dev: add a proof-erased sim test backend (real ledger, no proving)

Context

The test suite has two backends today: dry (pure in-memory circuit logic, fast, but mocks ledger semantics — e.g. placeholder mt_index: 0n) and live by https://github.com/OpenZeppelin/compact-contracts/pull/673 (full node + proving, real fidelity, hours per category). Bugs in the coin-tree / dust / serialization / ledger-apply layer are invisible on dry and expensive to catch on live (#656 is the canonical example: dry missed it, live caught it).

A spike by @andrew-fleming confirmed a middle tier is possible with the official @midnight-ntwrk/wallet-sdk-capabilities Simulator: deploy + state-mutating circuit call ran proof-erased, fully in-process — real ledger-v8 LedgerState (coin tree, nullifiers, dust, serialization), no proof-server, node, indexer, or docker.

Proposal

Add a third backend, MIDNIGHT_BACKEND=sim, behind the existing backend seam:

1. dry — circuit logic floor: guards, reverts, arithmetic. Inner dev loop.
2. sim (new) — real ledger semantics, proofs erased, instant blocks. PR-gate default for value / coin / serialization paths.
3. live — real proving + node. Nightly smoke for what sim cannot cover (circuit satisfiability, prover behavior, node acceptance policy).

Key design rule: no `` boolean. sim breaks the two-axis coincidence that `isLiveBackend()` relies on, so conditional sites re-key to capability predicates:

backend.hasRealLedger       // dry:false  sim:true   live:true
backend.usesRealNode        // dry:false  sim:false  live:true
backend.directLedgerAccess  // dry:false  sim:true   live:false

Value branches (getQualifiedShieldedCoinInfo placeholder-vs-real mt_index) key on hasRealLedger; node-dependent skips key on usesRealNode. Some live skips flip to runs on sim (e.g. burn-success: the coin tree is in-process and directly queryable, no wallet ciphertext discovery needed).

StrictnessConfig is per-axis, so balancing / signatures / conservation can stay ON with only proof verification OFF.

Open questions (must be answered by the prototype)

• Value flows. The spike proved deploy + a no-coin circuit call. With enforceBalancing ON, a deposit needs a funded input coin in the zswap state, and there is no wallet or faucet in-process. Fabricate genesis-like coin state, or relax balancing (losing fidelity)? This is the make-or-break question.
• Benchmark vs dry. Sim per-circuit-call cost is unmeasured (tx build + transcript + ledger apply + WASM serialize-bridge). Measure it; if sim lands within a small factor of dry, reconsider whether dry remains the PR floor.
• Serialize-bridge fragility. RtContractState.deserialize(simContractState.serialize()) couples two WASM instances (ledger-v8 vs compact-runtime). Pin wallet-sdk-capabilities and cover the bridge with a test so SDK bumps fail loudly.
• Multi-party flows. .as(alias) has no pooled wallets on sim; signer identity can map to witness-set keys as on dry, but coin ownership needs per-alias ZswapSecretKeys. Needs a design note before multisig specs pass.

Suggested first step

A one-file prototype: a SimSimulatorBackend registering as sim that runs one existing multisig spec green end-to-end, with the branches that spec hits re-keyed to capability predicates. The spike only proved the raw deploy+call round-trip; this proves the spec/DSL integration and answers the value-flow question.

Non-goals

• Replacing live: proving bugs (e.g. the unprovable guarded-branch class) and node acceptance policy remain live-only.
• Landing in #673: the harness PR stays as-is; this builds on its backend seam afterwards.