FA2 -> fresh Hopper program: exact selected instruction excerpts

These are representative regions of the two bundled noncausal functions.
They show logical components, NOT same-PC patch pairs. Instruction addresses,
registers, tile shape, thread roles and layout change during whole-body generation.
Disassembly has intervening instructions omitted. All addresses are relative
to the selected function. Native program IR is a separate diagnostic artifact.

INPUT: fa2-input.cubin
SHA-256 fbd4bc6cbc5b6f028cd30f23eb90783e97ca5a60d5e20b294dc83d08bb98ec52
Selected text: 49,792 bytes / 3,112 instruction words
Selected-text SHA-256
0a2a51d4e0a8309b49e4cbdd22d9dbfe941f75b90a4ca7f8e1ba045453947159

Transport:
/*0700*/ LDGSTS.E.BYPASS.LTC128B.128 [R246], desc[UR16][R4.64];
/*0de0*/ LDGDEPBAR;
/*0e80*/ DEPBAR.LE SB0, 0x0;
/*0e90*/ BAR.SYNC.DEFER_BLOCKING 0x0;

Matrix fragments and warp MMA:
/*1180*/ LDSM.16.M88.4 R12, [R3+0x4000];
/*11a0*/ LDSM.16.M88.4 R4, [R234+0x2000];
/*1230*/ HMMA.16816.F32 R56, R4.reuse, R12, RZ;
/*1240*/ HMMA.16816.F32 R116, R4, R14, RZ;

Intermediate source normalization (constructed forms, NOT final disassembly):
QK uses four HGMMA.64x128x16.F32 per M64 source half, then full gsb0 wait.
PV uses eight HGMMA.64x64x16.F32 per M64 source half for the N128 reduction
(two M64 halves in the source window), with a transposed V
shared descriptor. Independently constructed source forms do not copy native
FA3 matrix recipes. Complete program reconstruction follows this normalization.

OUTPUT: fa3-style-output.cubin
SHA-256 89050c25cdfbfecd0395ad88131eee670e095978583b94bbe777254b634e6a01
Selected body: 30,864 bytes / 1,929 regenerated instruction words
Selected-text SHA-256
d29b8338bfef612af95f6504d49fb8fd018d9488f145dfc5f7232ee63add88eb

Initial QK: M64 x N192 x D64, four K16 steps:
/*0b20*/ WARPGROUP.ARRIVE;
/*0b30*/ HGMMA.64x192x16.F32 R24, gdesc[UR32], RZ, !UPT;
/*0b40*/ HGMMA.64x192x16.F32 R24, gdesc[UR36], R24;
/*0b50*/ HGMMA.64x192x16.F32 R24, gdesc[UR40], R24;
/*0b60*/ HGMMA.64x192x16.F32 R24, gdesc[UR44], R24, gsb0;
/*0b70*/ WARPGROUP.DEPBAR.LE gsb0, 0x0;

Fresh scalar weights and shared probability publication (excerpts):
/*1990*/ FFMA.FTZ R24, R24, R0, -R10;
/*19d0*/ MUFU.EX2 R24, R24;
/*2ca0*/ F2FP.F16.F32.PACK_AB R24, R25, R24;
/*2cb0*/ F2FP.F16.F32.PACK_AB R25, R27, R26;
/*2cc0*/ F2FP.F16.F32.PACK_AB R26, R29, R28;
/*2cd0*/ F2FP.F16.F32.PACK_AB R27, R31, R30;
/*2ce0*/ STSM.16.M88.4 [R15], R24;
/*3080*/ MEMBAR.ALL.CTA;
/*3090*/ FENCE.VIEW.ASYNC.S;
/*30a0*/ NOP;
/*30b0*/ WARPSYNC.ALL;

Previous PV in a recurrent overlap iteration: twelve K16 steps for N192.
The intervening descriptor increments and other instructions are omitted:
/*34e0*/ HGMMA.64x64x16.F32 R120, gdesc[UR36].tnspB, R120;
/*3510*/ HGMMA.64x64x16.F32 R120, gdesc[UR36].tnspB, R120;
...
/*36f0*/ HGMMA.64x64x16.F32 R120, gdesc[UR36].tnspB, R120, gsb0;
/*3750*/ WARPGROUP.DEPBAR.LE gsb0, 0x1;
...
/*3e40*/ WARPGROUP.DEPBAR.LE gsb0, 0x0;

One K0 transfer: expect 0x6000 = 24,576 bytes, then TMA:
/*69a0*/ ELECT P4, URZ, PT;
/*69b0*/ VIADD R8, R5, 0x6000;
/*69c0*/ R2UR UR24, R8;
/*69d0*/ VIADD R9, R5, 0x30010;
/*69e0*/ R2UR UR25, R9;
/*69f0*/ UMOV UR26, UR40;
/*6a00*/ UMOV UR27, UR41;
/*6a10*/ UMOV UR28, UR42;
/*6a20*/ UMOV UR29, UR43;
/*6a30*/ IMAD.MOV.U32 R10, RZ, RZ, 0x6000;
/*6a40*/ @P4 SYNCS.ARRIVE.TRANS64 RZ, [R5+URZ+0x30010], R10;
/*6a50*/ UTMALDG.4D [UR24], [UR34], desc[UR38];

Native logical IR links:
113 / 123 / 133-134: FMA, exp2, FP16 P conversion, FP32 denominator reduction.
629: current QK.
646: previous PV.
647: current online step consumes previous PV's completed accumulator.
671: five-value carry.
673: final pending PV drain.

Physical contracts:
- K/V ready barrier count 1; empty barrier count 3, one release per math group.
- Carry stage epochs and alternating slot cursor across persistent query tickets.
- Named barriers 1 -> 2 -> 3 relay issue credit between math groups.
- Q reuses named barrier 10 with 416 participants, not the K/V empty cycle.
- Full matrix completion precedes stage release/P overwrite where required.
- Every selected published word is regenerated; the reference supplies the ABI shell.
- Shorter code and logical overlap are not by themselves a runtime validation.