ARCADIA-Studio/harness/experiments/default-fidelity-20260913/fidelity-review.md

Default-omission fidelity review (2026-09-13)

Scope: local-only independent static review on codex/text-defaults-9e8r6t, base 6f2aae0f76708e0d4ace1cd66ea019c07d3299df, of the new test-only harness/prototypes/source-defaults-render-probe-20260913.rs. Parent owns runtime capture, integration, and Gitea coordination. No production files, tests, Cargo settings, or shared records changed by this review.

Adapter result

No blocking correctness finding in the inspected adapter. Its typed declaration extraction matches language.rs::format_declarations for elements, relations, diagrams/nodes, artifacts and traces. The original renderer/parser establishes admitted canonical syntax before the adapter operates. The adapter then retains original bytes and deletes whole field spans selected by the isolated source-default policy; it does not recreate references, enums, numerical spellings, string escapes, or historical values. Recursive subset checks reject value replacement, array shortening, added fields, and changed zero signs. Every typed identity must be visited exactly once, every retained field must have rendered syntax, every ordinary declaration must be consumed completely, and deletions must be nonoverlapping.

The small syntax-shape reader matches the actual canonical writer's records, arrays, scalar values, live @alias references, and historical uuid literals. Unexpected syntax/shape fails an assertion instead of being interpreted or normalized. Historical records are traversed for matching spans, but this policy does not select their fields for deletion. Whole project-description omission is a separate explicit empty-string rule consistent with the parser's project default.

Oracle boundaries

• Project equality alone does not distinguish floating-point signed zero. Independent tests must check bit patterns or exact persistent serialization; the adapter's subset assertion protects policy comparison but cannot detect an existing canonical renderer/parser zero-sign defect.
• Round-trip success alone would also pass insufficient pruning. Reuse the independent literal field-policy matrix and include representative output-omission assertions; do not describe fidelity evidence as exhaustive pruning acceptance.
• This bridge begins with production canonical output. It cannot establish authored-source preservation, new-fragment generation, real modular export formatting, or production CompleteForEdit update wiring. Existing issue 11 source-erasure regressions and source8 issue 7 integration remain separate gates.

Independent fidelity test review is pending the worker's frozen file. No runtime result is asserted here.

Frozen test review

Reviewed crates/oa-core/tests/default_omission_fidelity.rs at SHA-256 09abc1bfab900e6f76ea8e0e09ba8f6733f01f38e99b7d7283c9240f3ec5b920; reviewed adapter SHA-256 25cf757c7c7a2e1703254e94a5d1ef140f0b33d35e0d50278170610ab6d3d213. These are pre-parent-format local files, not an integrated commit. No blocking correctness or oracle finding.

The nine tests use five unchanged fixture corpora and four bounded typed mutations. Fixture preconditions independently check the model features being claimed before applying the probe. Every run compares the complete typed Project after ordinary parse and public source-project compile, retains the exact one-file input bundle after compilation, and checks canonical/compact repeat stability plus nontrivial actual deletion. Expected Projects come from unmodified public parsing followed only by explicit witness mutation; they are not generated by the policy under test.

Explicit bit checks address the signed-zero limitation for x=-0/y=+0 in each canonical, compact, and bundle decode. Independent typed assertions preserve unbounded DataField maximum, present-empty collection default, present-empty optional artifact strings, and selected versus unresolved present OperationUse. Corpus checks cover OA/SA, legacy migration, nonempty historical Binding/Annotation facets and identity atoms, structured records/scenarios, state-machine actions/projection, operation ports, and a dense forty-exchange model. The canonical and compact strings are built without editing fixture files.

The >=10 field-removal assertion and size check demonstrate activity, not completeness of every omission path. Existing independent literal policy tests remain the correct bounded completeness evidence. Likewise the one-file compile_bundle calls validate source-project acceptance and stability, not generated multi-file format_bundle output. These tests do not assert complete engineering validation, all numeric paths/constructors, full-suite acceptance, or production source editor behavior. Those are scope limits rather than reasons to block this isolated fidelity experiment.

No tests were executed by this critic; parent owns the captured test/Clippy results and any failure investigation. The earlier pending-test sentence is superseded by this frozen review.