Physical Direct multiprogram composition — 2026-08-25

Historical artifact evidence. The active signed, compiled, canonical-owner successor is pinned in COMPILED_SIGNED_DIRECT_2026_08_25.md. The measurements below remain exact for their named source commit and bytes.

Scope

This evidence concerns source commit 803286409a4aeaf1ce052ffab98705088bff208b. It composes four real SBF artifacts in one transaction:

1. the Direct semantic-controller experiment;
2. the 1,872-byte Lean-owned claim executor;
3. a custody micro-adapter for the Lean-owned 40-byte transfer plan; and
4. the official legacy SPL Token program built from published crate 9.0.0.

The successful route updates replay/claim state and performs two TransferChecked CPIs. The hostile route allows the first token CPI to finish, then discovers a frozen fee destination and demonstrates transaction-wide rollback across the controller, claim, and token programs.

This is not a complete Direct successor. The controller still accepts pre-derived plans instead of authenticating signed intents and deriving the unique admissible plans itself. It also pins the legacy Token program directly instead of consuming an immutable Realm, and it does not authenticate exact controller ProgramData or a checked release manifest.

Reproducible artifacts

The three first-party programs were built from the exact source commit with:

cargo build-sbf \
  --manifest-path programs/dclutch-claims-proof-sbf/Cargo.toml \
  --lto --optimize-size --sbf-out-dir target/deploy
cargo build-sbf \
  --manifest-path programs/dclutch-controller-proof-sbf/Cargo.toml \
  --lto --optimize-size --sbf-out-dir target/deploy
cargo build-sbf \
  --manifest-path programs/dclutch-custody-proof-sbf/Cargo.toml \
  --lto --optimize-size --sbf-out-dir target/deploy

The token artifact was built without a native test processor from the spl-token 9.0.0 crate source:

cargo build-sbf \
  --manifest-path <cargo-registry>/spl-token-9.0.0/Cargo.toml \
  --sbf-out-dir target/deploy

The published crate archive SHA-256 is 878b0183d51fcd8a53e1604f4c13321894cf53227e6773c529b0d03d499a8dfd. Its .cargo_vcs_info.json binds program/ to the official solana-program/token commit dfb260231c761be7d9c8b63728e770a102b86495. The build used cargo-build-sbf 4.0.0, platform-tools v1.53, and SBF rustc 1.89.0.

Capitalization uses Rent::default(), one 36-byte Loader V3 Program account, and 45 bytes of ProgramData metadata per program. SPL Token's number is an equivalent local measurement, not dClutch deployment capital: the canonical legacy token program is already deployed. Transient buffers, state accounts, and transaction fees are excluded.

Real-SVM evidence

physical_direct_composition.rs used solana-program-test 4.2.1, preferred SBF, loaded all four exact ELFs, and registered no native processor or token mock. Reproduce it with:

SBF_OUT_DIR="$PWD/target/deploy" \
  cargo test \
  --manifest-path crates/dclutch-svm-harness/Cargo.toml \
  --test physical_direct_composition -- --nocapture

Observed routes:

• direct controller-PDA impersonation refused in 7 CU;
• the wrong maker replay-root bump refused in 4,530 CU;
• the complete physical fill committed in 24,901 CU; and
• a frozen second destination refused in 19,420 CU after the first real Token TransferChecked logged success.

The committed example increments the journal and both replay nonces, changes the claim balances to 3,000 and 2,200, transfers 1,000 base units to the seller and 2 to the venue, and consumes the source delegate's exact 1,002-unit allowance. The hostile example compares the complete controller journal, claim projection, source, seller, and venue accounts before and after refusal; all five are restored byte-for-byte.

Correctness boundary

Lean owns the admitted example's claim and custody plans, exact quote and fee arithmetic, conservation statement, canonical bytes, and abstract atomic transition. The claim executor has a separately pinned qedsvm successful-path theorem. The custody adapter and controller are safe Rust adapter experiments, and this campaign is runtime evidence rather than a proof of Solana, SPL Token, their CPI composition, or hostile-path completeness.

The experiment removes the prior false model in which executor-owned integers stood in for collateral. Real token accounts now move under the existing Direct V2 maker replay-root delegate. The next gate is to stop accepting effect plans from instruction data: the controller must authenticate the signed intent, Product, Realm, release, and chain frame, then derive the only plans it is willing to authorize.