[reissjason]

What this adds

examples/conformance/ — twelve hand-written Thing Descriptions, one binding construct each, plus tests/test_conformance_corpus.py, which fails if any construct the binding publishes has no example.

Why

Coverage of the vocabulary currently tracks vendor adoption, not the vocabulary. A lorav: term, a wire type or a layout is exercised only if some manufacturer happens to ship a schema that uses it. Terms that no vendor has adopted are published, documented, accepted by lorawan-form.schema.json — and have never been decoded once.

What the two existing corpora actually reach:

	curated examples/*.td.json	after this PR
lorav: terms	20 / 24	24 / 24
payload layouts	3 / 4	4 / 4
lorav:wireType values	5 / 32	30 / 32
lorav:endian values	1 / 2	2 / 2
lorav:presentWhen shapes	1 / 2	2 / 2
lorav:derived descriptors	0 / 5	5 / 5
lorav:byteLength forms	0 / 2	2 / 2

examples/devices/ reaches more, but it cannot play this role: it is generated, gitignored and pinned to a submodule, so nothing in it is reviewed and nothing in it is stable across a bump. Two things were covered in neither corpus at all — the ctv layout, and 19 of the 32 published wire types.

This mirrors, one layer up, the _language-conformance/ fixtures in device-payload-schema, which hold the four reference interpreters to the same behaviour for constructs no device schema uses — and which found two cross-language defects when they were added.

What it found

Building the corpus turned up two published constructs that do not work. Both are filed separately and are deliberately not papered over with a fixture:

• lorawan: lorav:wireType 'bool' does not advance the payload cursor, shifting every later field #103 — bool / xsd:boolean. The reference interpreter reads bool as a bit of the byte under the cursor and does not advance the cursor without consume. The binding has no term for bit or consume, and converter._WIRE_WIDTH gives bool a width of 1, so a second boolean re-reads the first one's byte and every later field is read one byte early.
• lorawan: a per-field lorav:endian override emits a be_/le_ type the interpreter rejects #104 — mixed lorav:endian in one Thing Description. converter._typed() emits a be_/le_ type prefix for a field disagreeing with the schema default; the interpreter has no such prefix and rejects the field. Across 167 generated and 7 curated Thing Descriptions, zero emit it, so the path had never run.

Both are recorded in UNCOVERED_WIRE_TYPES and examples/conformance/README.md, with a test asserting that gap list stays accurate — it fails if a listed type gains a fixture or leaves the schema.

How the fixtures are built

Every expected value in a *.vectors.json here was computed from the wire bytes by hand and is then held against the reference interpreter, rather than recorded from whatever the interpreter returned. Three of my hand-computed values disagreed with the interpreter on the first run; two of those disagreements were the defects above.

The fixtures are picked up by the existing suite through conftest.example_tds() / vector_files(), so each one is schema-validated, decoded and round-tripped like any other example. Discovery is an explicit union of examples/*.td.json and examples/conformance/*.td.json, never a recursive glob, so the gitignored generated catalog cannot steer the checked-in suite.

Incidental fixes

tests/snapshot.py and tests/test_decode.py both resolved a vectors file's "td" key against examples/ rather than against the vectors file's own directory. That only worked while every example sat in one flat directory; both now resolve beside the vectors file.

The golden snapshot is re-recorded: 19 payload schemas and 37 decoded vectors, with the conformance entries under a conformance/ prefix.

Verification

1270 passed  ->  1498 passed, 55 skipped
ruff check   ->  All checks passed
ruff format  ->  28 files already formatted

The guard was mutation-checked: deleting layout-ctv-default-tags.td.json and derived-compute-guard.td.json produces exactly

FAILED test_every_supported_layout_has_an_example[ctv]
FAILED test_every_derived_descriptor_has_an_example[compute]
FAILED test_every_derived_descriptor_has_an_example[guard]

🤖 Generated with Claude Code

[Hao-Siemens]

Thank you very much for the PR!

I ran the tests and they passes. In generall, I support adding these conformance tests alongside the real-device tests. Also future developers can treat this as labled examples. Though, we show binding-feature coverage, not real device compatibility here. My opinion is to prefer real device with derived vectors for all the labeled examples where available and use synthetic cases for specific gaps and edge conditions.

Beside, my AI agent found one mismatch: polynomial documentation (the form schema, ontology, and README describe coefficients in ascending order (c0 + c1x + c2x^2)) with the interpreter’s descending coefficient order. Please align this before merging.