nx_vecfetch.nx

nx_vecfetch.nx -- ACQUIRE EXTERNAL ANSWERS WITH THE AGENT OUT OF THE BYTE PATH. THE HOLE THIS CLOSES (measured 2026-07-31): every clause in the evidence apparatus verifies the COMPARISON and none verifies the PROVENANCE OF THE ANSWERS. An agent can write a gate against "RFC 8032 test vectors" by typing them from its own recall -- it compiles, runs GREEN, and yields a row with a real ref=, a real refdig= and a real execution. refdig= commits to the bytes we compared against; it CANNOT prove those bytes came from the IETF rather than from the model that wrote the file. ============================ v2, 2026-07-31: TWO DEFECTS FOUND AND FIXED ============================ ⚠DEFECT 1 -- I HASHED THE ENVELOPE, NOT THE DOCUMENT. v1 hashed whatever nx_https_get returned. By contract that is the ENTIRE HTTP RESPONSE: status line, headers, and chunked-transfer framing. So all 14 pinned "authority documents" were HTTP WIRE TRANSCRIPTS -- rfc2202.txt carried 707 bytes of Cloudflare headers and a `2ea6` chunk-size line before the RFC even began. nx_extvec_authentic_gate now reports 14/14 CONTAMINATED against the v1 corpus. Verified against a stack sharing no code (.NET WebClient): chunk size 0x2ea6 = 11942 = the true document length exactly. ★★★★★ THE LAW v1 GOT WRONG, CORRECTED: v1 said "hash the bytes before anything touches them". That is too crude -- it hashes the ENVELOPE. THE RULE IS: HASH THE OBJECT YOU ARE MAKING A CLAIM ABOUT, AT THE EARLIEST POINT IT EXISTS. The document does not exist until the transport framing is removed. Hashing before that point is not "hashing earlier and therefore safer" -- it is hashing A DIFFERENT OBJECT. The anti-tamper property v1 was protecting is PRESERVED: the decode below is a pure, deterministic, in-process transform. No agent, no second tool, and no disk sits between the socket and the digest. ⚠DEFECT 2 -- THE DECODER ALREADY EXISTED AND I DID NOT ADOPT IT. nx_fetch_unit.nx has had fu_body_off / fu_status / fu_find_ci / fu_dechunk, and a fully staged nx_fetch_staged, the whole time. v1 called the RAW primitive underneath them. ★THE ADOPTION GAP, NOT A MISSING PRIMITIVE -- the same class already in memory. The fix is to ADOPT the proven decoder, not to write a ninth private copy of one. (`nx_body_off` is currently duplicated across at least 9 files -- that Rule-15 extraction is filed separately; this organ deliberately adds no new copy.) ============================ CORROBORATION: THE JULY-2026 BAR ============================ A digest pin proves INTEGRITY-SINCE-FETCH. It cannot prove AUTHENTICITY-AT-SOURCE, because a corrupted or re-framed fetch pins exactly as cleanly as a good one -- the pin is self-consistent by construction. The state of the art does not solve this with a better hash; it solves it with ANOTHER PARTY: - CA/Browser Forum MPIC (mandatory for TLS/S-MIME issuance): validate from MULTIPLE INDEPENDENT NETWORK PERSPECTIVES, and if the perspectives DISAGREE, THE PROCESS HALTS. It does not pick one. - Sigstore/Rekor: a transparency log exists so verification need not trust a single party. So this organ takes an OPTIONAL third argument: a digest obtained by an INDEPENDENT STACK. If supplied