Independent render-artifact review, 2026-10-23
Historical source last updated, 2026-10-23. Runtime and implementation claims below describe that dated source; see current app notes and source for subsequent changes.
Verdict: REJECT the candidate's unrestricted executable-identity/cache
contract; do not integrate. Candidate
fd5e3fc9329711d3190bbd5c7046976c21d0b9a1, exact base/unchanged remote main
7f7358c160ae7185fc38f9a15447a6d20c00f97e. This is an independent source and
small-fixture protocol review, not native acceptance. No renderer/WAT, D3D
fixture 09dc88de, production helper, or main change is published by this review.
Concrete executable-identity failures
prepareArtifactchecks onlyrequire.cache[watx.js]. Importing the actualtest/compile-src.jsalso imports the actual closure adapter, but does not load WATX until compilation. Import that adapter, replace its disk file withthrow Error("CHANGED adapter must reject"), then prepare. Preparation succeeds using the previous cached function; the receipt seals the new, throw-only adapter and consumption accepts it. The before/after hashes are identical because the replacement is complete before preparation starts. This is a deterministic stale-code binding, without any concurrent writer during hashing or compilation. Checking the loader cache alone does not establish that the sealed adapters are the evaluated adapters.- Default mode hashes a changed compiler loader into a different cache key,
but
compile-srcandwatx-closureretain Node's previously loaded loader. First boot with the tiny loader, replace that loader with a throw-only file, verify thatidentity()changed, then boot again. The second boot succeeds using the old loader. Fresh disk identity invalidates the Wasm map entry, not the executable compiler cache. This control uses the real closure and compile adapter, with an eight-byte loader double and instantiation double.
The replay asserts both defects, rather than accepting any thrown exception as
evidence. See independent.js (historical artifact: d3d-render-artifact-independent-review-20261023/independent.js),
checks.log (historical artifact: d3d-render-artifact-independent-review-20261023/checks.log), and
execution receipts (historical artifact: d3d-render-artifact-independent-review-20261023/execution.jsonl).
Scope qualification: both failures require disk changes after a relevant module has been loaded. A wholly immutable tree for the entire Node process, started fresh, avoids these reproductions. The producer explicitly documents immutable inputs during an admitted command, which is a useful operating restriction; these findings do not allege failure under that narrower condition. However, the feature also expressly offers fresh identities and invalidation between default calls, and its prepare freshness guard does not enforce adapter freshness. The broad source/compiler identity claim therefore needs a narrower, enforced lifetime contract or an executable-cache correction before acceptance. The immutable-source requirement cannot by itself prove the requested behavior when exercising input invalidation between calls.
No speculative compiler input omission was found. The concrete omission is the relationship between hashed disk adapters/compiler and their cached evaluated versions. Merely deleting all require-cache entries is not a reviewed fix: other callers can retain exports from the same compiler, and default/prebuilt users currently share this process. A safe correction must explicitly settle that lifetime/compatibility behavior and retain the two reproductions as rejecting controls. This bounded review leaves production code untouched rather than claiming such a lifecycle change is qualified by loader doubles alone.
Independently examined successful properties and limits
- The canonical closure enumerates every top-level
.wat/.watxsource and supplies three VFS aliases.compile-srctransforms all three aliases of13-exports.wat. The producer inventories that whole source set, including unused fragments, and seals the entrypoint and adapters. Exact UTF-16LEextraWatand boolean tail-call options are distinguished. Independent matching, changed-extraWat and changed-tailCalls controls passed. tools/watx.jsevaluates precisely four local VM stages: parser, stages, codegen and compiler. All four are covered by the recursivewatx-srcinventory, as are loader and closure adapter. The VM receives builtins, not filesystem/require access. Synchronous production compilation uses production mode, standard WAT, runtimeBuiltins false, strictDeclarations true and the selected tail-call boolean. Our actual canonical adapter passed these options to the double. No additional disk dependency was found in this transitive path. Compiler identities (historical artifact:d3d-render-artifact-independent-review-20261023/compiler-audit.json) bind the inspected files to exact Git blobs and SHA-256. Node/builtin identity is not a field in this local receipt; it is not a portable build attestation.- Consume reads and hashes files without evaluating compiler code. A rejecting import hook confirmed this independently. No compile fallback exists. The producer's 34 protocol controls also passed under our cheapNode wrapper, including missing/stale/invalid input rejection before instantiation, default source-only invalidation, and prebuilt overload compatibility. Their closure/host doubles do not prove executable freshness of the real adapters; the new reproductions use the real closure adapter to examine that gap.
- Both receipt and module use open-descriptor type/size checks before
Buffer.alloc: 128 KiB and 8 MiB respectively. Module metadata is also bounded before the module read, and schema/options/hash mismatches reject. Independent descriptor-size doubles with an allocation sentinel confirmed both limits. No oversized staged input or module compilation was needed. Per-file source input reads are capped at 24 MiB; the inventory has no aggregate input cap, so future execution still needs external admission. - Helper preparation is synchronous and exits before its host imports, guest
memory, instantiation, or assertions. Both actual unchanged entrypoint files
ran through this branch with the real helper/closure/compile adapter and tiny
loader double. Top-level Device/fixture imports were intercepted; these runs
do not qualify those modules' side effects. Exact extraWat digests differed:
compaction
aedd14faff32e47f95292dd8b40bc0f39f13c04f97da96164e288d1521276a0b, arena reuseffbd12b13656c33399fb47914adb70f78f98584c4fd9fb4d81292ae508ffa2c1. Both exit-0 results are labelled PREPARATION ONLY, never native test passes. - Environment selection is process-wide. Prepare exits on the first harness;
multi-harness tests are unsupported by that command. Consume can service
repeated identical invocations but rejects differing extraWat/options instead
of selecting or compiling another artifact. Both modes reject prebuilt mixing;
empty/relative paths and conflicting modes fail closed.
process.exit(0)bypasses later assertions/cleanup and can truncate pending asynchronous work; the module writes themselves are synchronous. Preparation must be invoked and labelled separately from a suite runner that treats exit 0 as a passing test. - Prebuilt mode bypasses identity/compilation and retains the Module overload. Default mode now reads/hashes the complete source/compiler inventory on every call, even a cache hit: the base source inventory alone is 128 files and 20,706,777 bytes. This adds synchronous disk/CPU cost; source revisions also retain additional Wasm map entries. No full-source performance measurement was authorized. Moving host imports later changes error/side-effect ordering but preserves the ordinary construction path by inspection and protocol doubles.
- Descriptor growth checks and before/after identity comparisons do not prevent same-size rewrites, ABA edits, or cross-file snapshot races. An immutable tree remains necessary throughout an operation. Receipts are unsigned trusted-local integrity records; a caller able to rewrite receipt and module can forge them.
- The actual main classifier returns
unitfor the new regression; discovery matches its filename, andtest/run-all.shtakes unit/quick membership from that classifier. No test-tier omission was found. No broad tier/gate ran.
Evidence, resource bounds, publication, and remaining gates
Replay from this report's directory in its checkout:
timeout 60s python3 docs/re-notes/d3d-render-artifact-independent-review-20261023/run.py \
node --max-old-space-size=256 \
docs/re-notes/d3d-render-artifact-independent-review-20261023/independent.js
The adjacent sealed reviewed-inputs contains small exact candidate/main input
copies for replay after rejection; no executable was copied. Fixtures create
tiny temporary input files and always remove them. Node executes installed in
place. The full compiler is neither copied into fixtures nor evaluated. input-seals.json records copied inputs. The original independent harness had a
generated-regexp quoting error; its failed log/receipt remains, followed by the
corrected passing run. No guard failure was retried unchanged.