feat(conformance): close 206 uncovered constraints; coverage 18.8% → 40.7%

[fas89]

Follow-up to #44, which shipped 269 cases at 16.8% mutation coverage. Case count is not coverage: the question that matters is whether deleting a constraint from the schema turns anything red, and for most of the schema the answer was no. exposePolicy and semanticModel — both governance-critical — had no cases at all.

Seven new groups, written against measured gaps

Each agent was given the exact list of constraints that nothing pinned, not a vague brief.

Group	Cases	Targets closed
expose-policy	14	13 / 13
semantic-model	22	21 / 21
orchestration	47	51 / 56
acquisition	71	69 / 69
lineage-ml	27	26 / 26
binding-governance	18	17 / 17
expose-mcp	11	9 / 9

206 of 211 closed. The five left open are airflowTask allOf branches that are genuinely unreachable — allOf[0]'s then carries only a description, so no document can make it fail. They are reported as unreachable rather than papered over with a case that proves nothing.

The number

18.8% → 40.7% (192 → 415 of 1020 falsifiable constraints).

Better framed by what kind of constraint: the ones that carry real meaning — enums, required members, closed objects, bounds — went from 248 unpinned to 38. What remains is dominated by type deletions (567), which need a wrong-typed value for every property in the schema and are individually low-value.

The metric got more honest, too

mutation_coverage.py now distinguishes uncovered from redundant. A constraint sitting beside an enum that already restricts the instance to values satisfying it is unfalsifiable — no case could ever catch its removal, so counting it as a gap blames the corpus for something it cannot fix.

fluidVersion is the clearest example: it carries pattern, type: string and an enum of three version strings. Two of the three are dead weight. 120 such constraints are now excluded from the denominator and reported as findings about the schema instead.

envelope.json also gains the four missing required-member cases (kind, id, name, exposes) plus an empty-name case. The mutation report caught that the original corpus tested only two of the six required members going missing — a gap in the file I wrote by hand, found by the tool.

How it was checked

Every group was written by one agent and then adversarially reviewed by a second that independently re-measured the coverage claim rather than trusting it — checking that every invalid case raises exactly one error, and that reversing only the declared defect makes the document valid. All seven came back sound with the author's claim confirmed. (That step exists because in the first wave a verifier caught an author reporting "18 passed, exit 0" for a file that actually failed.)

Tested

Clean venv built from scratch, Python 3.14, jsonschema[format] only — not the dev environment:

corpus                 exit=0   480 cases, 480 passed
meta-test              exit=0   19/19 checks
compat 0.7.1 -> 0.7.2  exit=0
compat 0.7.2 -> 0.7.3  exit=0
compat 0.7.3 -> 0.7.4  exit=0
compat 0.7.4 -> 0.7.5  exit=0
reference impl         478 agree, 0 disagree, 2 optional not implemented

Notes

• Additive: seven new files, plus envelope.json, mutation_coverage.py and tests/README.md. No schema changes, so schema-sync is unaffected; link-check is path-filtered to docs/**.
• No local paths, credentials or real email addresses in the 211 newly published vectors.
• Still worth doing separately: required_status_checks on main is empty, so these gates are advisory at merge time.

🤖 Generated with Claude Code