ARCADIA-Studio/harness/experiments/default-policy-20260913/policy-followup-review.md

Source-7 omission policy: narrow follow-up review

Independent reviewer /root/default_policy_critic, 2026-09-13. This supplements policy-review.md; it does not replace the original review or expand production scope. Reviewed only the two appended tests, unchanged prototype, and captured policy-2 evidence. No Cargo, Clippy, Gitea, or compiled-input edits were performed by this reviewer.

Result

The original nested unknown-field/constructor oracle gap is resolved for the requested representative paths. No blocking defect or new oracle overconstraint found.

• nested_future_fields_prevent_whole_default_removal independently specifies three expected maps: default String with future=[], default Exchange with future=false, and default view_options with future=null. Each verifies known default removal while requiring the unknown field and its containing composite to survive. String also retains its required minimum_length. The existing compact helper additionally checks idempotence and input immutability.
• unknown_and_multikey_constructor_wrappers_are_not_reinterpreted verifies unknown and ambiguous wrappers for DataType and Occurrence remain intact. Occurrence also exercises CompleteForEdit, proving its restoration step does not invent operation=null under an unrecognized or multi-key wrapper. Requiring intact wrappers follows the bounded helper contract; these sentinels do not claim current parser admission or future-schema semantic compatibility.

Captured execution evidence inspected

Read policy-2/started.json, result.json, stdout and stderr. The recorded Windows locked/offline focused command ran at HEAD 136917b47e50f7c07da975b93f879493cab81e21 plus captured additions, 10:14:58-10:15:02 UTC. Stdout names both new tests as passing and reports 18 passed, 0 failed, 0 ignored; receipt records exit 0 and 608 inputs unchanged. I independently matched SHA-256 hashes of the manifest, source ZIP, stdout and stderr to that receipt, and hashed both reviewed ZIP source entries against their manifest entries and current files. I did not rerun the tests or independently replay the full 608-input verifier.

• Prototype SHA-256: cf22a23d84e9abe949d38bd2f8771d9c677149006d8c436084b154d8eaa1d460, identical to policy-1 manifest and the original reviewed prototype.
• Final test SHA-256: 39fce98786ca5c3cc56ae8e1dea7499739a3290f8f1c9b15eeefeadfc22cd113.

The concurrent policy-clippy-2 outcome is outside this follow-up; parent owns its closure and verification.

Unchanged limits

Existing nonempty operation and port_anchors preservation under CompleteForEdit, and CompleteForEdit idempotence, still lack dedicated tests. Static entry(...).or_insert(...) behavior remains sound; this is a nonblocking coverage limitation. These map tests still do not exercise actual serializer extraction or parser re-admission. Public compact formatting, bundle exports, compact edit fragments, source-version handling, real semantic baselines, and the three known authored-default deletion regressions require their separate integration gates. No source8, GUI, full-suite, or release-completion claim follows from this result.