PLAN v1 disposition

Opus 5 returned REQUEST CHANGES at 2026-10-13T02:01:04.647084Z. Exact input: 130675 bytes, SHA-256 3b17ac27ecf44c2e211f1e87a14f0f914825b53aabac6f488862f9cc701e9ea4. Raw stdout JSON SHA-256: 77ffab3cddb12497fe1468751032b67b583f18671a28ef51843e855c0e93e400. Original input, invocation, completion, stdout/stderr and portable v1 capsule remain unchanged. This disposition is new evidence; it does not retroactively amend v1.

1. Accepted: missing Hemera cfg/API/backend analysis. The complete 18-site inventory is now retained (three std, fourteen architecture guards, one lint declaration). Hash::to_hex and OutputReader's Read implementation are actual conditional APIs; architecture selection has no std guard. Reachability and source scan limits are explicit in feature-audit.md.
2. Accepted: repository benchmark gate was omitted. Actual parent record 49 measures all nine reduction cases with the declared Criterion settings. Candidate uses the same settings; raw estimates/samples and a bounded paired-repeat policy are required. Algorithm identity is not a gate waiver.
3. Accepted: reverse consumers need feature inheritance, not only NeBu identity. The named exact manifests independently preserve Hemera defaults; isolated nox-cli/Trident-silicon have no affected direct API use. Product execution remains explicitly unverified, despite the manifest-level findings.
4. Accepted: durable no_std guard missing. V2 proposes a project-specific CLAUDE gate using real thumb cargo build, measured on the exact parent in record 38. No new script or CI dependency layout is introduced.
5. Accepted: ring/decision ambiguity. Computed ring is 1; product/Kelvin/tag controls are not green and the disposition is decision. The owner already authorized ordinary work/review/merge in this session. Root's task message accepts the dependency feature scope and this correction; no new user approval is invented. Versions, tags, publication and releases remain excluded.
6. Accepted: parent baselines were incomplete. V2 retains actual command baselines before any production edit, including their inherited and harness failures. Automatic API runs with a shared compiler target are not counted as isolated proof. Separate current/parent JSON builds succeeded; exact-tag --locked fails on its stale committed lock and remains unmet. This is not masked by a generated lock or a parent API comparison.

Advice adopted: thumb code generation, exact feature variants, separate source and package epochs, publication limits, and explicit current/tag API scope. Registry NeBu's older Dot code is a risk, not an automatic prediction of failure: the actual same four package-H264 controls pass in both debug and release (records 66/67). No broader H264 domain equivalence follows from those four controls.

The optional-honeycrisp advice's cache premise was stale: record 56 resolves acpu and fails in nox on absent acpu::field::ntt_forward and multilinear_eval symbols. This is now the precise inherited optional-feature boundary. All parent sources and locks remain unchanged.

Preparation corrections are visible: records 19–37 invoked Homebrew cargo with unsupported +toolchain syntax; 38–56 use the rustup shim. Record 46's Clippy mixed Homebrew driver with rustup rustc; 57 uses an explicit toolchain-bin PATH and exposes the real existing needless_borrow lint. Record 58 similarly pins rustfmt. The package-H264 builder initially assumed spaced manifest syntax and stopped before writing its manifest; corrected builder uses a new isolated directory. Those are harness preparation errors, not production findings. Portable v1 first had two Nu parse errors; both are documented with its original verifier retained, and actual portable replay passes after correction.