2026-08-19 - The execution model, decided and stubbed

Done. The concept flagged as needing input (D096 aftermath) came back as predicated, with structured stubbed loudly, and the foundation is laid.

• orbistoun-spirv: a module builder that emits words directly - no assembler, no text. Identifiers are allocated by the builder so the header bound cannot disagree with them.
• orbistoun-translate: the strategy seam. Predicated is a skeleton that reports translating zero instructions; structured is an error that says so at length.
• tools/validate-spirv.sh: spirv-val as the oracle, emitting on the host and validating in the VM so neither machine needs the other's toolchain.

Verified. Gate green. The minimal compute module validates first try and disassembles exactly as intended - every opcode number proposed from memory was right, which spirv-val is in a position to confirm and this crate is not.

Surprises.

• A one-off gate failure in orbistoun-abi, a crate untouched by any of this. It passed 8/8 on its own and three subsequent workspace runs were clean. The plausible cause is contention between test binaries that reserve fixed addresses - abi executes generated code at a fixed address and mem reserves ranges - and adding two crates changed how many run concurrently. Not chased, because it did not reproduce, but it is a latent fragility rather than noise and is now in the backlog.

• spirv-val lives in the VM and cargo does not. Splitting emit from validate across the two machines is better than installing a Rust toolchain in the VM, and it falls out naturally: the emitter writes a file, the validator reads one.

Not done. The predicated translator translates nothing - it emits a correctly shaped module and reports zero instructions, so nothing downstream can mistake a skeleton for a translation. Register modelling, the dispatch loop, and instruction emission are all ahead. Structured reconstruction is deliberately absent.