vgsp: CONT-99 - sema_gl_mme_prelude_ir_only + sema_gl_mme_prelude_pack (partitioned IR: each SA region its own upload run clipped at next SA entry; SA pairs + one contiguous 0x114 ONE_INC run ~1270 words so SA+IR+calls ship in same <=2048-word block; static gsp_phys_gr_mme_ir_pack[4096]) - decisive offline: capture is sum of OVERLAPPING per-macro upload runs; run 34 (ptr=0x265, 75 words) == driver mmeCodeOGL_C597_DRAW_IAB word-for-word (0/75 diffs) but later runs overwrite its tail so capture-order image differs at 50/75 words; r1343/44 packed same-launch: MME-PRELUDE pack ir_words=937 runs=55 sa=54, held launch retires words=1515, first call-bearing launch words=1633 STILL POISONS in 8.7ms; unified failure 6 boots: MME-FAST end events=1 (poison) + ILLEGAL_OPCODE + ESR 0x404490=0x80000004 (bit2 NV_PGRAPH_PRI_MME_HWW_ESR, dev_graphics_nobundle.h:2046); r1340 nop_drop {48,51,57} 40x drops still faults {0,2,34,39,40}; r1341 block_cap=4096 words=3441/3623 never retire (block ceiling); sema_gl_mme_version=0 identical (version exonerated); GATE1 black 0/900 pages; next: macros' runtime STATE wrong - they branch on MME shadow-RAM STATE (slot-57 reads 0xd04..0xd2a) replay leaves zero + prelude's 6 non-SA/IR scratch writes (0x3470/0x34a8/0x3484) + client 0x124 mode not established; capture golden MME ctx state from real run and replay before calls; fallback per-boot nop_drop bisect {0,2,34,39,40} to name faulting macro then read g_tu100mmecode.h - SESSION 153 r1340-r1344

0 files changed, 0 insertions, 0 deletions