Coldstart has not been released. This section describes verification and regeneration in the development build at the pinned revision; there is no release to install.
The export route has changed, the historical-ID work is queued, and the rejected adapter is gone. Now the project needs two different kinds of upkeep: check the rules the harness actually represents, and rebuild machine-made views from the authored facts they summarize.
Coldstart keeps those jobs separate. A green check is a receipt for a named run. A current generated map is a receipt that one view matches a fresh rendering of its sources. Neither means the whole project is correct.
The three pages
| # | Concept | What it answers |
|---|---|---|
| 7.1 | What green means | Which checks ran, under what conditions, and what the result supports |
| 7.2 | Re-deriving projections | How authored facts produce replaceable maps without creating a second source of truth |
| 7.3 | Sweeping and re-deriving | Which two entrances you actually type to sweep the harness and rebuild its generated views |
The ordinary alternative
For a small app, run the relevant tests and update a hand-maintained index when something moves. That is often the clearest choice. Friction appears when a growing harness can print green without showing what ran, or when a copied index slowly disagrees with the files it summarizes.
Coldstart makes both receipts inspectable.
Green belongs to one run
The verification receipt names:
- the registered check roster;
- which checks applied to this tree;
- whether the operator selected a smaller subset;
- the mode and environment; and
- exit
0,1, or2.
Exit 0 means the applicable checks completed without errors under those conditions. Exit 1
means a represented rule found a problem. Exit 2 means the runner or a check could not make an
honest judgment.
A broken checker must not look like a pass. A skipped check must not look like a passing check. The report matters as much as the color.
Generated views point one way
Warehouse maps and routing indexes help people or tools find authored material. Their facts still belong to the authored cards. A generator scans those sources, renders the complete expected view, and compares it with the bytes on disk.
If the bytes differ, check-only mode reports staleness without writing. Write mode replaces the generated view. If a source cannot be read, generation stops instead of publishing a partial view that merely looks complete.
The jobs compose without becoming one claim
| Receipt | Narrow conclusion | Unsupported conclusion |
|---|---|---|
Verification exit 0 | Applicable represented checks passed here | The project has no defects |
| Generated bytes match | The view matches this rendering of readable sources | The authored facts are correct or current |
Run both when the harness or its authored records changed. Use a smaller manual routine when the project has no generated views and only a few obvious checks.
People still decide whether the roster represents the right risks, whether a warning should become an error, and whether an authored fact is wise. Upkeep makes disagreement and broken machinery visible; it does not remove judgment.
Next
Start with the first receipt: what green means.