Hemera feature-boundary execution

Approved input: PLAN v2 1f2028db98726cea04732856de00b2b058501b930112df0811f09c83c1ac60d3. Fresh Opus 5 APPROVE completed 2026-10-13T02:23:59.183394Z; raw stdout 1bd21c68c348163df8bd58a0205772732e50dfc733d30374aa61cdf88d09c223. Invocation/completion/full raw are retained in audit review. Root read/accepted this exact plan and authorized the two-file implementation.

Historical PLAN inputs remain unchanged. All parent commands completed before the production edit, including extra records 68–71: normalized-package thumb build fails on Hemera/std, while parallel/brakedown/wgpu inverse feature trees succeed. This supplements the 67 frozen PLAN records, not a retroactive PLAN measurement.

Adopted advice: side-by-side default/std trees plus existing variants, durable PR names for the two std-only Hemera APIs, a named exact-tag gate follow-up, and the verbatim delivery boundary: Kadek still resolves unchanged registry nox 0.2.0. The optional install hint is omitted to keep the approved gate delta exact; the gate already states its target prerequisite. No toolchain is installed.

The review's approval concerns this source repair only. Consumer findings are manifest/source evidence, not full product builds; dependency std API behavior is outside cargo-semver-checks' model. Existing user authorization covers ordinary source merge with decision disposition, never publication/version/tag/release. Candidate results are pending at this execution note's initial creation.

Measured candidate and retained reds

The candidate exactly matches the two approved proposed files; 48 owning Rust files, all sibling source and all source locks remain byte-identical. Candidate exports use /tmp/nox-no-std-candidate-inputs; parent and candidate compiler roots are distinct, recorded per command. Records 72–101 are the scoped candidate gates; 102/103 are the single predeclared paired benchmark repeat.

• Actual no_std code generation passes in both sibling-source (72) and normalized package (96) epochs; their exact parent builds fail at Hemera extern crate std (38/68). Default inverse tree (90) has no Hemera std feature; explicit std (91) and parallel/brakedown/wgpu (92–94) do. No dependency was removed or substituted.
• Workspace debug/release and std each pass 161 owning tests (74–76); parallel/brakedown each pass 164 (77/78). Nightly-2025-11-26 invocation82 succeeds with zero doctests; it is not extra behavioral coverage. CLI build/test and wgpu check pass, with zero CLI tests. Existing four H264 controls pass debug/release against both sibling and package dependency domains (84/85 and98/99); package library tests97 pass161. All counts name their raw record and exact frozen source.
• Current-parent default-surface API comparison101 passes229 checks, skips31, consuming the two exact JSON files recorded by63/100 from distinct compiler targets. No claim is made about dependency feature API inheritance by that tool. Exact-tag gate64 remains unmet; follow-up owns the explicit blocks edge. Automatic diagnostics60/61 remain unaccepted shared-target evidence, not substituted for the valid JSON comparison.
• Required inherited reds remain: fmt79 equals parent58 after only path-prefix normalization; strict docs81 has the same error headings as47; honeycrisp89 has the same missing-acpu-symbol errors as56. Clippy80 and57 have the same needless_borrow source diagnostic at cli/main.rs:247:52, but Cargo's final wrapper differs (bin nox test versus bin nox); full stream identity is not claimed. Parent Trident59 stops at cached clap4.5.57; no candidate product build is claimed.
• Performance is red/unverified, not inherited-green and not explained away as noise. Initial comparison49/83 has five cases crossing the plan's threshold; one paired repeat102/103 has six. Repeat medians include tree+89.64%, particle +43.12%, hash-chain+46.90%, hash+5.28%, inv+21.08%, look+15.82%; add and mul reverse direction versus the first pair. Raw estimates, samples and both comparisons are retained. The shared host was not a controlled benchmark lab. Root paused its heavy Cargo work during the repeat and this agent ran no concurrent compilation; other system activity/frequency was not measured. Cause is unestablished. No further repeat or source tuning is hidden. Post-repeat binaries have equal 3028144-byte size but different full-file SHA-256; this does not establish emitted code equality or identify a performance cause.

Ring1/decision remains necessary. No whole-product, clean release, speedup or registry-delivery claim follows. The PR must state the two affected dependency APIs (Hash::to_hex, Read for OutputReader) and the future graph/feature-aware release floor. Kadek still resolves unchanged registry nox 0.2.0.