Skip to content

Generate reviewable evidence inventories instead of hardcoded counts #217

Description

@woksin

Problem

Several high-value V4 evidence gates intentionally fail closed on fixed counts: Storybook expects 64 story modules, 261 stories and 64 autodocs pages per renderer, while the package graph records 34 kernel modules. The counts caught drift during V4, but a legitimate addition requires editing verifier constants by hand and can make a stale number look like architectural policy.

Direction

Generate evidence inventories from authoritative inputs while retaining reviewable golden snapshots:

  • Storybook inventory from built indexes plus source story modules;
  • renderer set from validated adapter metadata;
  • kernel inventory from the declared kernel-boundary source;
  • parts/state counts from the generated parts manifest;
  • package subpaths from the export map.

CI should compare generated inventory bytes with a committed, reviewable snapshot. A change must fail until the snapshot and rationale are reviewed, but verifier source should not need an unrelated constant edit.

Acceptance criteria

  • No story, autodocs, renderer, kernel, parts/state, or export count is duplicated as an unexplained verifier constant.
  • Generated inventory is deterministic and checked into source or uploaded with an exact base/head diff.
  • Removing a story/renderer/kernel entry fails and names the missing identity, not only a count.
  • Legitimate additions require an explicit reviewed inventory update.
  • No sampling or silent truncation is introduced.

Metadata

Metadata

Assignees

No one assigned

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions