nx_pattern_emit15.nx

nx_pattern_emit15.nx -- PATTERN EMITTER: PNG_INGEST (shape 19 -- the COMPOSE-shape). ==================================================================================== DECLARED-SCAFFOLD (operator-authorized R3, X-AUT-NCF-001 / spec 2026-06-13-keystone- newcontrolflow-route-decision.md, which names "the 5-stage ingest compose" VERBATIM as the gap no existing data-driven shape emits). This emitter is HAND-WRITTEN -- it is the GENERIC compose primitive. It is DEBITED vs autonomy: its register row is tagged author=tutor (cl_register_dual), counted in A2 auth_all but NEVER in auth_emit. Tagging it author=emitter to inflate A2 would be the cheat and is REFUSED. BOOTSTRAP-PROVISIONAL: it stays a scaffold until >=2 distinct downstream organs ride it hands-off (X-AUT-NCF-BOOT). The first rider -- nx_store_ingest -- is authored BY nx_auto_builder hands-off from a DATA spec (author=emitter), NOT hand-written. WHY THIS IS A NEW SHAPE, NOT A VARIATION: every existing emitter authors ONE leaf function over ONE buffer (e.g. pe8 STRUCT_WALK authors <name>_locate(b,n,out) -- a straight-line field walk; NO loop over repeating records, NO calls into other organs, NO store side-effects, NO idempotency control flow). The ingest organ must (a) LOOP over ALL PNG chunks, (b) CALL _pe_pngchunk/_pe_pngtext per chunk, (c) CALL canon_encode+cid_of, (d) CALL ss_get (idempotency) then ss_begin/ss_add/ss_commit. That cross-organ composition + per-chunk loop is precisely the new control flow the transducer cannot synthesize. WHAT IT AUTHORS (GENERIC, data-driven from the spec param vector): <name>_ingest(buf: *u8, flen: i64, prefix: *u8) -> i64 Walks every PNG chunk in buf[8..flen) (skips the 8-byte signature), and for each chunk whose FourCC is in the spec's CHUNK-TYPE set, reads its keyword/value via the green _pe_pngtext_locate. If the keyword is one of the spec's GENREC keys, the (key, value) pair joins the canonical record. The assembled vector is canon_encode'd, the CID computed (cid_of), the record key built ("img:"+CID). ss_get probes for the key: FOUND -> return 0 (idempotent skip, rule 10, no dup); ABSENT -> ss_begin/ss_add(kind 1=put, key, canonbytes)/ss_commit -> return 1 (new record stored). Return: 1 = stored a NEW record; 0 = idempotent skip (already present); -1 = no GENREC fields found / malformed PNG (REFUSED, never a clipped store). DATA-DRIVEN (the spec is the config, rule 11): the GENREC key STRINGS to harvest, the chunk FourCC(s) to walk, the record-key prefix, and the segid are all carried in the param vector -- the scaffold is reusable for any image-metadata ingest, not a PNG one-off. THE KAT (emitter-computed, no-fake-green): the test builds a REAL multi-tEXt PNG in memory (one tEXt chunk per GENREC key, each with the EMITTER-COMPUTED crc32), ingests it to a temp store prefix, then asserts: (1) CID round-trip -- ss_get returns the stored canonical bytes and re-hashing them yields the SAME CID the emitter computed off-line;