test(conformance): adequacy criteria for a vector set, applied to every set here

[lywinged]

A conformance vector set is a claim: a verifier that does not implement these rules
will fail this set. Nothing in the repository checked that claim, and it is not implied
by the vectors passing.

#169 and #170 both came out of asking it. Both were sets that were green.

• test(provenance): a depth vector set must also separate how far the list is read #169: build-provenance-depth reported the same verdict at every depth, so a
  verifier that resolved one link and assumed the rest matched passed it.
• test(canonicalization): a second vector for the form a careful implementer reaches #170: canonicalization-boundary separated its closest form by a single vector,
  and that vector put the divergence at the top level, where the obvious implementation
  already sorts.

Neither was found by reading a vector. Both were found by asking what a set fails to
distinguish.

The criteria

Four, each from a defect on a real set rather than from first principles:

Criterion	The defect behind it
A set must fail both unconditional implementations	a set of all-rejections is passed by a verifier that rejects everything
Every boundary needs more than one vector	one vector cannot separate a check that reads the head of a list from one that reads all of it
Every set on disk is measured, or named with the test that measures it	a hand-maintained list of what gets graded silently stops being complete
Shortfalls are recorded exactly	a gap that is skipped can widen; a gap asserted to its exact extent cannot

What it says about this repository today

Applied to every set, by the same loader, with the results recorded where they fall.

build-provenance-depth      margin at every boundary, nothing recorded
canonicalization-boundary   one-directional: every vector expects acceptance
action-receipts             covered by tests/test_vector_completeness.py

canonicalization-boundary is mine. Every vector in it expects acceptance, so the
set cannot tell a conformant verifier from one that accepts unconditionally. That second
implementation is a real failure mode, not a hypothetical, so it is a gap and not a
design. It is recorded in KNOWN_ONE_DIRECTIONAL with the record asserted, so it cannot
widen quietly, and the entry is deleted when the set gains a record signed over a non-JCS
form that a conformant verifier must reject.

I mention this first because a standard that only ever measures other people's work is
advocacy. The only shortfall this currently records is on a set I wrote.

The guard on the instrument itself

SETS is a hand-maintained list of what gets graded, which is the third criterion's
defect, sitting in the one place it would otherwise be invisible: a set added later would
simply not be graded, and nothing would fail to say so.

test_every_vector_set_on_disk_is_measured_somewhere compares SETS and
MEASURED_ELSEWHERE against what is actually in examples/, in both directions, since a
stale entry hides a renamed set as effectively as a missing one.

Rather than assert these fail for the right reason, I broke each and recorded which test
went red:

Mutation	Test that caught it
an unlisted set directory appears	test_every_vector_set_on_disk_is_measured_somewhere
an entry names a set that is not on disk	the same
the directory scan stops finding anything	the same
MEASURED_ELSEWHERE names a test file that does not exist	test_the_test_named_as_measuring_a_set_elsewhere_exists

What this costs, said before it happens rather than after

A future PR that adds a thin vector set will go red here, and the failure will name
the boundary and the vector rather than a line number. That is the intended behaviour and
it is a policy consequence of merging a test, so it should be a decision rather than a
surprise. The escape hatch is deliberate and visible: record the shortfall in
KNOWN_THIN with its exact extent, which documents the gap instead of hiding it, and the
entry is deleted when someone writes the second vector.

Scope

Tests and an informative document. No normative text, no RFC 2119 keyword, no schema
change, so no sponsor is needed under GOVERNANCE.md. Same class as #169, #171 and #175.

485 passed, 1 skipped; ruff check src tests scripts clean; mypy clean. Branched off
main at 697e20a.

The criteria are also usable outside this repository, since they read expected and
nothing else. If they are more useful in trace-tests than here, say so and I will move
them.

[github-actions]

❔ Contributor Check: UNKNOWN

Check	Result
Profile	UNKNOWN
Credential	LOW
Overall	UNKNOWN

Automated check by AgenTrust Contributor Check.