A zero nobody has ruled on is not a measurement

[ChelseaKR]

Closes #83.

The defect

Five DINS fields are FieldSpec(..., numeric=True) with no declared vocabulary, so each carries Basis.NONE and publishes marker_basis: "none". So does a field nobody has ever opened. docs/MARKERS.md section 7 named it in its own last sentence — "the gate that reads this file has no way to tell a reviewed decision from an unexamined one" — and the registry, the artifacts and both pages had no word for the difference.

Field	Recorded zeros	Share of what it records	Reading section 7 already carried
NOOFCARSONPROPERTY	55,831	99.4%	a count of no cars
NOOUTBUILDINGSDAMAGED	59,628	96.3%	a count of no damaged outbuildings
NOOUTBUILDINGSNOTDAMAGED	58,888	95.0%	a count of no undamaged outbuildings
NUMBEROFUNITPERSTRUCTURE	58,411	94.3%	contested
ASSESSEDIMPROVEDVALUE	6,613	5.2%	contested

What this does

FieldSpec.zero_reading is a second axis, not more Basis members. Basis answers where a declared vocabulary came from; ZeroReading answers what a zero is, and a field can be in any combination of the two. Four words, because a reader acts differently on each:

Word	What it says	What a reader should do with a 0
measurement	Somebody read the form and the distribution; a 0 is a finding	Use it as a count of none
undecidable	Somebody read them and the published documentation does not settle it. The zero is left present — declaring it a marker would move tens of thousands of records on a reading nothing supports	Do not use it as a finding, and do not use it as an absence
marker	Declared: 0 stands in for a value that was not recorded, so the zeros are counted as recorded-as-unknown and the field publishes none	There is no 0 to read; it is already counted as an absence
unreviewed	Nobody has ruled. A 0 would be published as a value and might be an absence	Nothing — and the build fails if a field in this state publishes one

The readings are section 7's own, transcribed rather than invented. Three counts are measurement, the two contested fields are undecidable — the third answer #83 asks for by name — and YEARBUILT is marker. Section 7 now states the word beside each row and tests/test_schema.py holds document and registry to each other by an exact token, in both directions. UNREVIEWED is the default, because a field added tomorrow has not been reviewed.

Two numbers everywhere, derived at render time from the rows beside them. recorded_zero_review in each artifact; the sentence above each field table on dins.html and perimeters.html; each field's own reading on its own row.

Coverage, before and after

surface	before	after
DINS fields measured as numbers, carrying a reading	0 of 8 — the state was not representable	6 of 8 (LATITUDE, LONGITUDE named as the rest; both publish no zero)
of those publishing a zero	0 of 5	5 of 5
FRAP	0 of 1	0 of 1, published rather than silent; GIS_ACRES publishes no zero
artifact per-field	zero count only	zero count and the reading
page	nothing	the pair above each table, the word on each row

Schema and consumers

artifact_schema_version 1 → 2. recorded_zero_review is a new required top-level key and recorded_zero_reading a new optional per-field one; a consumer validating against version 1 — whose schema closes with additionalProperties: false — would reject the new artifact. The rule for a bump is stated in src/perimeter/artifacts.py and in the CHANGELOG so neither can be edited alone.

site/ is regenerated from the pinned 2026-08-07 retrieval, and make site-check reports it byte-identical to a fresh build from data/raw/ (1182 and 81203 leaves compared, none differing).

Negative controls

Three sabotages, each with the occurrence count and git hash-object recorded before and after, each restored to a byte-identical tree.

sabotage	result
ZeroReading defaults to MEASUREMENT, so an unexamined field's zeros publish as reviewed	fired — 59 tests red
withdraw NOOFCARSONPROPERTY's ruling, so it publishes 55,831 zeros declaring nobody ruled on them	did not fire — all 534 tests in tests/test_schema.py green
count every examinable field as reviewed, so the block publishes 8 of 8	did not fire — all 1,158 tests green

Two of the three controls found a gate that could not fail, and both are fixed in this PR.

The second passed because every gate on the zero question reads the committed artifact, and the only thing that rebuilds it — make site-check — needs data/raw/ and cannot run in CI. So a ruling withdrawn in schema.py would have shipped green. spine() in tests/test_published_site_is_current.py already holds marker_basis between a fixture build and the committed copy; it now holds recorded_zero_reading beside it. Re-run after the fix: fired, dins-coverage.json shape mismatch.

The third passed because every gate on recorded_zero_review recomputes it from the same payload the block was written into — a summary checked against itself. zero_review now has two tests that read its own answer over a field set whose two numbers are known to differ, one of them specifically that a field with no zeros today still counts as a question. Re-run after the fix: fired, 2 tests red.

Verified

pytest 1160 passed, 25 skipped (browser/jsdom gates skip without make node-sync); ruff check, ruff format --check, mypy --strict src all clean, each exit code captured off the bare command rather than through a pipe. make site-check byte-identical.

Prepared with AI assistance; reviewed before submission.