FINAL5 architecture and ownership review

Reviewed 2026-09-12 against frozen FINAL5 source archive SHA256 7c63fcad9d12a220036b531f85eacc1a91984ff87c18687703aeeac57b3df415. This review does not publish a release or replace the separate 198-proof gate. Production was initially inspected read-only; the two concrete shared-TIR findings below were subsequently authorized for repair in live source. Frozen FINAL5 is unchanged.

Actual ownership

Trident's Cargo manifest has no Triton VM, tasm-lib or Neptune dependency. Its catalog/vm/triton/owner.toml and catalog/os/neptune/owner.toml contain external-owner discovery metadata. The owner names and terrain/union matching in src/config/target/discover.rs are dispatch and identity checks, not machine implementations. Other catalog designs do not establish executable support. There is no portable trident/lib/os implementation to migrate. Language-level os.state.read is a semantic operation; the state certificate, runtime and proof boundary belong to the runtime owners.

Installed resources and API 3

Trident build.rs embeds sorted lib and catalog source/descriptors. Trisha rs/build.rs separately embeds its lib, targets and networks. rs/target.rs exports Triton modules for triton, adds Neptune modules/states and canonical policy intrinsics only for neptune, and seals the package. Joy rs/target.rs uses the canonical upstream nox ABI and embeds its own runtime capabilities. The retired handwritten Neptune proof implementation is outside the production SDK; the production recursive entry belongs to Trisha.

TargetPackage::validate checks compiler API/schema, bounded ABI/module data, module declaration identity and hashes, selected OS namespace, union/terrain and state identities. CompileOptions::with_package takes the package sources and ABI together; validate rejects substituted ABI, including noncanonical nox shapes. Generated generic hash/I/O declarations derive their dimensions from the selected ABI. Reachable intrinsic capability checks remain necessary: an unused portable module declaration does not assert runtime availability.

Resolver precedence is explicit overlay, supplied package sources, explicit dependency, compiler embedded library; unsupplied reserved std/vm/os modules do not silently fall back to an ambient checkout. An overlay is an explicit editor input, not a runtime provider replacement.

Sixteen selected implementation/reference files (resources, builders, package/discovery, API, Trisha entry/target/build, Joy target/capabilities and three canonical references) were compared byte-for-byte between live and frozen FINAL5 before repair and were identical. Installed FINAL5 Linux and Darwin smoke receipts additionally exercise these resources outside the source checkout. See Trisha audit/final5-linux-validation.md and audit/final5-darwin-receipt.json for exact binaries and proof interoperability; these receipts are stronger evidence than a catalog listing.

Concrete findings and live repairs

1. src/ir/tir/builder/expr.rs tracked XFieldMul results as three words even for a different selected extension width. FINAL5 Triton uses three words; direct nox does not use this TIR path. The live repair uses target_config.xfield_width. The regression checks widths 2, 4 and 7 and observes the exact depth of a retained neighboring value after the product.
2. src/ir/tir/builder/legalize.rs redundantly lowered deep Dup/Swap into compiler scratch RAM according to the target stack window. Although typed and not literal TASM, this violated the agreed owner-only machine-window boundary. The live repair removes that module and the build-file final pass. Trisha's existing machine legalizer remains authoritative. Compiler scratch for semantic temporaries, indexing and generic spill management is retained. Native program identities/costs may change; FINAL5 proofs cannot be relabeled as proofs of changed native programs.

Documentation contradictions requiring cleanup

• reference/os.md says Joy rejects state requests and lacks authenticated public proof support. Current Joy capabilities and actual FINAL5 smoke support bounded authenticated public BBG certificates, JOYST001, and hidden query coordinates over complete public tables through JOYZK003. There is still no live Cyber state synchronization/deployment or private database guarantee.
• The early API example in reference/warrior-api.md recommends TerrainConfig::triton() and old vm/<engine>/target.toml paths. The former is test-only; runtime integrations must use the owner package resolution APIs. Its claim that supplied module sources only apply when no file exists also disagrees with the explicit package-source precedence above.
• Earlier wording in that reference calls the migration interfaces merely proposed, while its later API 3 section documents the implemented contract.
• Trisha architecture previously described compiler final-pass deep access legalization. The live description now assigns it to the owner lowerer.
• audit/target-ownership.md contains explicitly dated historical coverage and platform gaps. Those are history, not current support claims; use current candidate receipts and the active release ledger for readiness.

The supported runtime matrix is narrower than the catalog: Trisha provides Triton/Neptune CPU execution and STARK proofs; generic deployment remains false, with a separate validated-intent Neptune CLI demonstrated only at the isolated local node boundary. Joy provides nox and its cyber alias, public execution, private Triton-backed execution, and bounded authenticated public-state proofs. GPU mining selection does not select a GPU proving backend. None of this review claims public-network admission, block confirmation or unrestricted proof shape.

Live repair validation

The authorized repairs completed with these bounded checks:

• Core TIR suite: 51 passed, zero failures/ignored, 629 filtered.
• Trisha target_ownership: 11 passed, zero failures/ignored. Raw deep Dup/Swap checks every word at depths 16, 17, 31 and 64; the new source case reverses a 40-word array through a callee and verifies the returned array, original caller array and scalar sentinel in both debug and release profiles.
• Alternate XField widths 2/4/7 preserve the neighboring stack value's exact access depth. No arithmetic semantics for an unimplemented target are claimed.
• Independent custom-token canonical codec/own-program fixture generator: one selected test passed. The initial unchanged-fixture run retained the expected three stale self-hash rejections (130/133 passed). Only input and secret fields of the three positive custom-token fixtures were regenerated; all hand streams, expected outcomes, amounts and negative fixtures are unchanged.
• Fresh full execution with both exact FINAL5 and live binaries: each 133/133 fixtures and 43/43 baselines passed. Of 99 positives, 48 use fewer cycles and 51 are unchanged; none increases. All hand cycles and 34 rejection statuses are unchanged. Total source cycles across these fixtures decrease by 15,283.
• Actual native Program hashes change in 40 of 84 distinct source/target pairs; 44 are identical. This is a real program-identity change, not whitespace. There were no new proof generations in this repair gate.

Machine-readable before/after native hashes, all 133 cycle rows, binary hashes, fixture rebinding identities and log hashes are preserved in owner-stack-legalization-receipt.json. Frozen FINAL5 proof receipts continue to describe FINAL5 only; changed live programs require fresh proof evidence before their own release candidate.

Files changed for this repair: src/ir/tir/builder/expr.rs, src/ir/tir/builder/mod.rs, deletion of src/ir/tir/builder/legalize.rs, src/ir/tir/builder/tests/advanced.rs; in Trisha, rs/tests/target_ownership.rs, docs/explanation/architecture.md, and the three baselines/triton/fixtures/custom-token-v2-{burn,mint,transfer}/vector.bench.toml files. No owner ISA algorithm needed modification: the existing Trisha deep access implementation passed the source and raw-IR checks unchanged.

Rust compilation emitted no compiler warnings in these targeted gates. The existing target_ownership negative/source fixtures print two language warnings for unused imports (vm.crypto.hash and os.neptune.xfield); these are recorded, not counted as zero diagnostic output. Whitespace checks passed for the edited Rust and architecture files.