evm_fork_cache::cache::EvmCache) -> Result<(), Box<dyn std::error::Error>> {
let target = Address::repeat_byte(0x42);
let constructor_args = encode_constructor_args((Address::ZERO,));

let etched = etch_foundry_artifact_or_create(
    cache,
    target,
    "out/MyContract.sol/MyContract.json",
    Address::ZERO,
    constructor_args,
)?;

println!("installed {} bytes at {}", etched.code_size, etched.target_address);
# Ok(())
# }

Performance & honest trade-offs

It is easy to post huge multipliers against a naive loop (a fresh cold fork per candidate that re-fetches everything and deep-clones to isolate). That is not the loop a competent revm user writes. Measured against a competent baseline — one shared foundry-fork-db SharedBackend (which caches and deduplicates fetches) plus checkpoint/revert isolation on a single fork — this crate is roughly at parity on raw within-block speed, and we say so plainly:

The value of this crate is not a within-block speed multiplier. It is correctness, cross-block freshness, and a structured control plane the bare primitives don't give you:

① Cross-block freshness — the one quantitative win (exact, CI-pinned integer). foundry-fork-db's cache is not block-keyed: re-pinning to a new block does not invalidate cached slots, so a refresh-by-refetch loop must re-read every slot that changed each block to stay correct. Decoding the block's logs into targeted writes keeps that hot state correct with 0 RPC fetches/block — the log→write path runs fully offline in tests/event_pipeline.rs, and the zero-extra-fetch integer is tallied by a real fetch counter in tests/fetch_minimization.rs. Sampled reconcile() re-reads a fraction to catch drift (the honesty backstop). See reactive_cache.

Honest caveat: an equally sophisticated peer running their own log subscription + delta applier also reaches 0 fetches/block. The crate's contribution is the packaged, cold-aware, reorg-safe, reconcilable vocabulary — not an unreachable number.

② Parallel fan-out — available, modest, workload-dependent. snapshot() is an immutable Send + Sync view; cloning the Arc hands each thread its own overlay, so candidates fan out across cores — which a single mutable fork cannot do. The measured speedup is honest and modest: ~1.2× across the 64–1,024-candidate sweep (cargo bench --bench fanout) on a 10-core M1 Pro, because these micro-sims are bound by per-candidate allocation, not EVM compute. The ratio scales with both core count and per-candidate compute weight. Heavier candidates (real txs doing substantial execution) parallelize better; trivial ones barely. We don't headline a core-count multiplier we can't reproduce. cargo bench --bench fanout; parallel_overlays.

③ Point-in-time consistency. Every overlay reads one frozen, consistent block state. A lazily-filled shared backend can interleave reads taken at slightly different moments unless carefully pinned; the snapshot removes that class of bug.

④ Act-then-validate control plane (structure, not speed). Run optimistically against current state, return immediately, and validate the volatile read-set in the background — re-running only the sims whose slots changed (rerun_count), with Confirmed/Corrected/Unverified verdicts and block-pinned validation. A searcher who simply acts on warm state is equally fast; the value is the safe, selective re-run, not a latency multiplier. See freshness_optimistic.

⑤ Cold-load economics — the second quantitative win (live-measured). Since 0.2.0 the default batch storage fetcher packs slot reads into single eth_calls whose target code is overridden with a 23-byte extractor (Dedaub's bulk storage extraction — full credit to their write-up and reference implementation), so a cold working set loads in a handful of calls instead of one billed read per slot. Live-measured on Alchemy mainnet (RPC_URL=… cargo run --release --example bulk_storage_bench, medians of 3):

CallDispatch::CallMany drops any batch to a flat 20 CU on Erigon-lineage endpoints (Alchemy included); custom storage programs go further and derive what to read in-EVM (a one-shot V3 observation-ring loader ships as a worked example); the code-seeding row rides the same transport's account-fields extractor — one call settles every pending bytecode claim against on-chain EXTCODEHASH and materializes real balances, with zero code bytes on the wire. Providers without state-override support degrade automatically to the classic point-read path. Methodology, latency tables, gzip measurements, and every limitation found: docs/bulk-storage-extraction.md.

The CU costs are deterministic and exact — re-verified live three times through 2026-07-05, identical every run. The wall-clock latencies above are a conservative floor: they were captured across constrained, variable networks, so well-provisioned connectivity should meet or beat them (repeat runs measured the same loads up to several× slower purely from network load, never faster CU).

[!NOTE] Methodology & candor. Offline (mocked provider, state injected — no network). We deliberately do not lead with the headline multipliers a naive baseline would produce (~500× fewer reads, ~545× throughput, ~3,800× latency): all three collapse toward ~1× against a competent SharedBackend + checkpoint/revert loop — which is the very primitive this crate wraps. The zero-extra-fetch integer is real and CI-pinned (cargo test --test fetch_minimization, with the log→write path exercised offline by cargo test --test event_pipeline); the parallel-fan-out ratio is a Criterion median on an Apple M1 Pro, read as a ratio not an absolute. Live-RPC checks live behind the RPC_URL gate.

Phase 8's trace-backed resync path has separate live-RPC measurements in docs/trace-resync-benchmarks.md: on Alchemy CU pricing, one block trace breaks even with two storage reads and wins at three or more; in latency tests, batched storage stayed faster for small known slot sets, while gzip materially reduced large Alchemy trace response latency. eth_getProof — the one call with no bulk substitute (storage roots and nonces are not EVM-visible) — is kept off the per-block path entirely: root-gate probes fire on a cadence (default every 16 blocks, RootGateCadence — 16× fewer probes than per-block, by construction) and each firing batches every gated account into one fan-out (EvmCacheBuilder::max_concurrent_proofs, default 8 — live-measured ≈4.7–7.3× over serial for a 50-account sweep across runs, bounded by the cap and larger when per-proof latency is higher). Reproduce the fan-out measurement with the RPC-gated test E2E_RPC_URL=… cargo test --test liveness_root_gate -- --ignored (default_proof_fetcher_fans_out_concurrently).

Production safety checklist

The defaults favor doing something reasonable over failing; production deployments should opt into the strict/observable variants deliberately:

• Multi-thread tokio runtime. RPC-backed calls bridge sync→async via block_in_place; on a current-thread runtime they degrade to typed RuntimeErrors. Use #[tokio::main(flavor = "multi_thread")].
• Pin a concrete block (EvmCacheBuilder::block / EvmCache::at_block) for anything that must be reproducible; latest pins are for exploration.
• Strict block context. builder.strict_block_context(true) + try_build() so a missing basefee/prevrandao fails loudly instead of silently defaulting the EVM env (per-field knobs via BlockContextRequirements for pre-London/pre-merge chains).
• Set the chain id explicitly (EvmCacheBuilder::chain_id) rather than relying on eth_chainId inference with its mainnet fallback.
• Watch the health surface. Poll ReactiveRuntime::health() and treat Degraded/Unhealthy as "stop trading until resynced"; export metrics() counters (deep_reorgs, resync_failures, coverage_gaps, missed_ranges) to your dashboards.
• Track balance/nonce-sensitive accounts. Storage-only freshness cannot see a native-balance/nonce/code move that shifts no storage slot. Root-gate such accounts with ReactiveRuntime::track_account + TrackingPolicy::WholeAccount/Scalars: an eth_getProof root probe catches the drift and reactive resync repairs it, keeping the cache fresh. This is the reactive runtime's account-freshness path — it is separate from the speculative freshness validator, whose success verdict is ConfirmedStorage (storage slots only). ConfirmedFull (storage and verified account fields) is defined but not yet emitted — validator-side account verification is a tracked follow-up (see the verdict taxonomy in freshness).
• Gate on code-seed verification. If you seed bytecode (seed_account_code), require the round's CodeVerifyReport.unverifiable bucket to be empty and pending_code_seeds() drained before serving sims; audit deliberate divergence via etched_accounts().
• Size reorg horizons deliberately. ReorgConfig::depth and ReactiveConfig::journal_depth bound purge/rollback reach: a reorg within the journal is rolled back precisely, but effects from blocks that have already aged out of the journal are not auto-purged. A reorg that deep escalates health to Unhealthy (with a warn!) and leans on freshness validation as the backstop — treat it as "resync before trusting sims" and size the horizons above the deepest reorg you intend to recover precisely.
• Know your provider. The default bulk storage loader needs eth_call state-override support (major providers have it; the fetcher latches to point reads after two fully-failed batches — a warn! you should alert on, or poll it directly by building the fetcher via bulk_call_storage_fetcher_with_status and checking BulkFetcherStatus::fallback_latched); the trace resync accelerator needs the debug namespace and falls back to point reads without it. Enable gzip on the HTTP client for both (docs/bulk-storage-extraction.md).
• Persisted state is trust-gated, not trusted. Load disk state with a RootBaseline (roots.bin) so restart drift is detected via root probes instead of silently simulating on stale slots.
• Read docs/KNOWN_ISSUES.md — the accepted limitations (BLOCKHASH-in-overlays, decoder assumptions, deep-reorg bounds) are documented there rather than discoverable by surprise.

Benchmarks

Criterion benchmarks live in benches/ and run fully offline (mocked provider) so they are reproducible:

cargo bench                      # all offline benches
cargo bench --bench fanout       # one suite

The rpc_mainnet bench runs against live mainnet state to validate real-contract performance (USDC balanceOf, totalSupply, and allowance). It is gated behind the RPC_URL environment variable and is skipped (not failed) when it is unset, so cargo bench stays offline and CI-reproducible by default:

RPC_URL=https://eth.llamarpc.com cargo bench --bench rpc_mainnet

Crate boundary

evm-fork-cache is the generic simulation engine: cache, snapshots/overlays, freshness control, access lists, revert decoding, ERC-20 helpers, multicall, deployment, CREATE3, and event-pipeline primitives. AMM state tracking, protocol-specific storage layouts, and DeFi adapters belong in the companion evm-amm-state crate or downstream applications.

Stability

evm-fork-cache is pre-1.0. Until 1.0, breaking changes may land in minor releases — the roadmap deliberately reshapes the API before the surface freezes. Each release documents its breaking changes in CHANGELOG.md.

• MSRV: Rust 1.88 (enforced in CI). Edition 2024.
• Semver: pre-1.0 minor versions may break; patch versions will not.
• Roadmap: see docs/ROADMAP.md for the path to 1.0.
• Known issues / limitations: see docs/KNOWN_ISSUES.md.

Contributing

Contributions are welcome — see CONTRIBUTING.md for branch conventions, the green-bar CI expectations, and the commit format.

License

Licensed under either of

• Apache License, Version 2.0 (LICENSE-APACHE)
• MIT license (LICENSE-MIT)

at your option.

Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in this crate by you, as defined in the Apache-2.0 license, shall be dual licensed as above, without any additional terms or conditions.