nx_record.nx

nx_record.nx -- NXR1: the TYPED, SELF-DESCRIBING record encoding for sovereign planes. WHY THIS EXISTS (debt seq1326). The TSV->plane migration moved the CONTAINER, not the ENCODING: nx_store_seed_lib stores each flat LINE as an opaque blob under a POSITIONAL key q:<seq> and is contractually required to rebuild the flat file byte-identically. So a "migrated" plane still holds tab-delimited untyped text. A plane-ified TSV is still a TSV. Its failure modes are structural, not sloppiness: a TAB inside a value corrupts the row; a record missing one column SHIFTS THE MEANING OF EVERY COLUMN AFTER IT (the live debt- plane already carries mixed 5-col and 7-col rows); there are no types, so every reader re-parses text and hopes; and there is no way to add a field without rewriting every existing record. NXR1 fixes all four AT THE REPRESENTATION, which is the only place they can be fixed: - FIELD IDS, not positions -> a missing field is simply absent. It cannot shift its neighbours. - LENGTH-PREFIXED values -> a tab, a newline, a NUL inside a value is just bytes. No escaping, no quoting, no delimiter to collide with. - EXPLICIT TYPES -> i64 is stored as i64, not as text a reader must trust. - SKIP-UNKNOWN + DEFAULTS -> a reader steps over field ids it does not know (forward compatible, rule 19 at the DATA layer), and an absent field reads as a caller default. That is Iceberg v3's default-value semantics: a new column costs zero backfill. LAYOUT (big-endian throughout, matching the seg-store's on-disk convention): record: ['N']['X']['R']['1'] [u16 nfields] then nfields x field field: [u16 field_id] [u8 type] [u32 len] [len bytes] This is a LEAF: it imports nx_syscalls ONLY. A record codec that had to pull in the storage engine just to read a u32 would be backwards -- the record layer sits UNDER the store, not beside it. license_tier: ORIGINAL No hw writes (Rule 26).