CR-2026-036: implement the metadata enrichment block as specified

[reissjason]

Paired with the spec-side MR on la-payload-schema (cr/2026-036-metadata-enrichment), which adds PS-309 to PS-325 for a construct that had sat in schemas/payload-schema.json since the initial release with no clause defining it.

The count was wrong, and the reason matters

I reported "five undocumented meta-schema keys". The real answer is a whole block of twelve: metadata.include (name, source) and metadata.timestamps (name, mode, field, offset_field, elapsed_field, time_base, source, format), plus encode_formula and input_metadata.

The five I found were the five whose names aren't ordinary words. mode, source, field, name and format all appear in the spec for unrelated reasons, so a name-match reported "documented" for keys nothing described. That's now a rule in AGENTS.md: check whether the construct is described, not whether its key names occur.

Specified as OPTIONAL

Only the Python reference implements any of it — Go, Java, C#, the TS013 generator and C have nothing, and no corpus schema used it. Requiring it would make five implementations non-conformant on a feature nothing exercises. It's also optional in kind: the block contributes no bytes, reads no payload, and runs after decoding, so an implementation ignoring it decodes every field identically.

Three defects, all silent, found by probing to describe it

	Was	Now
include named for a decoded field	overwrote it — a u16 decoding to 60 came back as an ISO timestamp string	PS-313: decoded field wins, warns
mode: rx_time, no recvTime	measured_at: null — indistinguishable from a silent device	PS-314: key omitted, warns
malformed recvTime	4 swallowed exceptions, no value and no warning	names the key and why

_enrich_metadata is rewritten around one place() applying both rules, so a fifth mode can't reintroduce either by forgetting to check.

Making the requirements reachable

• validate_schema.py validates the block for the first time (PS-315–317). mode: elapsed_to_absolut was previously accepted and silently derived nothing; 8 malformed shapes are newly rejected.
• vector-verdicts.py passes a vector's input_metadata to decode (PS-324). Without it no vector could reach the block on any path — the CR-2026-014 lesson applied up front instead of a CR later.

Why the fixture asserts only the payload

metadata-enrichment.yaml carries a metadata block but lists only its two decoded fields in expected. Asserting the OPTIONAL enriched keys in the shared corpus would fail the four implementations that correctly don't produce them. The fixture's job is PS-310/PS-311 — a schema carrying the block decodes its payload identically everywhere — and all five demonstrate that, C included. Enrichment semantics are asserted in the CR test against the one implementation that has them.

Verified

• Python 2819 passed / 4 skipped (was 2783; +36)
• Both conformance paths 1253/1253, 0 disagreements
• Go ok · Java 46 · C# 92
• C harness 491 of 491 attempted, 0 differ — up from 488 because it builds the new fixture too

🤖 Generated with Claude Code