Cycle-close architect review of 16e31fe..HEAD against docs/design/INDEX.md +
CLAUDE.md. Regression gate: none configured (commands.regression: []), so the
architect is the sole gate.
What holds:
- C18 honoured — aura-registry stores (manifest, metrics) RunReports append-only
to runs/runs.jsonl, reproducible-from-manifest; no git/Gitea duplication, no
multi-project manager. The depth deferrals (lineage, re-derive, run-diff,
run_id, Aura.toml runs-dir) are named in the spec, not silently dropped.
- C1 preserved — the sweep stays disjoint/lock-free; "sweep == N independent
runs" now compares full RunReports; serde_json output is deterministic.
- C12/C19/C20 fit — SweepPoint.report closes cycle C's deferred manifest-per-
point gap; the manifest stays World-supplied (engine `sweep` is `Fn -> RunReport`,
domain-free).
- Workspace green (cli_run 11, registry 5, engine 93, …); clippy clean.
Resolution:
- FIX (this commit): docs/design/INDEX.md "External components" (the data-server
row) still asserted the struck "external-dependency firewall / engine crates
stay external-dependency-free" framing — stale after the C16 amendment (aura-
core/aura-engine now link serde). Rewritten to the amended per-case policy.
The iter-1 commit's "no ledger text still asserts it" missed this mirror; it is
now true (re-grep clean across the ledger and live source).
- FOLLOW-UP (#54, idea): RunReport has two divergent JSON encoders — serde for
the on-disk store, hand-rolled to_json for stdout. This was a deliberate,
spec-named deferral (keep stdout goldens stable this cycle); filed for the
stdout→serde unification that retires the last hand-rolled writer.
Ratify: the C16 amendment (blanket zero-dependency -> considered per-case policy)
is the intentional contract change of this cycle, justified in the 0029 spec and
the iter-1 commit; this audit ratifies it as drift-clean (the codebase no longer
contradicts it anywhere).
closes#33
Brummel
654b23ad1f