nx_jpegmem_gate.nx

nx_jpegmem_gate.nx -- THE JPEG DECODE MEMORY GATE: repeated decodes are FLAT-RSS and overrun-free. Subject: _offc/nx_jpegmem_probe.elf forked over a REAL banked donor texture (never a synthetic stand-in -- the asset class that exposed the defect IS the donor corpus). The probe decodes the same jpeg N times in one process, prints VmRSS after each round, and surfaces any ARENA-OVERRUN diagnostic from the allocator canary. WHY THIS GATE EXISTS (measured 2026-08-23, nx_jpegmem_probe): nx_jpeg_idct_8x8 and the zigzag inverse-table builders each allocated 512 B PER 8x8 BLOCK; 512 > NXA_SMALL_MAX routes past the bump arena to the page allocator, so every block cost a 4 KiB page nothing freed -- 3,203,232 kB RSS growth PER 4096x4096 decode (it OOM-killed a build VM). Separately, NX_JPEG_DEC_CTX_BYTES=112 under-sized the 15-field NxJpegDecCtx, so every decode wrote 8 B past its context allocation (canary: ARENA-OVERRUN prev_alloc_size=112). After the root fixes the probe measures ~50 kB/round, bounded by bump-arena chunk growth. FAIL DIRECTION IS PROVEN BY MUTATION, not by a synthetic neg-control tooth: nx_gate_bite with nx_jpeg_idct.nx as the mutated source and nx_jpegmem_probe as the rebuilt subject restores the per-block allocation and the flat-rss tooth MUST go RED (bite receipt in the ship record). NOT ROSTER-ADMITTED on purpose: a real decode takes seconds and the roster beat kills at ~1.8 s (the voxchunk precedent) -- run via the job lane / ship loop.