Skip to content

Research-0658: Project modernization audit

  • Status: Active
  • Workstream: ADR-0658
  • Last updated: 2026-05-20

Question

How can the fork keep finding real backlog, scaffold, stub, limitation, and modernization work after many one-off sweeps have already landed, without turning stale prose or model-artifact blockers into noisy PR churn?

Sources

  • .workingdir2/OPEN.md and .workingdir2/BACKLOG.md — current local gitignored planning state.
  • docs/state.md — durable project status, deferred rows, and recently closed implementation history.
  • docs/ai/signal-mix-audit.md and ADR-0650 — precedent for advisory, table-side audit tooling.
  • scripts/AGENTS.md and scripts/lib/AGENTS.md — read-only local state and stdlib-only helper expectations.

Findings

  • Pure rg searches still work, but they mix very different things: actionable NotImplementedError / ENOSYS code, stale docs, accepted ADR history, smoke-only fixture rows, and real blockers such as external model weights or dataset access.
  • The current project state is split between committed docs and ignored local files. Any audit that ignores .workingdir2 loses the merge-train context; any audit that rewrites .workingdir2 risks deleting editorial judgement.
  • AI modernization is not only individual TODO markers. Large one-off families such as train_*, export_*, eval_*, and extract_* scripts are useful cluster signals for future manifest/config consolidation.
  • Model registry smoke rows are a separate class of debt. They should be visible in reports, but placeholder fixtures are often intentionally blocked until a real model or redistribution decision exists.

Alternatives explored

Ad hoc search won on speed but lost on continuity. It is useful while working inside one file, but it does not preserve prioritisation across sessions.

CI gating was rejected because the text markers intentionally appear in docs, historical ADRs, disabled-build contracts, and model cards. A failing required gate would incentivise wording games instead of useful code work.

Automatic backlog rewriting was rejected because .workingdir2 is the human-edited state of record. The audit can produce JSON and Markdown, but the operator decides what belongs in OPEN.md or BACKLOG.md.

Open questions

  • Whether future reports should learn explicit suppressions for intentional contracts such as disabled optional-backend -ENOSYS.
  • Whether the script-family cluster output should grow into a dedicated AI manifest/schema consolidation PR.
  • Whether the model-registry smoke rows should be split into "permanent smoke fixture" and "promotion debt" fields in the registry schema.