CR-2026-034: flagged in the C interpreter, and in the harness that measures it

[reissjason]

The construct #32 named as the largest remaining gap. Attempted vectors went 453 → 488, all passing, and flagged is no longer a reason a corpus schema cannot be reached.

The representation

Reuses what match and tlv already have: match_var holds the name of the field carrying the mask, and each group is a case_def_t whose match_value is its bit position and whose body is placed above field_count — the invariant #32 established, for the same reason.

The mask field must declare var_name. The interpreter records a value in its variable table only where a field declares one, while a YAML flagged refers to a field by name — so no corpus schema carries var: for it, and the harness patches it in when building.

A builder that forgets gets SCHEMA_ERR_MATCH, not a silent decode of nothing: var_get returns 0 for a miss, which for a bitmask is indistinguishable from "no bits set". A new var_has() tells the two apart. The reference interpreter raises in the same case ("Flagged field reference not found"), which is what this matches.

Coverage

Six of the 34 flagged schemas build now; the rest are blocked by transform (26), bitfield_string (24) or computed fields — not by flagged. There's a test asserting a flagged schema is genuinely built and its var_name patched, because "0 differ" proves nothing if no group ever fires — I checked one decodes with flags = 0x0003 and both bits firing.

A message that misattributed

no constructor for type 'u32le16' read as a C gap when C has FIELD_TYPE_U32LE16 and decodes it correctly — it's a harness limitation. The report now says which side each limit is on, the same care the tlv and flagged reasons already took. That distinction is load-bearing in a report whose whole purpose is telling C's gaps from the harness's.

repeat is now the last construct C has no field type for, and at 3 schemas it's worth less than widening the harness for transform and bitfield_string.

Slips, all mine, all in scaffolding

• CR-2026-033: tlv in the C interpreter, and in the harness that measures it #32's test anchored on "FIELD_TYPE_TLV\n} field_type_t;" — pinning TLV as the last enum member, incidental and broken the moment this appended FIELD_TYPE_FLAGGED. Compares positions now.
• One assertion here read a doc comment through a byte-count window that caught its tail and missed its opening, and searched for a phrase the wrapped comment breaks across a line. Failed twice on a header saying exactly what it asserts. Anchored on the comment block with whitespace normalised.
• Twice while editing AGENTS.md I wrote a multi-edit script that aborted on a bad anchor and silently wrote nothing — the second time my --amend then committed unchanged content. Now I verify every anchor before writing any.

Verification

Suite	Result
Python	2706 passed, 4 skipped
make test-c	3 C binaries pass; harness 488/488, 0 differ
make selftest	ALL 3 SELFTESTS PASSED
vector-verdicts	1250 vectors, both paths 100%, 0 disagreements
encode-round-trip	1163 of 1239, 0 unexplained
validate-devices / score-check / docs-index-check	pass

Only the C header and the harness changed, so the other four implementations are untouched.

🤖 Generated with Claude Code