Let unsafe-relaxed JSON audits express scoped private trust boundaries

[Widthdom]

Summary

The unsafe-relaxed JSON audit reports a private local JSONL writer without distinguishing its controlled local trust boundary from external/untrusted parsing.

Audit IDs: R-20. Reproduced with locally built cdidx v1.40.3 at 404db09489c305f17d94f37026e00564d3ac527a.

Existing issue relationship

Closed #3965 is adjacent trust-boundary/audit recipe work; no exact issue covers this current writer context.

Expected behavior and PR scope

Allow an explicit, reviewable local/private trust scope in recipe evidence without making unsafe-relaxed parsing broadly disappear. Keep this to trust-origin/direction metadata and the recipe.

Implementation guide for Codex 5.6 sol xhigh

• Distinguish serialization/writing from deserialization/parsing and local/private sinks from external inputs.
• Prefer structured origin/sensitivity/trust annotations or configuration over path/name allowlists.
• Require explicit evidence for suppression and surface that reason in results.
• Add positive controls for external input parsing, public APIs, network/files, and ambiguous trust.
• Add focused positive/negative regression tests, run the full net8.0 and net9.0 suite, and add the required bilingual changelog.d/unreleased/ fragment.

Acceptance criteria

• The controlled private writer is classified with its trust rationale rather than generic high risk.
• Untrusted parsing remains a finding.
• Users can audit/override the trust decision explicitly.