nx_medrec.nx

nx_medrec.nx -- MEDIA-RECOVERY ADMISSION CORE: decide whether an archived capture is REALLY the media, or the host's tombstone wearing a valid image header. module: nishi-core.archive.medrec depends: nx_cdx_parse.nx (rows/fields -- NOT re-parsed here, Rule 15) capability: PURE DECISION CORE (zero I/O) WHY THIS EXISTS -- MEASURED LIVE 2026-08-06 against the Wayback CDX index, never assumed (Rule 4): nx_media_xrecover admits a recovered capture on is_image(): GIF/JPEG/PNG/BMP magic bytes. That is a STRUCTURE test and it CANNOT fail on a host placeholder, because a placeholder IS a structurally valid JPEG. Measured: i.imgur.com "image not found" -> digest VXEDWGPHSNERWHDOUD6YWRWNT4ZOLEX4, status 200, image/jpeg, sizes 335 / 339 / 457 / 789 / 827 B -- the SAME digest across captures 2013-04-20 .. 2026-06-17. Real media recovered from that same host -> 6PYVXUDUV5WT3VSTAK33KZYTSI4BA56Z at 1,090,473 B and INAME2WHG3WPFAVNJ3URHYKXTB6FH46R at 937,110 B -- each digest UNIQUE to its image. So a pipeline gated only on magic bytes reports SUCCESS on an unbounded number of identical ~400-byte "not found" images, and its success count RISES as the host deletes MORE. *A ROW COUNT IS NOT EVIDENCE. Structure is not identity. *TWO SIGNALS, BOTH READ FROM THE CDX ROW ITSELF -- a tombstone is rejected for ZERO bandwidth: (1) DIGEST IDENTITY -- a known placeholder digest is never the media. (2) SIZE FLOOR -- supplied by the CALLER, never baked (Rule 11). *AND THE GENERAL FORM, which needs no list -- but ONLY as a CONJUNCTION, and the reason matters. A placeholder is byte-identical every time it is served, so its captures collapse to ONE digest at a SMALL size. Digest-collapse ALONE proves nothing: measured, the real 1 MB image OT6mUiy.jpg also carries one digest (6PYVXUD..) across all 3 of its captures -- an unchanged file always does. What separates them is SIZE, and note the CDX length column is the COMPRESSED WARC RECORD size, not the payload (OT6mUiy reads 1,090,473 / 1,083,440 / 1,083,439 for bit-identical content), so it is a coarse floor and never an equality test. mr_rowset_is_tombstone therefore requires small AND collapsed, and answers 0 whenever either leg fails. Build intelligence, do not maintain a blocklist (Rule 25) -- but do not let a heuristic claim evidence it does not have. *FAIL-CLOSED: an absent or unparseable field is MR_UNKNOWN, never ADMIT. A CDX length of "-" (revisit records carry it) is UNKNOWN, not zero. "We cannot tell" and "it is the media" are different answers. license_tier: ORIGINAL LIB.