Files
prometeu-runtime/docs/vm-arch/IDENTITY_HASHING_QUALIFICATION.md

6.6 KiB

Identity and hashing qualification

Qualification date: 2026-10-23. Target: aarch64-apple-darwin. Toolchain: rustc 1.98.1 (48a229cea 2026-09-01), cargo 1.98.1 (797e8a9bc 2026-08-05). The machine-readable record contains exact source/artifact SHA-256 values and commands. The canonical contract is normative; the producer handoff defines consumer work.

Delivery identity

Base revision: 57992cf214e8a5f9de53a40dba9aaa7fab038c91. This qualification covers working-tree changes identified by source fingerprint a0d202e46fbca3319d7b04fe09906388c015da7b606216e75bd7c62041011885. The base revision alone does not contain this delivery. The JSON lists the 29 source/test files and their individual digests. To reproduce the aggregate, visit sorted paths and hash the concatenation of path UTF-8, NUL, exact file bytes and NUL for each file.

Artifact SHA-256
Cargo.lock cffad9ed62f38a591538abd81dd39270d6022397342c8189f4ac6c7baaa8e795
INTRINSICS.csv 7c1c363f5f0f941b3ab27cfc72af41e13e44cd4371c687db8300721288f47e63
BUILTIN_CONSTANTS.csv e407e4718127aa7ef7b1097b5b4b67d86dd979235e73a551a7c36b628cddc039
hash-conformance.tsv a863486eaea14ab1a010f076f355eed0d5738207dad3651f7d10e92fc09f74d4

Schema 2 has 25 columns and 108 intrinsic rows. Comparison against the base revision confirms that all 18 prior fields of each of the 105 old entries are unchanged. Three hash entries were added. Builtin constants are byte-identical. Existing historical math/random qualification records retain their original revision claims.

Executed checks

All commands use the checked-in lockfile and offline dependency cache.

Command Result
cargo test --locked --offline -p prometeu-bytecode -p prometeu-vm --no-fail-fast 453 passed, 0 failed, 3 existing ignored.
cargo test --locked --offline -p prometeu-vm --features resource-diagnostics --no-fail-fast 401 passed, 0 failed, 2 existing ignored.
cargo test --locked --offline -p prometeu-quality-checks -p prometeu-layer-tests --no-fail-fast 19 passed, 0 failed.
cargo test --locked --offline --release -p prometeu-vm --test identity_hashing --test intrinsic_profiles --test execution_allocations 12 passed, 0 failed.
cargo check --locked --offline --workspace --all-targets Passed, including headless/desktop, firmware, tools and shared consumers.
cargo fmt --all --check Passed.
git diff --check Passed.

Independent oracles and covered behavior

The 118 vectors come from the official C XXH64 v0.8.3 reference, not the Rust implementation. The generator and provenance pin the upstream header by SHA-256 and verify its version. Raw byte vectors cover every length through 70 and longer stripe/tail boundaries through 4097; the remaining cases cover signed integer extremes, finite binary64 bit patterns, UTF-8/NUL/empty strings, reference index/generation pairs and direct null zero. Every vector is checked by the Rust algorithm. Scalar/String/null vectors also execute through serialized PBX, typed admission and the real interpreter.

Independent raw instruction bytes check IDs, LE immediates, truncations, assembler/disassembler behavior and Registry widths. Neutral PBX tests cover Struct/Array/Fn/Random identity, distinct closure creation, aliases, mutation, Random advancement, multiple Trait views, valid implementation zero, all eligible null groups and exact successful instruction costs. Admission rejects ineligible source types, bad immediates/IDs, wrong numeric carriers and old String nulls, including invalid metadata in unreachable code where appropriate.

Private fault injection covers non-finite Float bits, stale/malformed references, wrong concrete types, invalid/partial Trait pairs, left-before-right validation, invalid right operands despite null left, underflow, foreign String ownership, checked String cost overflow and malformed/missing/excess handler results. Assertions preserve complete operands and String owners, resource usage, Float NaN payload bits and charged cycles. Misbehaving output handlers use a different owner namespace to exercise the shared invocation protocol. Descriptor tests reject unsupported layouts, inconsistent metadata and undeclared capabilities.

Real GC retains live aliases and permits collection when only an Int64 hash is retained; reuse increments the generation and old handles still fault. Existing generation-exhaustion tests cover retirement without wrap. Hashing equal String content across image, session concatenation and explicit import owners agrees. Sequential handle distribution exercises 256 reduced buckets; deliberate bucket collisions still require reference equality. This is distribution evidence, not a uniqueness or adversarial-resistance claim.

Allocation, resources and scheduling

Ten new warmed instruction regions cover three scalar hashes, five reference hash categories, Struct equality and Trait equality. After ten warmup iterations, 1,000 measured iterations per region allocate/deallocate nothing natively, including reallocations, and create no guest heap objects. Original values remain in locals so owner destruction does not contaminate the measured region. The same checks pass in debug, resource-diagnostics and selected release execution.

Quota-injection tests exhaust guest/handle/auxiliary/temporary/String budgets before reference/scalar queries and check unchanged resource use. New queries do not trigger GC, grow their result stack, request scratch or call the host. Fault checks cover guest resource/owner preservation; the native allocator measurement covers successful warmed execution, not the existing diagnostic formatting path. Zero fuel leaves PC/state unchanged; one positive cycle admits the entire instruction, including a 65-byte String hash costing 73, with no extra generic INTRINSIC cycle. Foreign String owners are rejected only after full byte cost.

Scope of evidence

This record covers the local target and configurations above. Other hosts, architectures and cross-host execution comparisons were not run. The explicit LE implementation and independent vectors support portability but do not replace qualification on additional advertised targets. Studio source compilation, normal/optimized source conformance, catalog pins/mirrors and packaged SDK resources remain the consumer handoff. No Lua frontend, Set/Map or structural/Dynamic equality is reported as delivered.