fpga-archive-reports-qwen · EAGLE4 FPGA archive

Source reports/results/QWEN_GPTQ2_20261116/BUILDS_20261116.md · source md5 9f0605ff3c8ed22ff16bba80cffda3f5 · archived 2026-11-18

Download PDFOpen PDFRaw markdown

QWEN_GPTQ2 v0 builds, 2026-11-16 (qgptq2-build-watch for qwen-gptq2-arch)

Program: 7-card 4C1 drafter for the Qwen3.8-27B target (EAGLE-4 GPTQ-2 drafter, A16). All nine from verilog dadc560a, par8 hook. Ledger /mnt/hdd_data/eagle4_scratch/qwen_gptq2_20261116/stage/LAUNCHED_QGPTQ2.tsv; launch lines + stage md5s in STAGING_20261116.md §1/§9. Stage wall times are copied from each build's own <stage>/STAGE_TIMES_<id>.pg<PGID>.tsv (qg_stage_timer.py, sampled process-group RSS). DONE = positive setup AND hold WNS. Gates are report-only.

0. Snapshot (latest probe first)

probe UTC 139B MLPV L128 trantor 139A ATT L256 midyard 139D MLPV L256 asgard 139C ATT+emb buildy00 139E ATT hedge buildy01 139F MLPV L256 hedge buildy04 13A0 ATT+FR_A8 buildz01 13A1 MLPV L256+FR_A8 buildz02 13A2 MLPV L128 hedge buildz04 13A6 ATT L512+FR_A8 record_be buildz00 13A7 ATT L512+FR_A8 Explore buildz05 13A8 ATT L512+FR_A8+emb record_be buildy05 13A9 ATT L256+FR_A8+emb AD2 buildxd00
15:55Z route Phase 4 done ROUTED +0.000/+0.009 margined route Phase 5 post-route phys_opt place write_device_image place place Phase 4 init DEAD at place (URAM SLR2 682 > 624) opt route Phase 2
15:30Z route Phase 2 (start 15:18:34Z) done route Phase 5 (2 h) route Phase 4 ROUTED +0.017/+0.010 margined place post-route phys_opt place place Phase 3 init 61 min place done init 59 min post-place phys_opt
15:05Z post-place phys_opt done route Phase 5 (since 13:33Z, asgard MemAvailable 23 G) route Phase 2 route Phase 5 opt ROUTED +0.072/+0.010 margined place place (PSIP) init (since 14:29Z) opt (init took 58 min) init (since 14:31Z) place Phase 4
14:02Z place (opt done) PDI DONE (pipeline exited) route Phase 3 (asgard MemAvailable 33 G) place Phase 2 post-place phys_opt (hold fix) synth_1 route Phase 2 synth_1 init_design synth_1 init_design synth_1 synth_1
12:56Z synth_1 wdi DONE 12:53, post-build gates place (Phase 2) synth_1 place (Phase 2) synth_1 place (Phase 2) synth_1 synth_1 OOC 47/63 OOC 62/63 pkg pkg
12:29Z OOC 63/64 route Phase 16 (post-route) impl opt_design OOC 62/63 impl opt_design OOC 50/64 impl opt (finalization) OOC 63/64 OOC 63/64 pkg synth OOC pkg synth (start 12:10:26Z) pkg synth (start 12:10:29Z)
12:03Z OOC 63/64 route_design Phase 4 (place 1,601 s) OOC 63/64 OOC 62/63 OOC done 11:56 (2,857 s, peak 84.1 GB) pkg post-synth gates OOC done 11:57 (3,010 s) pkg post-synth gates (PKG gate PASS 0 failures) OOC 16/64 pkg synth pkg synth
11:37Z pkg synth (95 min) place_design (Phase 4 detailed) OOC 63/64 OOC 58/63 OOC 62/63 pkg (synth done 11:32) OOC 62/63 pkg (synth done 11:37) pkg (synth done 11:35) STEP1 pkg synth (start 11:19:59Z) STEP1 pkg synth (start 11:19:57Z)
11:07Z STEP1 pkg synth (tech mapping, 65 min in) impl_1 opt_design STEP1 pkg synth STEP1 packaging (util written) BD create/OOC launch STEP1 pkg synth OOC synth (0/63) STEP1 pkg synth STEP1 pkg synth – –

Notes 11:07Z:

1. STEP1 package results (util_pre_package.rpt, A1/A3, F11)

id LUT FF BRAM36 URAM DSP58 rpt md5 component.xml md5 A1/A3 --pkg --kind --dual 0 F11 BD synth wrapper literals
0xE4B7139A 777,062 (30.19%) 702,209 (13.64%) 1,651.5 (44.15%) 913 (47.43%) 2,757 (25.41%) ee64db91 379e3c2c PASS (aw/araddr [9:0], reaches FOLD_BASE_HI 0x1FC, addressBlock 65,536 B) 10/10 count 1 (wrapper md5 7f833b16)
0xE4B7139E 777,050 702,203 1,651.5 913 2,757 aa9e9680 8f91b2e9 PASS (same rows) 10/10 count 1 (7f833b16)
0xE4B713A0 778,339 (+1,277 vs 139A: FR_A8) 702,175 1,651.5 913 2,757 f71d3fb8 ccd4d8cd PASS (same rows) 10/10 count 1 (7f833b16; FR_A8 is a literal bake, not a top param)
0xE4B7139C 777,058 702,209 1,651.5 913 2,757 c3cb1015 beaf92a6 PASS (same rows) 10/10 count 1 (7f833b16)
0xE4B7139B (MLPV L128) 1,241,392 (48.22%) 1,062,144 (20.63%) 1,048 512 2,946 (27.16%) 020a58c8 c433e958 PASS (aw/araddr [8:0], MLPV_MANIFEST_RD 0x1FC) 4/4 count 1: .LANES(128) .GEMV_SCALE_FRAC(16) .TREE_V_MAX(248320) .ARRAY_FAST_CLK(1'D0) (wrapper md5 c583614e)
0xE4B713A2 (MLPV L128 hedge) 1,241,390 1,062,144 1,048 512 2,946 5e6122a5 446ca0af PASS (same rows) 4/4 count 1 (c583614e)
0xE4B7139F (MLPV L256 hedge) 1,491,988 1,241,164 1,048 512 3,458 acb94b13 1d14737d PASS 4/4 count 1 (9f4a8e0e)
0xE4B713A1 (MLPV L256 + FR_A8) 1,491,997 (+9 vs 139F) 1,241,166 1,048 512 3,458 fea1547b 3bcd69c8 PASS 4/4 count 1 (9f4a8e0e)
0xE4B713A7 (ATT L512 + FR_A8, Explore) 1,293,553 (50.25%) 1,029,000 (19.99%) 1,651.5 1,169 (60.73%) 3,013 (27.77%) e8a89f16 fee7fc68 PASS (aw/araddr [9:0], FOLD_BASE_HI 0x1FC) 10/10 count 1 with .LANES(512) (wrapper md5 c392ea4c)
0xE4B713A6 (ATT L512 + FR_A8, record_be) 1,293,556 1,028,997 1,651.5 1,169 3,013 bf26be55 f059aaa8 PASS 10/10 count 1 with .LANES(512) (c392ea4c)
0xE4B713A8 (ATT L512 + FR_A8 + emb v2, record_be; verilog 46d68700) 1,293,551 1,029,000 1,651.5 1,169 3,013 19639a57 7e948541 PASS 10/10 count 1 with .LANES(512) (c392ea4c)
0xE4B713A9 (ATT L256 + FR_A8 + emb v2, AD2; verilog 46d68700) 778,338 (= 13A0 778,339; +1,276 vs strap-off 139A) 702,175 1,651.5 913 2,757 34bb6d9f 734af6cf PASS 10/10 count 1 with .LANES(256) (7f833b16)
0xE4B7139D (MLPV L256) 1,491,985 (57.96%) 1,241,164 (24.11%) 1,048 (28.01%) 512 (26.60%) 3,458 (31.88%) d1873ca9 7b9fbd45 PASS (aw/araddr [8:0], reaches MLPV_MANIFEST_RD 0x1FC, addressBlock 65,536 B) 4/4 count 1: .LANES(256) .GEMV_SCALE_FRAC(16) .TREE_V_MAX(248320) .ARRAY_FAST_CLK(1'D0) (wrapper md5 9f4a8e0e)

F11 literals (ATT): .LANES(256) .ATTN_N_HEADS(10) .ATTN_D_HEAD(128) .HEAD_PAR(5) .G5_HPS(2) .HFC_INBOX_FRAC(7) .GEMV_SCALE_FRAC(16) .K_MAX(10240) .DEPTH_N(1) .ENORM_D0(1) in <clone>/build/prj.gen/sources_1/bd/top/ip/top_eagle4_drafter_lb_0_0/synth/top_eagle4_drafter_lb_0_0.sv (whitespace-stripped grep). Util rows: pkg_tmp_oneflip_p2s_head/util_pre_package.rpt lines 38/41/61/66/79 (Registers / CLB LUTs* / Block RAM Tile / URAM / DSP Slices).

2. Per-build stage times (s) and place/route

0xE4B7139A ATT L256, midyard (PGID 34274, root ~/build4c1/s3LAq2BE9A)

stage start end s pgroup RSS peak MB vivado peak MB
pkg_synth 10:08:13Z 10:27:42Z 1,169 45,045
ooc_synth (63 runs) 10:32:31Z 10:57:02Z 1,470 68,596
synth_1 10:57:02Z 11:00:40Z 218 32,876 6,521
init_design 11:00:43Z 11:06:48Z 365 24,692 16,348
opt_design 11:06:48Z running

0xE4B7139E ATT hedge, buildy01 (PGID 3163515)

| pkg_synth | 10:19:10Z | 10:57:58Z | 2,328 | 44,620 | |

0xE4B713A0 ATT + FR_A8, buildz01 (PGID 2541840)

| pkg_synth | 10:19:10Z | 10:56:54Z | 2,264 | 45,442 | | | ooc_synth | 11:07:27Z | running | | | |

0xE4B7139A at 11:37Z

| opt_design | 11:06:48Z | 11:18:19Z | 690 | 35,678 | 17,537 | | place_design | 11:18:19Z | running (Phase 4 detailed placement) | | 42,628 | | midyard MemAvailable 43 G during place. | place_design | 11:18:19Z | 11:45:01Z | 1,601 | 42,631 | 35,035 | | route_design | 11:51:58Z | running at 12:03Z (Phase 4 initial routing) | | 42,931 | | MLPV L256 vs L128 at pkg (139D vs 139B): +250,593 LUT, +179,020 FF, +512 DSP, BRAM/URAM equal.

Other builds at 11:37Z (from their STAGE_TIMES)

id pkg_synth s (pgroup peak MB) ooc_synth start, runs done (peak MB)
139B 5,695+ (52,332), still in pkg
139C 3,043 (45,311) 11:22:31Z, 58/63 (62,931)
139D 2,850 (27,205; timer started late, 11:10Z) 11:19:10Z, 63/64 (42,215)
139E 2,328 (44,620) 11:08:36Z, 62/63 (84,129)
139F 4,373 (52,632) packaging
13A0 2,264 (45,442) 11:07:27Z, 62/63 (58,274)
13A1 4,678 (52,111) packaging
13A2 4,535 (52,834) packaging

FR_A8 literal bake (staged top/eagle4_config.svh): 13A0 :525 and 13A7 :525 `define FR_A8_V 1'b1; 13A1 :827 1'b1 (+ QG_STRAPS_EFFECTIVE.env:26 CFG_FR_A8_V=1'b1); 139F :827 1'b0 (control). MLPV use = one site, top/eagle4_drafter.sv:7931 FR_A_FRAC = (Q5120 && EAGLE4_FR_A8) ? 8 : 9 (a requant shift), which is why the MLPV pkg delta is only +9 LUT; ATT (a_align_shr SH 3 vs 4) costs +1,277 LUT (13A0 vs 139A).

13A8 / 13A9 (added 12:10Z by qwen-gptq2-arch; verilog 46d68700, A2 PASS at derive): A1/A3 PASS, F11 10/10 (above). Emb aperture v2: clone src/bd/create_bd_design_layout_b.tcl (md5 58e5328a on both, = 139C) carries the three ;# EMB_APERTURE_V2 lines assign_bd_address -offset 0x004680000000 / 0x0046C0000000 / 0x004700000000 -range 0x40000000 ... m_axi_emb] [... S08_AXI/HBM13_PC0 / HBM13_PC1 / HBM14_PC0] (1 each), the built prj.srcs/.../top.bd names each segment, and the in-build row emb-aperture-v2 PASS ('4 m_axi_emb segments'; 13A8 13:03:27, 13A9 12:49:21). 13A6 (no emb) BD tcl 3e434ed9 has 0 EMB_APERTURE_V2 lines (control). FR_A8: staged top/eagle4_config.svh:525 define FR_A8_V 1'b1 on 13A6/13A8/13A9. G4N: staged primitives/stage_geometry.svh:67 SG_ATT_TP4_G4N 4608 on 13A6/13A8 (L512), 3840 on 13A9 (L256, = 139A). B3 (13A9 drafter OOC dcp 0a488755, box1): araddr/awaddr[0..8] loaded (17..636), [9] 0 loads on both (= 139A, INFO: no CSR > 0x1FC) (/mnt/data/eagle4_scratch/qg_buildwatch_20261116/13A9_B3.txt 333e87f4). B4 vs strap-off parent 139A (FR_A8 0, same L256 A16): drafter 779,524 vs 777,919 LUT (+1,605), and the delta sits on the FR_A8 sites: g_resid_fullh.u_r3a_hfc_align (80 LUT in 139A) is ABSENT (its shift F2_HFC_FRAC - FR_A_FRAC folds when FR_A_FRAC = 8, eagle4_drafter.sv:9017/9070/9091), u_seam_r4u 1,419 -> 1,787, u_seam_r5 401 -> 665, u_resid_r3 1,525 -> 1,557, top (inst) 70,267 -> 71,507; rope / enorm / array unchanged -> PASS (non-zero and differs from the parent) (13A9_B4_hier.rpt e916a526). 13A8 OOC not done at 14:10Z. F1: at route.

13A6 / 13A7 (added 11:20Z by qwen-gptq2-arch: ATT L512 A16 + FR_A8_V 1'b1, pair with 13A1): timers running (STAGE_TIMES_0xE4B713A6.tsv, ..A7.tsv). The staged forks/head_hxf_cons/build/src_staged_oneflip_p2s_head/primitives/stage_geometry.svh:67 reads SG_ATT_TP4_G4N 4608 on both (the only 3840 is dryrun_qgptq2/ref_stage_qgptq2_att, the dry-run reference). F11 for these two expects .LANES(512).

3. 0xE4B7139A (ATT L256, midyard): FINAL PDI written; FINAL_TRUE +0.109 / +0.010 (meets the record bar +0.030 / +0.010); F1 srpol rerun in flight

item value source
route WNS +0.069 / TNS 0 / WHS +0.001 / THS 0 (1,772,606 setup endpoints) impl_1/route_report_timing_summary_0.rpt md5 98188145
post_route_phys_opt +0.069 / +0.001 (equal to route; these are the MARGINED numbers: ^KVC2_MARGIN_NONVACUITY clocks=2 runme.log:825, ^KVC2_MARGIN_CANCELLED clocks=2 :2296/:2312) post_route_phys_opt_report_timing_summary_0.rpt md5 e4712475
FINAL_TRUE (as-saved, uncertainty 0) WNS +0.109 / WHS +0.010; ^ASSAVED +0.109/+0.010 (UNC_ASSAVED empty on every clock = margin cancelled). Worst setup aur_link1/start_d_reg -> gtm_tx_gearbox_66_64_3/data_in_d1_r_reg[46]/CE ch0_txoutclk_3, 0 levels, SLR1 (shell). Worst hold ld_hop_st/U0/ip2bus_data_i_D1_reg[2] -> s_axi_rdata_i_reg[29]/D clk_pl_0, 0 levels, SLR1->SLR0, FDRE->FDRE (shell AXI GPIO). Hold top 20: clk_pl_0 AXI GPIO (ld_hop_st/ld_st_a/ld_st_b) +0.010..+0.015, aurora rx/tx +0.012..+0.014, TP link tp_l4 g_port->g_eng rw_r clkout1_primitive +0.013..+0.014; no drafter path in the 20 worst holds (same classes as the record 0x13BE: TP link +0.011, SMBus +0.012). Hold by clock: clk_pl_0 +0.010, rxoutclk_3 +0.012, txoutclk_3/4 +0.013, clkout1_primitive +0.013, usr_0 +0.021, usr_2 +0.038. Setup by clock: txoutclk_3 +0.109, usr_2 +0.146 (drafter kvf afifo), clk_pl_2 +0.162. Route status 0 nets with routing errors. Run on buildx04 /shared/scratch/qg_bw_20261116/139A/ with licence 2103; evidence copied to /mnt/data/eagle4_scratch/qg_buildwatch_20261116/139A_ft/ (hold20 52886f28, setup20 82e729ea, tsum dfe3e7b8). Earlier text: running on buildx04 /shared/scratch/qg_bw_20261116/139A/ (copy of top_wrapper_postroute_physopt.dcp md5 782c6953 both ends); hold20/setup20 lists with clock pair + SLR + ref scripts/4c1/qgptq2/watch/final_true.tcl
stage times place 1,601 s · phys_opt 416 s · route 2,961 s · write_device_image 489 s; launch -> wdi end 2 h 34 m STAGE_TIMES pg34274
per-SLR (SLR0/1/2) SLICE 58,351/67,963/55,306 = 51.85/66.29/51.82 % · LUT 205,149/301,965/285,274 · FF 341,647/249,781/151,127 · RAMB36 767.5/501/499 · URAM 145/276/492 (SLR2 78.85 %) · DSP 2,270/247/177 place_report_utilization_0.rpt md5 1008eb64 §15
SLL SLR2<->SLR1 6,581 (34.88 %) · SLR1<->SLR0 9,096 (48.20 %) of 18,870 same §13
GATE_VERDICTS midyard:~/build4c1/repo_qgptq2_20261116/reports/results/GATE_VERDICTS_DUALCLK_s3LAq2BE9A_head_ASGARD.txt, 256 lines at 13:10Z (still appending)
FAIL rows so far gate-selftest p2s_selfref INCONCLUSIVE; 8-3848 bake-drop (100); OSPREYCHECK g_r7_resid_on absent / r7_sum_clip_cnt absent (x2 incl. pkg-census); enforce-build-manifest vs 0xE4B71128; portw reg0 range 0x1000; STEP3 THREEC_A h1-only export mask + G6-ends-at-R3-done hookup NOT in staged drafter.sv; v1-netlist SIBW-TP col_tag=0 (x2); enorm-d0-engine(watcher) INCONCLUSIVE rc=3 (x2); config-manifest vs INTENT (x2); pblockgate 27 dead filters to be diffed against the record F_BE.txt (/mnt/data/eagle4_scratch/handoff_cut_20261112/adj_13be/F_BE.txt, 40 rows) once the pipeline ends
B3 (drafter OOC dcp, box1) araddr/awaddr [0..8] all have leaf loads (2..638); [9] has 0 loads on both = INFO, not FAIL: the highest ATT CSR is FOLD_BASE_HI 0x1FC (A3), DUAL_MODEL 0, so bit 9 decodes nothing (0x200..0x3FC alias 0x000..0x1FC: the D2 alias-probe rule applies before any write above 0x1FC) /mnt/data/eagle4_scratch/qg_buildwatch_20261116/139A_B3.txt e0d5e51b; dcp prj.gen/.../top_eagle4_drafter_lb_0_0.dcp 182,258,108 B
B4 (same dcp, report_utilization -hierarchical -hierarchical_depth 4) drafter 777,919 LUT / 702,158 FF / 913 URAM / 2,757 DSP; Qwen blocks non-zero: u_attn_full/u_rope 2,177 LUT, 8,349 FF, 108 RAMB36, 256 DSP; g_enorm_depth_requant.u_rmsnorm_g1_e_d 2,204 LUT / 48 DSP; LUTG lanes g_dual.g_lutg.u_lane 1,563 LUT each; u_array 443,358 LUT / 768 URAM; u_attn_full 119,605 LUT / 145 URAM / 2,221 DSP 139A_B4_hier.rpt 3afd4c7e
FINAL PDI midyard:~/eagle4_artifacts/uni_oneflip/oneflip_p2s_head_asgard_0xE4B7139A_66m_c537f6ab.pdi = build root aved_oneflip_p2s_head_asgard/hw/amd_v80_gen5x8_25.1_pdi23/amd_v80_gen5x8_25.1.pdi; md5 c537f6abf185e4df51931341f9146014 (both), 59,661,088 B (56.9 MiB, inside 30-120 MiB), magic 16a5f792. .pdi.QUARANTINED marker e5e61f47 (45 FAIL lines); GATE_VERDICTS copy 4f273aa5 md5sum, `head -c4
F1 srpol pipeline run = UNVER (rc 2, must_pass_gate --srpol refused on midyard RAM: GATE_VERDICTS :373-374). Rerun on buildx04 (licence 2103; copies postroute_physopt 782c6953 / opt a016420c, md5-equal to midyard): tooth 1 PASS SRPOL_SUMMARY mixed=0 minority_loads=0 selftest=PASS FDRE=0 FDSE=0 FDCE=0 FDPE=0 (13:23:45-13:31:16Z, tcl d56fe71b); tooth 2 PASS (srpol-optdiff fast, 13:23:45-13:36:47Z) /mnt/data/eagle4_scratch/qg_buildwatch_20261116/139A_ft/GV_139A_srpol.txt
FAIL set vs record 0x13BE (adj_13be/F_BE.txt, 40 rows, path-normalised) 41 rows. Gone vs record: OSPREY-NETLIST ATTN_N_HEADS (now PASS: chkval 10/128/7), marginal body cone RESULT, pblockgate 21 (now 27). New: (1) 8-3848-bake-drop (100) = 100 x g_resid_fullh.g_forreason.fr_rem_buf[*] undriven (eagle4_drafter.sv:9728): it is written only under ACCFIX_SEAM_P (Q5120 pre-bake 0, config.svh:1039) and its only consumer rem_chunk is folded to 0 by SEAM_FIX_P && seam_en (layers/for_reasoning_producer.sv:195) -> dead, BY DESIGN; (2) LUTG-netlist wants lutg_a8q defs >=1 but A16 = A8Q 0 (256 lutgemm_lane_w2 defs present) -> BY DESIGN, stale row in this gate (not the census STAGING §3 updated); (3) portw-netlist-araddr/awaddr sim_netlist decl [8:0] (want [9:0]) = bit 9 has 0 loads (my B3) -> OK while no CSR > 0x1FC (A3 PASS); (4) STEP4c final strap netlist gate = STRAP_GATE_VERDICT 'wrapper has NO .DOT_E3/.ATT_TP/.ATT_G7/.ATT_TP_RANKS/.KVO_QSKIP/.MODEL_H/.MODEL_INTER/.FR_A8/.MODEL_N_HEADS/.PIPE_CUT_SILU/.ROPE_ROM_BRAM' = the header-localparam class STAGING §4 A4 says is shipped by literal bake + 19-line readback -> BY DESIGN; (5) enorm-d0-engine INCONCLUSIVE rc 3 (x3): rerun of scripts/check_enorm_d0_engine.sh on the clone = instruments P=320 I=1 S=10 present, [Q] ^module ...rmsnorm_sumsq defs = 4 (wants 5) -> 3/4 votes; the engine IS in the netlist (B4: g_enorm_depth_requant.u_rmsnorm_g1_e_d 2,204 LUT / 981 FF / 48 DSP; straps ENORM_D0 1, DEPTH_N 1 SELECTED) -> the [Q] module-def count is not calibrated for the Q5120 parameterisation (instrument drift, not a missing engine); §FIXB info: u_rmsnorm_g1_e is a clone of u_rmsnorm_g1 (FIX B not landed, informational as on the record); (6) srpol tooth1/2 rc 2 -> rerun above /mnt/data/eagle4_scratch/qg_buildwatch_20261116/139A_gv/

Watch tools (committed): scripts/4c1/qgptq2/watch/{probe.sh, impl.sh, f11.sh, emb.sh, b34.tcl, final_true.tcl, run.sh, hosts.txt}.

4. 0xE4B713A0 (ATT L256 + FR_A8, buildz01): ROUTED (15:05Z)

item value source
route Router Completed Successfully; Post Routing Timing Summary WNS +0.072 / TNS 0 / WHS +0.010 / THS 0 (margined, same KVC2 hook as 139A) impl_1/runme.log (Route 35-20)
per-SLR (SLR0/1/2) SLICE 97,720/75,112/4,998 = 86.84/73.26/4.68 % · LUT 398,974/370,046/24,223 · FF 478,879/251,493/12,266 · RAMB36 1,010.5/722/35 · URAM 328/519/66 · DSP 2,423/257/14 place_report_utilization_0.rpt md5 18eb31f6 §15
SLL SLR2<->SLR1 5,550 (29.41 %) · SLR1<->SLR0 11,955 (63.35 %) same §13
note same place/route directives as 139A (place_design -subdirective {Floorplan.BalancedSLR.high GPlace.ReduceCongestion.high DPlace.ReducePinDensity.high}, route/phys_opt AggressiveExplore, both runme.log), yet the placer packed the design into SLR0/SLR1 and left SLR2 at 4.68 % SLICE (139A: 51.85/66.29/51.82 %). Both route positive
post_route_phys_opt +0.072 / +0.010 (margined) post_route_phys_opt_report_timing_summary_0.rpt
build-root PDI buildz01:/shared/build/s3LAq2A0/aved_oneflip_p2s_head_asgard/hw/amd_v80_gen5x8_25.1_pdi23/amd_v80_gen5x8_25.1.pdi md5 b5db7626961c79a34332946d302f40b8, 59,275,584 B (56.5 MiB), magic 16a5f792, 15:58Z
FINAL_TRUE +0.112 / +0.011; worst setup in the drafter g_kvf.u_kvf/g_ch[0].g_afifo count_value -> fifo RAM (usr clock); worst holds aurora tx/rx +0.011..+0.012 (shell) (buildx03 copies postroute_physopt 8f3ae30c / opt 03f84211, md5-equal) /mnt/data/eagle4_scratch/qg_buildwatch_20261116/13A0_ft/ (hold20 48ef3cf5, setup20 54818674)
F1 srpol tooth 1 PASS mixed=0 minority_loads=0 selftest=PASS (16:15:48Z); tooth 2 PASS (16:21:19Z) 13A0_ft/GV_13A0_srpol.txt 682dc233

5. B3/B4 on the L512 images (drafter OOC dcps on box1)

6. Added 14:38Z: 0xE4B713AA (MLPV L256 + FR_A8, arm explore; hedge of 13A1) on midyard

Stage midyard:~/build4c1/qgptq2_mlpv_stage_13AA_r2, root ~/build4c1/s3LCq2AA, PGID 172063, verilog 46d68700, CFG_LANES_V 256 (STAGING §14). A rescue waiter R13AA started 14:41:23Z (EVENTS.log).

7. 0xE4B7139E (ATT L256 hedge, arm record_be, buildy01): ROUTED (15:30Z)

item value source
route Router Completed Successfully; Post Routing Timing Summary WNS +0.017 / WHS +0.010 (margined: the KVC2 setup margin is 0.040 on the usr clocks, so the as-saved number is expected higher; FINAL_TRUE owed) impl_1/runme.log
place place_design -subdirective {Floorplan.ForceSpreading.high GPlace.ForceSpreading.high GPlace.ReduceCongestion.high DPlace.ReducePinDens...} runme.log
per-SLR (SLR0/1/2) SLICE 107,768/78,332/37,197 = 95.77/76.40/34.85 % · LUT 345,938/309,369/138,756 · FF 456,048/213,144/73,513 · RAMB36 867.5/642/258 · URAM 211/444/258 · DSP 2,329/275/90 place_report_utilization_0.rpt md5 2f5559fc
SLL SLR2<->SLR1 1,993 (10.56 %) · SLR1<->SLR0 8,975 (47.56 %) same
FINAL_TRUE +0.057 / +0.012 (as-saved = FINAL_TRUE); worst setup TP link tp_l4/u_l4/want_reg[23] -> g_port[1].u_tf ... mem_reg_0_15_0_13/RAME_D1/I clkout1_primitive 8 levels; worst hold aurora aur_link1 gtm_rx_gearbox -> pre_r1_rxdata...[36]/D ch0_rxoutclk_3 +0.012; hold top 5 = aurora + TP link only (buildx02 copies postroute_physopt d89fcb8d / opt 9712977b, md5-equal) /mnt/data/eagle4_scratch/qg_buildwatch_20261116/139E_ft/ (hold20 61c9e0f6, setup20 454fcb31)
F1 srpol tooth 1 PASS mixed=0 minority_loads=0 selftest=PASS (16:17:55Z); tooth 2 PASS (16:28:56Z) 139E_ft/GV_139E_srpol.txt 0fd429a6
build-root PDI buildy01:/shared/build/s3LAq2BE9E/aved_oneflip_p2s_head_asgard/hw/amd_v80_gen5x8_25.1_pdi23/amd_v80_gen5x8_25.1.pdi md5 723ab73b43f5e945dd45d42f043e7896, 66,380,224 B (63.3 MiB), magic 16a5f792 (pipeline still in post-build gates at 16:35Z)

L512 init_design wall: 13A7 58 min (13:36->14:34Z, single-thread XDC parse of eagle4_drafter_r4silu_slr0_evac_pblock.xdc); 13A6 and 13A8 still in init at 15:30Z (61 / 59 min). L256 init: 139A 6 min.

8. 0xE4B713A7 (ATT L512 + FR_A8, Explore, buildz05): DEAD at place_design 15:25:54Z

9. 0xE4B7139D (MLPV L256, asgard): ROUTED (15:55Z)

route Post Routing Timing Summary WNS +0.000 / WHS +0.009 (margined) runme.log
per-SLR (SLR0/1/2) SLICE 110,110/95,521/94,230 = 97.85/93.17/88.30 % · LUT 619,137/526,608/360,912 · FF 345,374/745,556/221,741 · RAMB36 359/772.5/8 · URAM 464/48/0 · DSP 977/2,305/79 place_report_utilization_0.rpt md5 a290f71d
SLL SLR2<->SLR1 4,936 (26.16 %) · SLR1<->SLR0 11,400 (60.41 %) same
route wall 13:33:53Z -> ~15:50Z (rip-up at congestion level 6; asgard MemAvailable 23 G through it)
FINAL PDI asgard:/shared/build/s3LCq2619D/aved_oneflip_p2s_tail_asgard/hw/amd_v80_gen5x8_25.1_pdi23/amd_v80_gen5x8_25.1.pdi md5 d6394fb60adc067d4afda2af6678beb3, 96,477,536 B (92.0 MiB), magic 16a5f792, 16:25:03Z
F1 srpol both teeth on buildx03 (copies from the rescue waiter: postroute_physopt 3e3a89df = asgard; opt 97668675 = asgard re-checked), started 16:23:15Z /shared/scratch/qg_bw_20261116/139D/
pipeline DONE 16:11:38Z (wdi 590 s; launch -> wdi end 5 h 02 m); uni_oneflip copy asgard:~/eagle4_artifacts/uni_oneflip/oneflip_p2s_tail_asgard_0xE4B7139D_66m_d6394fb6.pdi md5 d6394fb6 (= build root), .QUARANTINED 45 FAIL lines
FINAL FAIL diff vs M1361 (37 vs 33 normalised rows; /mnt/data/eagle4_scratch/qg_buildwatch_20261116/139D_gv/) adds TCLKCHECK 'routed clk_usr_1 (ap_clk_fast) period is not 7.57' (GATE_VERDICTS 11:25:03 local): the routed timing summary has NO clk_usr_1, only clock_reset_clk_usr_0 {0.000 7.576} 15.151 ns 66.001 MHz (ARRAY_FAST_CLK 1'D0, F11) -> stale record row, BY DESIGN; gone vs M1361: DUALCLK-wrapper, tpsafe-loop-stage, tpsafe_staged_loop. Margined post_route_phys_opt WNS 0.000 / WHS +0.009 (usr_0 +0.004)
FAIL rows vs record M1361 (/mnt/data/eagle4_scratch/g8own_silicon_20261112/...M1361_e24deb3f.GATE_VERDICTS.txt 01865bb0), partial (139D file e262c874, 16 rows at 16:23Z) new: 8-3848 = 100 x g_g8_splice.nrm_buf[0..99] (Vivado message cap), written only under !G8_W2 (eagle4_drafter.sv:9737), G8_W2_V 1 (eagle4_config.svh:855), record config has no G8_W2 -> BY DESIGN; STEP3 v1.2 csa_tree16_nw / gemv_lane_scaled_stream_dual(_ext)_opt not in staged src = LUTG-replaced modules -> BY DESIGN; p2-netlist A8Q 0 = A16 -> BY DESIGN; enforce-build-manifest vs 0xE4B71128 (ATT record class); u_lmdown g_wbank count 0 (want 128) also in M1361 (record class)
F1 srpol tooth 1 PASS mixed=0 minority_loads=0 selftest=PASS (16:23:15-16:38:47Z, buildx03, copy 3e3a89df); tooth 2 PASS (16:54:48Z, opt 97668675). Applies to the BUILD dcp; the rescue dcp needs its own srpol /shared/scratch/qg_bw_20261116/139D/
FINAL_TRUE (R139D waiter, EVENTS.log 16:49:20Z) WNS 0.000 / WHS +0.009, nfail 0/0, unrouted 0 -> below +0.030/+0.010 -> margin rescue STARTED 16:44:35Z on buildx03 (ARM B, MS 0.040, MH 0.012, NFAIL 200, JOBS 8). The build PDI d6394fb6 is therefore NOT the ship candidate; the rescue PDI will be /mnt/data/eagle4_scratch/qg_rescue_20261116/EVENTS.log

10. Coordinator actions 15:49-16:02Z (qwen-gptq2-arch)

13A6 and 13A8 were killed on purpose at 15:49:37/38Z (xdc fe2a53bd guarantees Place 30-5423), and 13B0 (buildz00) at 16:02:12Z as a coordinator RED (id collision with box1-two-v2; marker /shared/build/qgptq2_att_stage_13B0/KILLED_BY_ARCH.txt). Fixed L512 siblings (verilog 46d68700, xdc 99a8d299 = one lane group per SLR): 0xE4B713AE buildy05 (record_be; root /shared/build/s3LAq2AE, PGID 2014413) and 0xE4B713AF buildz05 (explore_fs; root /shared/build/s3LAq2AF, PGID 2540150), both launched 15:55Z.

11. 0xE4B7139C (ATT L256 + emb v2, buildy00): ROUTED (17:00Z)

route Post Routing Timing Summary WNS +0.084 / WHS +0.011 (margined) runme.log
per-SLR (SLR0/1/2) SLICE 98,049/74,906/5,467 = 87.13/73.06/5.12 % · LUT 399,502/369,166/24,242 · RAMB36 1,035.5/645/87 · URAM 301/519/93 · DSP 2,427/249/18 place_report_utilization_0.rpt md5 bf082514
SLL SLR2<->SLR1 6,583 (34.89 %) · SLR1<->SLR0 13,149 (69.68 %) same
Same SLR2-nearly-empty shape as 13A0 (both AD2 default arm).
FINAL_TRUE +0.124 / +0.012 (worst setup TP link tp_l4/u_l4/g_port[1].u_tf fifo rdp -> RAM; worst hold tp_x/u_hc/u_s1/u_lnk accL -> lo3 clkout1_primitive +0.012, then aurora +0.014) on copy postroute_physopt ed1c7605 / opt d58adc47 (buildx02) /mnt/data/eagle4_scratch/qg_buildwatch_20261116/139C_ft/ (hold20 573081f9, setup20 3553a2c6)
--- --- ---
F1 srpol tooth 1 PASS mixed=0 minority_loads=0 selftest=PASS (17:29:58Z); tooth 2 PASS (17:35:59Z) 139C_ft/GV_139C_srpol.txt fd9c87b0
F11 enorm routed (qwen-gptq2-arch arch_enorm_routed{,2}.tcl c8bdadfd/d72347ca) u_rmsnorm_g1_e_d 5,353 prims / 946 FF, out pins 279 loaded 266; u_rmsnorm_g1_e 279 out pins loaded 9 (E_LOADED x_req_chunk[0..8]); ED_UNLOADED = done / out_chunk_idx_o[0..8] / out_chunk_strobe / out_valid / x_req_valid -> identical to 139A's routed output (diff = line order + premlp 5,467 vs 5,468 prims). Note: the census rows print LUT=0 / DSP=0 because the tcl buckets by PRIMITIVE_GROUP LUT/DSP, which Versal does not use (same defect I hit in slr_census v1) 139C_ft/f11a.out c4a5b7cc, f11b.out da31f4eb
FINAL PDI not written yet at 17:50Z (wdi started 17:18:34Z)

12. Added 16:19-16:52Z (ledger): 0xE4B71397 ATT (asgard, record_be, root /shared/build/s3LAq297, PGID 757670) and 0xE4B71399 MLPV L256 (buildz01, root /shared/build/s3LCq299, PGID 2652997; rescue waiter R1399). R13A3 waiter started 16:53:20Z for buildy01 s3LCq2A3 (stage present, not in the ledger at 17:00Z).

13. Post-place SLR gate (scripts/4c1/qgptq2/qg_place_slr_gate.py md5 592c2d74, report-only; FAIL = SLICE > 95 %, URAM/RAMB > 100 %, SLL > 90 %), run 17:10Z on each build's impl_1/place_report_utilization_0.rpt (rows /mnt/data/eagle4_scratch/qg_buildwatch_20261116/slr_rows.txt)

id rpt md5 SLR0 SLICE/LUT/URAM/RAMB36/DSP % SLR1 SLR2 SLL 2-1 / 1-0 % gate
0xE4B7139A 1008eb64 51.85/22.79/21.42/55.41/56.98 66.29/36.81/44.23/40.58/7.15 51.82/33.41/78.85/41.58/5.19 34.88 / 48.2 PASS 0
0xE4B7139B 165cc37e 99.94/53.87/13.59/11.71/4.77 FAIL 92.28/55.17/51.92/74.33/72.02 75.06/38.19/15.38/6.17/4.99 41.39 / 53.33 FAIL 1
0xE4B7139C bf082514 87.13/44.38/44.46/75.02/60.92 73.06/45.01/83.17/53.0/7.2 5.12/2.84/14.9/7.25/0.53 34.89 / 69.68 PASS 0
0xE4B7139D a290f71d 97.85/68.78/68.54/18.12/24.52 FAIL 93.17/64.2/7.69/62.0/66.7 88.3/42.27/0.0/0.67/2.32 26.16 / 60.41 FAIL 1
0xE4B7139E 2f5559fc 95.77/38.43/31.17/62.57/58.46 FAIL 76.4/37.72/71.15/52.67/7.96 34.85/16.25/41.35/21.5/2.64 10.56 / 47.56 FAIL 1
0xE4B713A0 18eb31f6 86.84/44.32/48.45/73.23/60.82 73.26/45.12/83.17/59.33/7.44 4.68/2.84/10.58/2.92/0.41 29.41 / 63.35 PASS 0
0xE4B713A1 4562747d 99.45/49.05/11.82/8.58/6.43 FAIL 92.97/69.15/41.03/65.75/77.14 97.82/59.77/28.21/7.58/12.88 FAIL 50.79 / 49.13 FAIL 2
0xE4B713A2 6e1942ae 63.76/25.06/44.31/16.41/7.25 92.32/62.51/33.97/69.33/68.34 99.24/61.67/0.0/5.92/5.81 FAIL 34.09 / 48.67 FAIL 1
0xE4B713A9 bca9bd54 87.08/44.71/45.35/75.84/59.94 52.82/30.68/55.77/37.83/6.25 25.43/16.25/41.35/21.5/2.64 11.07 / 48.09 PASS 0
Module census of the FAIL SLRs (placed dcp, scripts/4c1/qgptq2/watch/slr_census.tcl v2 = REF_NAME filters; v1 printed BLOCKRAM only because Versal PRIMITIVE_GROUP is not LUT; files /mnt/data/eagle4_scratch/qg_buildwatch_20261116/census/<id>_census.txt), LUT placed inside the hot SLR, drafter hierarchies >= 40K LUT:
id (hot SLR) owners (LUT in that SLR) census md5
--- --- ---
139D SLR0 97.85 % g_array_scaled.u_array 558,659 · g_tree_pwide.u_tree 123,262 · g_tree_pwide.g_ser.u_cheap_ser 43,190 087465ee
139B (L128) SLR0 99.94 % no >100K owner: u_seam_r5 90,648 · u_seam_r1_h 59,057 · u_seam_r1_e 57,676 · u_array 53,139 · g_tree_out.u_tree_dma 47,909 0f3e8663
139E SLR0 95.77 % u_attn_full 139,648 · g_att_g6.u_tp 47,192 · u_array 45,530 604e8aa3
13A1 SLR0 99.45 % u_array 95,478 · u_seam_r5 92,716 · u_seam_r1_e 57,589 · u_seam_r1_h 55,170 af762944
13A1 SLR2 97.82 % g_g8_splice.g_g8p.u_g8 332,726 · u_array 210,973 · u_tree_dma 45,873 af762944
13A2 (L128) SLR2 99.24 % g_g8_splice.g_g8p.u_g8 339,899 · u_seam_r5 90,157 264b8558
13AA SLR0 96.70 % u_array 515,977 · g_tree_pwide.u_tree 122,401 · g_ser.u_cheap_ser 43,105 b85c4a6d
13AA SLR2 95.15 % g_g8_splice.g_g8p.u_g8 331,953 b85c4a6d
13A3 SLR0 99.10 % u_array 134,209 · u_seam_r5 93,268 4e343f40
13A3 SLR2 98.92 % u_g8 332,725 · u_array 167,325 · u_tree 123,262 · u_tree_dma 45,873 · u_cheap_ser 43,191 (tree moved to SLR2 as intended; SLR2 now carries G8 + tree + 1/3 array) 4e343f40
13AF (L512) SLR0 99.99 % u_array 383,393 (= grp0 at L512 under 99a8d299) · u_attn_full 139,651 · g_att_g6.u_tp 82,196 · u_ctrl 45,256 8a93beea
Census was run on md5-equal placed-dcp copies (139D fe91aa69, 139B b68c4437 on buildx01; 13A1 cf58c576, 13A2 80114052 on buildx04; 139E 6c39678d on buildx03).

14. ~17:54Z

id state numbers
0xE4B7139B MLPV L128 (trantor) ROUTED, post-route phys_opt since 17:51:44Z Post Routing Timing Summary WNS +0.002 / WHS +0.010 (margined); SLR0 SLICE 99.94 % (gate FAIL, §13) -> FINAL_TRUE decides (R13x waiter n/a: 139B has no rescue arm)
0xE4B713A9 ATT L256 + FR_A8 + emb v2 (buildxd00) ROUTED, post-route phys_opt since 17:49:55Z WNS +0.032 / WHS +0.010 (margined)
0xE4B71397 ATT (asgard, record_be) placed, route since 17:54:12Z SLR gate PASS: SLICE 93.94/69.94/55.96 %, URAM 21.42/55.77/67.31 %, SLL 2-1 47.23 % / 1-0 43.54 % (rpt ba140d40)
0xE4B713AA MLPV L256 + FR_A8 explore (midyard) place Phase 4
0xE4B713A1 / 13A2 / 139F route
0xE4B713AF ATT L512 (buildz05, 99a8d299) synth_1 (13AE, same xdc, KILLED_BY_ARCH 17:07:18Z: wave-4 probe measured Place 30-5425 SLR0-1 SLL 22,544 > 18,870 on this floorplan)
1398 / 1399 / 1392 / 1393 STEP1 pkg synth

15. ~18:20Z

16. 0xE4B713A9 (ATT L256 + FR_A8 + emb v2, AD2, buildxd00) checks (buildx02, copies md5-equal; opt 660ad5a4)

FINAL_TRUE +0.072 / +0.012; worst setup aurora aur_link3 gtm_tx_gearbox_66_64_2 count_i -> data_in_d1_r[14] ch0_txoutclk_5 4 levels; worst hold AXI-lite ld_ctl_a clk_pl_0 +0.012, aurora +0.013 (shell) /mnt/data/eagle4_scratch/qg_buildwatch_20261116/13A9_ft/
F1 srpol tooth 1 PASS mixed 0 / selftest PASS (18:10:23Z); tooth 2 PASS (18:21:24Z) 13A9_ft/GV_13A9_srpol.txt
F11 enorm routed ENORM rows identical to 139A (sorted diff empty) 13A9_ft/f11{a,b}.out
FINAL PDI not written at 18:40Z (wdi 17:57:40Z)

17. 0xE4B7139B (MLPV L128, trantor) checks (buildx03; opt copy b1fce937)

FINAL_TRUE +0.002 / +0.010 = positive but BELOW the record bar +0.030 / +0.010. Worst setup = shell GT: gtw1 ... quad_inst/CH0_RXUSRCLK -> aur_link1/gtm_rx_gearbox_64_660/data_in_d1_r_reg[50]_bret/D ch0_rxoutclk_3, 3 levels, SLR1; the next setup paths are +0.013 / +0.014 (same gearbox). Worst hold = drafter u_array g_grp[1].g_lane[5] ... u_lane/sc_sr_ -> DSP_A_B_DATA58 usr_0 +0.010 (SLR1->SLR2), then g_grp[0].g_lane[33] qz_d_reg +0.011 /mnt/data/eagle4_scratch/qg_buildwatch_20261116/139B_ft/
F1 srpol tooth 1 PASS mixed 0 / selftest PASS (18:15:10Z); tooth 2 PASS (18:33:41Z) 139B_ft/GV_139B_srpol.txt
build-root PDI trantor:/data/qgptq2_build/s3LCq2619B/aved_oneflip_p2s_tail_asgard/hw/amd_v80_gen5x8_25.1_pdi23/amd_v80_gen5x8_25.1.pdi md5 75231ef44ccc485f7f48e282efac5c70, 88,804,912 B (84.7 MiB), magic 16a5f792
139B has no rescue arm (RESCUE_STAGING §2 covers 139D / 139F / 13A1 / 13AA / 1399 / 13A3); proposed to qwen-gptq2-arch: an R139B margin-rescue arm (the WNS path is a shell GT gearbox, 1 path below +0.013).

18. ~18:48Z

19. 0xE4B71397 (ATT, asgard, record_be) checks (buildx02)

FINAL_TRUE +0.069 / +0.011; worst setup = shell CPM QDMA cips/cpm_0/qdma_1_wrapper_i/AXIST.u_demux int_demux_rst_n_2_reg -> ... lock_reg/D; worst hold TP link tp_x/u_hc/u_s1/u_lnk/u_mloc/u_ram RAMB36E5_INT -> FDRE clkout1_primitive +0.011, aurora +0.013 /mnt/data/eagle4_scratch/qg_buildwatch_20261116/1397_ft/
F1 srpol tooth 1 PASS mixed 0 / selftest PASS; tooth 2 PASS 1397_ft/GV_1397_srpol.txt
F11 enorm routed identical to 139A (sorted diff empty) 1397_ft/f11{a,b}.out
build-root PDI asgard:/shared/build/s3LAq297/aved_oneflip_p2s_head_asgard/hw/amd_v80_gen5x8_25.1_pdi23/amd_v80_gen5x8_25.1.pdi md5 a5c815e1d9f6a6752cc2ed6123e3c05a, 68,279,216 B (65.1 MiB), magic 16a5f792

20. ~19:14Z

21. 0xE4B71392 / 0xE4B71393 (MLPV L256 TREE_P 16; 1393 + FR_A8) STEP1 (19:45Z)

id LUT FF BRAM URAM DSP component.xml md5 util md5 A1/A3 TREE_P
1392 1,428,863 (-63,122 vs 139D TREE_P 32) 1,239,445 1,048 512 3,458 7eb3349a 6cad8b99 PASS component.xml MODELPARAM_VALUE.TREE_P 16; wrapper .TREE_P(16) x1, .LANES(256), .TREE_V_MAX(248320), .ARRAY_FAST_CLK(1'D0), .GEMV_SCALE_FRAC(16) x1 each; GATE_VERDICTS:10 [PASS] tailfast-pkgtcl-treep: set TREE_P 16 x1; derived pkg tcl :265 set TREE_P 16, :618 -generic TREE_P=$TREE_P
1393 1,429,056 1,239,405 1,048 512 3,458 6a4dd7bc 8e5975c7 PASS same (16 / .TREE_P(16) / tailfast-pkgtcl-treep PASS)
=> both are the TREE_P 16 hedge they claim to be (1399 = the TREE_P 32 control, §18). pblock_bcast_slr1 rows owed at init.

22. ~19:45Z

23. ~20:11Z

Clock note (20:13Z, flagged by qwen-gptq2-arch): section-header times from §14 on were estimated, not read from date -u, and ran 6-34 min ahead; headers §14/§15/§18/§20/§22/§23 corrected to the git commit time of each entry. Every timestamp INSIDE the entries is copied from a log/tsv and is unaffected. From here on headers carry date -u at probe time.

24. 20:37Z (date -u)

25. 21:03Z (date -u)

26. 21:10Z (date -u): 0xE4B713A2 (MLPV L128 hedge) build checks (buildx02)

FINAL_TRUE +0.004 / +0.010 (below the +0.030 bar; R13A2 rescue arm exists, see §23 note). Worst setup = shell aurora aur_link1 gtm_tx_gearbox_66_64_3 count_i -> data_in_d1_r[17] ch0_txoutclk_3 4 levels; next +0.015. Worst hold = drafter u_array g_grp[0].g_lane[2] lutg result_q usr_0 +0.010 SLR0->SLR1 /mnt/data/eagle4_scratch/qg_buildwatch_20261116/13A2_ft/
F1 srpol tooth 1 PASS mixed 0 / selftest PASS; tooth 2 PASS 13A2_ft/GV_13A2_srpol.txt
build PDI not written at 21:10Z (wdi since 20:36:53Z)

27. 21:14Z (date -u): 0xE4B71399 pblock rows (impl_1/runme.log pblockgate, xdc eagle4_drafter_bcast_slr1_pblock.xdc a07764c3)

28. 21:29Z (date -u)

29. 21:43Z (date -u): 0xE4B713A3 pblock rows (same xdc family as 1399)

pblock_qg_tree_slr2 NCELLS=2 (*g_tree_pwide.u_tree :102 OK 1, *g_tree_pwide.g_ser.u_cheap_ser :103 OK 1) = non-empty; pblock_bcast_slr1 660 (g_grp_pipe 322/177/161, tile_idx_p 560, stage_sel_p 90); also pblock_g8sorter_coloc_slr2 26,616, pblock_lmdown_coloc_slr1 6 hier (u_lmdown 191,290 cells, u_seam_r8 111,163), g1buf_tight 2,088. pblockgate 68 checked / 8 waived / 12 failed = the same inherited dead rows as 1399 (a_chunk_p :94 + empty pblock_r4_silu_slr1) -> adjudicated no effect (§27).

30. 21:56Z (date -u)

31. 22:22Z (date -u)

32. 22:48Z (date -u)

33. 23:14Z (date -u): rescue progress (^RESCUE_TIMING, latest line of each out/vivado_rescue_<TAG>.log)

arm step margined WNS / WHS true WNS / WHS log time
R139D p2_route 0.000 / +0.010 +0.040 / +0.022 (bar +0.030 / +0.010 met) 20:55:31
R139Bb p1_physopt +0.001 / 0.000 +0.041 / +0.012 (139B build was +0.002 / +0.010) 22:37:42
R13A2b p0_margin -0.036 / -0.002 (17 / 12 failing) +0.004 / +0.010 (= the build) 22:11:12

34. 23:32Z (date -u): 0xE4B71398 (ATT L256 emb v2 non-FR explore_fs) checks (buildx02) - CORRECTED 01:25Z: the FINAL_TRUE below is VOID, the design is NOT fully routed (see §39)

FINAL_TRUE +0.064 / +0.012; worst setup shell aurora aur_link gtm_tx_gearbox_66_64_3 count_i -> data_in_d1_r[50] ch0_txoutclk_2 4 levels; worst hold TP link tp_x/u_hc/u_s2/u_lnk ha3 -> sum4 clkout1_primitive +0.012 SLR0 /mnt/data/eagle4_scratch/qg_buildwatch_20261116/1398_ft/
F1 srpol tooth 1 PASS mixed 0 / selftest PASS; tooth 2 PASS 1398_ft/GV_1398_srpol.txt
F11 enorm routed identical to 139A 1398_ft/f11{a,b}.out
build PDI not written at 23:32Z (wdi since 23:03:51Z)

35. 23:40Z (date -u)

36. 00:06Z 2026-11-17 (date -u)

37. 00:32Z 2026-11-17 (date -u)

38. 00:59Z 2026-11-17 (date -u): R139D rescue PDI (0xE4B7139D margin rescue)

item value source
FINAL PDI buildx03:/shared/build/qg_rescue_20261116/R139D/out/amd_v80_gen5x8_25.1.pdi md5 1ae08166fbc9999f674c2a7a24305074, 96,520,576 B (92.05 MiB, inside 30-120), magic 16a5f792, AMC_IN_BAND_DBG strings 32 (= amc.elf 32), written 00:55:36Z (re-checked by me with stat / md5sum / xxd / strings) DRIVER.log FINAL PDI ..., GATE_VERDICTS_R139D.txt MAGIC PASS + AMC_IN_BAND_DBG PASS pdi=32 amc.elf=32
rescue FINAL_TRUE WNS +0.030 / WHS +0.022, nfail 0/0, MARGIN_OFF (00:12:19); full-precision re-measure owned by qwen-gptq2-arch out/vivado_rescue_R139D.log
final dcp out/R139D_final.dcp md5 5a603085 (copy on buildx02 md5-equal)
F1 srpol (rescue final dcp 5a603085, opt 97668675) tooth 1 PASS mixed=0 minority_loads=0 selftest=PASS (00:48:45Z); tooth 2 PASS OPTDIFF_SUMMARY changed=0 rep_diff=0 root_changed=0 legit_moved_inversion=0 reset_absorbed=59 candidates=59 raw_1to0=59 raw_0to1=0 common=1311453 new=1218 removed=0 unresolved=0 selftest=PASS (01:04:16Z) /mnt/data/eagle4_scratch/qg_buildwatch_20261116/R139D_ft/GV_R139D_srpol.txt ba6724a9
Other: R1374 FT_TRIGGER build FINAL_TRUE WNS=0.005 WHS=0.010 -> rescue 00:52:07Z (buildx04); 0xE4B713A1 write_device_image 00:57:24Z; rescue lines R139Bb / R13A2b / R13AA unchanged since §37.

39. 01:25Z 2026-11-17 (date -u): 0xE4B71398 DEAD at write_device_image - route incomplete

40. 01:52Z 2026-11-17 (date -u): rescues

arm state latest true WNS / WHS
R139D DONE: PDI 1ae08166, srpol x2 PASS (§38) FINAL_TRUE +0.030 / +0.022
R139Bb p2_route (00:14:35) +0.041 / +0.019
R13A2b p2_route (01:37:13) +0.089 / +0.016
R13AA p1_physopt (00:29:04) +0.040 / +0.011
R1374 started 01:27:07Z (buildx04), p0_margin 01:51:45 +0.005 / +0.010 (build)
R13A1 FT_TRIGGER 0.010 / 0.010 + QUALIFIED -> START 01:43:45Z (buildx03) -
R13A1b FT_TRIGGER 0.010 / 0.010 01:42:11Z (buildz03), not qualified yet -
R1398 armed + stopped by qwen-gptq2-arch (v3 NOT-NEEDED test reads FT only); R1398v4 (route-clean gate) pending -
Builds: 13A1 wdi since 00:57:24Z; 139F / 13AF / 1399 / 13A3 route Phase 5; 1392m / 1393m synth_1.

41. 02:05Z 2026-11-17 (date -u): route-clean re-derivation from the RAW route_status lines (fail-closed fix)

final_true.tcl (d9a6eb3fd) defaulted re 0, so an unparsed report read as clean (fail-OPEN, found by qwen-gptq2-arch's v4 worker); now set re -1 and ^FT_ROUTE_ERRORS <n> PASS|FAIL (PASS only at exactly 0). Raw report_route_status lines of every FINAL_TRUE I posted (<id>_route_status.rpt, same dcp copy as the FT):

id routable nets fully routed nets with routing errors FT verdict stands?
139A 1,784,616 1,784,616 0 yes
139B 2,510,333 2,510,333 0 yes (FT +0.002, below bar; R139Bb)
139C 1,785,082 1,785,082 0 yes
139E 1,786,114 1,786,114 0 yes
13A0 1,786,030 1,786,030 0 yes
13A2 2,509,851 2,509,851 0 yes (FT +0.004, below bar; R13A2b)
13A9 1,785,805 1,785,805 0 yes
1397 1,788,344 1,788,344 0 yes
1398 1,786,509 1,786,474 35 VOID (§39)
139D build FT came from the R139D waiter (route_status unrouted 0 in its FT_ROUTE_STATUS row); R139D rescue final = 0 of 2,932,083 (qwen-gptq2-arch). Rescue finals R139Bb / R13A2b / R13AA / R13A1 / R1374: raw route_status to be read as each lands. R13A1b STOPPED by qwen-gptq2-arch before admission (R13A1 = the one 13A1 rescue). R1398v4 armed on driver v4 (51985309, route-clean gate; waiter -> buildx01).

42. 02:18Z 2026-11-17 (date -u)

43. 02:45Z 2026-11-17 (date -u)

44. 03:00Z 2026-11-17 (date -u): R139Bb rescue PDI (0xE4B7139B MLPV L128 margin rescue)

item value source
FINAL PDI buildx03:/shared/build/qg_rescue_20261116/R139Bb/out/amd_v80_gen5x8_25.1.pdi md5 659231eef62bc6100190e24a3fc2dd39, 88,815,408 B (84.70 MiB), magic 16a5f792, AMC 32/32, 02:50:05Z (re-checked by me: stat / md5sum / xxd / strings) DRIVER.log FINAL PDI
rescue FINAL_TRUE +0.041 / +0.019, nfail 0/0 (02:14:28) vivado_rescue_R139Bb.log
route R139Bb_final_route_status.rpt: 2,510,364 routable / 2,510,364 fully routed / 0 routing errors
F1 srpol (final dcp b9f7dfe2, opt b1fce937) tooth 1 PASS mixed 0 / selftest PASS (02:45:26Z); tooth 2 PASS changed=0 rep_diff=0 root_changed=0 reset_absorbed=49 candidates=49 raw_1to0=48 common=1139457 new=1410 removed=3 unresolved=0 (02:58:56Z) /mnt/data/eagle4_scratch/qg_buildwatch_20261116/R139Bb_ft/GV_R139Bb_srpol.txt

45. 03:11Z 2026-11-17 (date -u)

46. 03:37Z 2026-11-17 (date -u)

47. 04:03Z 2026-11-17 (date -u)

48. 04:06Z 2026-11-17 (date -u): 0xE4B71399 tiler -> tree path (as-saved postroute_physopt copy 14a382aa, uncertainty 0; watch/tree_path.tcl)

49. 04:22Z 2026-11-17 (date -u): R13A2b rescue PDI (0xE4B713A2 MLPV L128 hedge margin rescue)

item value source
FINAL PDI buildy02:/shared/build/qg_rescue_20261116/R13A2b/out/amd_v80_gen5x8_25.1.pdi md5 9aa6b0d0b8df86a60065434b8f5b15ef, 86,776,928 B (82.76 MiB), magic 16a5f792, AMC 32/32, 04:10:31Z (re-checked) DRIVER.log
rescue FINAL_TRUE +0.091 / +0.016, nfail 0/0 (03:37:26) vivado_rescue_R13A2b.log
route 2,509,874 routable / 2,509,874 fully routed / 0 routing errors out/R13A2b_final_route_status.rpt
F1 srpol (final dcp eefe0fa6, opt a11da350) tooth 1 PASS mixed 0 / selftest PASS; tooth 2 PASS changed=0 rep_diff=0 root_changed=0 reset_absorbed=85 candidates=85 raw_1to0=78 common=1139459 new=1405 removed=0 unresolved=0 /mnt/data/eagle4_scratch/qg_buildwatch_20261116/R13A2b_ft/GV_R13A2b_srpol.txt

50. 04:29Z 2026-11-17 (date -u)

51. 04:55Z 2026-11-17 (date -u)

52. 05:00Z 2026-11-17 (git commit time; header had been estimated): 0xE4B713E5 (first margined FR TREE_P 32 MLPV tail, midyard, PGID 356346, root ~/build4c1/s3LCq2E5)

53. 05:04Z 2026-11-17 (date -u): 0xE4B713E5 build FINAL_TRUE (R13E5v3, driver v3, buildy03)

54. 05:21Z 2026-11-17 (date -u)

55. 05:40Z 2026-11-17 (date -u): 0xE4B713E5 (margined FR TREE_P 32 MLPV tail) - clean as built, no rescue

item value source
build-root PDI midyard:~/build4c1/s3LCq2E5/aved_oneflip_p2s_tail_asgard/hw/amd_v80_gen5x8_25.1_pdi23/amd_v80_gen5x8_25.1.pdi md5 76e0aff5c87611ac07806cbfd902d855, 96,945,872 B (92.45 MiB), magic 16a5f792, AMC_IN_BAND_DBG 32 stat / md5sum / xxd / strings
FINAL_TRUE (my re-measure, copy c3ae18fd == midyard) +0.052 / +0.022 (= R13E5v3); worst setup shell TP link tp_l4/u_l4/g_port[1].u_tf fifo rdp count -> RAM, ch0_txoutclk_4, 1 level (next +0.053 g_port[0]); worst hold drafter u_array g_grp[1].g_lane[69] lutg csa usr_0 +0.022 /mnt/data/eagle4_scratch/qg_buildwatch_20261116/13E5_ft/
route ^FT_ROUTE_ERRORS 0 PASS; raw: 2,935,617 routable / 2,935,617 fully routed / 0 routing errors 13E5_ft/13E5_route_status.rpt
kvc2 margin PASS 48 / 48 x2 / 48 (§52)
F1 srpol (opt 28f3ed31) tooth 1 PASS mixed 0 / selftest PASS; tooth 2 PASS changed=0 rep_diff=0 root_changed=0 reset_absorbed=21 candidates=21 raw_1to0=21 common=1311473 new=1350 removed=1 unresolved=0 13E5_ft/GV_13E5_srpol.txt

56. 06:05Z 2026-11-17 (date -u)

57. 06:15Z 2026-11-17 (date -u): 0xE4B713E5 preflash evidence (midyard, report-only, --dry-run; tools = box1 versions staged at midyard:~/build4c1/preflash_evidence_20261117/tool/gate_new: preflash_gate 2130df59, strapcomb coverage 9640a689, link_cdc b63f6804/86c42146, seam efficacy 738aa3dc; evidence ~/build4c1/preflash_evidence_20261117/0xE4B713E5/)

Inputs: PDI 76e0aff5, dcp impl_1/top_wrapper_postroute_physopt.dcp c3ae18fd, intent INTENT_CONSOL_TAIL_tail_ASGARD_s3LCq2E5.txt. Seam efficacy vivado 05:46:50-05:55:50Z (EXIT 0); link_cdc 05:55:50-06:06:15Z (EXIT 0).

row run 1 (strap 3903af07) run 2 (strap b3c748b0, 45d0c070b role-aware)
PDI-1 / 2 / 2b / 3 / 4 PASS (md5 76e0aff5, magic 16a5f792, 92 MB 96,945,872 B, 32 x AMC_IN_BAND_DBG) same
PDI-5 UNVER (QUARANTINED marker acknowledged, not cleared) same
PDI-6b / PDI-6 strap-legality FAIL 1 DESIGN (R6_ACCFIX_SEAM_P: Q5120 image: wrapper 1, want 0; STRAPCOV ce8d3ec5) PASS 0 DESIGN / 0 INSTR of 11 rows (R6_ACCFIX_SEAM_P: Q5120 MLPV (TAIL, PIPE_ROLE 2'd2): .ACCFIX_SEAM_P=1 is DORMANT ... 0 g_q5120_cfg_err; STRAPCOV_v4 79da8b1d)
PDI-7a census PASS (fail_rows 16, unverified-only 4 of 23) same
PDI-7 seam-cdc-efficacy FAIL 16 rows (SUMMARY pass=3 fail=15 unverified=4; file fbdb60fa) = the seam_xport selector class (this BD has hop_xport/tp_l4/tp_x, no seam_xport; same as 139C/13A9/1397) same
PDI-8 link-cdc FAIL pass=18 fail=2 superseded=2 (file e3edeb91): hop_aurora_cdc.xdc:51 and :54 set_max_delay SELECTOR VACUOUS (0 cells both ends); :45/:48 SUPERSEDED by att4_link.xdc:61/:55 (slack 11.485 / 13.957) = the same counts as 139C same
RESULT FAIL (IMAGE DEFECT), rc 1 (out bcd720cd) FAIL (IMAGE DEFECT), rc 1 (out c4dae4cc): PDI-6 now PASS; remaining FAILs = PDI-7 + PDI-8, the classes adjudicated on 139C

58. 06:31Z 2026-11-17 (date -u)

59. 06:58Z 2026-11-17 (date -u)

60. 07:23Z 2026-11-17 (date -u): R13AA rescue PDI (0xE4B713AA MLPV L256 + FR_A8 explore, margin rescue)

item value source
FINAL PDI buildx03:/shared/build/qg_rescue_20261116/R13AA/out/amd_v80_gen5x8_25.1.pdi md5 8c2e6a5a63280d2f41db7c178a03835a, 97,482,768 B (92.97 MiB), magic 16a5f792, AMC 32/32, 07:07:57Z (re-checked) DRIVER.log
rescue FINAL_TRUE +0.041 / +0.022, nfail 0/0 (06:26:54) vivado_rescue_R13AA.log
route 2,940,308 routable / fully routed / 0 errors out/R13AA_final_route_status.rpt
F1 srpol (final dcp 677548b0, opt from R13AA/opt) tooth 1 PASS mixed 0 / selftest PASS; tooth 2 PASS changed=0 rep_diff=0 root_changed=0 reset_absorbed=263 candidates=263 raw_1to0=254 common=1311468 new=1296 removed=2 unresolved=0 /mnt/data/eagle4_scratch/qg_buildwatch_20261116/R13AA_ft/GV_R13AA_srpol.txt

61. 07:24Z 2026-11-17 (date -u)

62. 07:52Z 2026-11-17 (date -u)

63. 08:17Z 2026-11-17 (date -u)

64. 08:39Z 2026-11-17 (date -u): Qwen silicon WINDOW 2 started (host levers L1/L3)

64. 08:43Z 2026-11-17 (date -u)

65. 09:40Z 2026-11-17 (date -u): 0xE4B713AF route FAILED legality

66. 10:11Z 2026-11-17 (date -u)

65. 10:42Z 2026-11-17 (date -u): Qwen silicon WINDOW 2 serve results (139C x4 + 139D x2, kit 19466e586, golden md7532 64/64 fp e5718cec)

67. 12:45Z 2026-11-17 (date -u): R139F rescue PDI (0xE4B7139F MLPV L256 hedge margin rescue)

item value
FINAL PDI buildx03:/shared/build/qg_rescue_20261116/R139F/out/amd_v80_gen5x8_25.1.pdi md5 2fc631326b01c182fa580576606d8d3f, 99,436,992 B (94.83 MiB), magic 16a5f792, AMC 32/32, 12:33:41Z (re-checked)
rescue FINAL_TRUE +0.041 / +0.022, nfail 0/0 (11:49:51)
route 2,932,752 routable / fully routed / 0 errors
F1 srpol (final dcp c05cdfa8) tooth 1 PASS mixed 0 / selftest PASS; tooth 2 PASS changed=0 rep_diff=0 root_changed=0 reset_absorbed=16 candidates=16 raw_1to0=16 common=1311446 new=1005 removed=3 unresolved=0 (/mnt/data/eagle4_scratch/qg_buildwatch_20261116/R139F_ft/GV_R139F_srpol.txt)

68. 14:16Z 2026-11-17 (date -u): 1392m / 1393m (margined MLPV L256 TREE_P 16; 1393m + FR_A8) build checks (buildx04, copies md5-equal)

id FINAL_TRUE route (raw) srpol t1 srpol t2 worst setup
0xE4B71392m +0.051 / +0.022 (^FT_ROUTE_ERRORS 0 PASS) 2,874,050 / 2,874,050 / 0 PASS mixed 0 PASS changed 0 root_changed 0 (74 reset_absorbed) shell TP link tp_l4 g_port[0].u_tf fifo rdp -> RAM
0xE4B71393m +0.054 / +0.022 (^FT_ROUTE_ERRORS 0 PASS) 2,884,633 / 2,884,633 / 0 PASS mixed 0 PASS changed 0 root_changed 0 (13 reset_absorbed) shell aurora aur_link1 txheader -> gtm_tx_gearbox_66_64_3, ch0_txoutclk_3
Copies: 1392m postroute 3ce3ac69 / opt a118935d; 1393m postroute 26b83f04 / opt 2abacfdc. Evidence /mnt/data/eagle4_scratch/qg_buildwatch_20261116/{1392m,1393m}_ft/. => both clean as built (no rescue needed), like 13E5. Build PDIs not yet written at 14:16Z.

69. 15:21Z 2026-11-17 (date -u): R13A1 rescue PDI (0xE4B713A1 MLPV L256 + FR_A8 margin rescue)

item value
FINAL PDI buildx03:/shared/build/qg_rescue_20261116/R13A1/out/amd_v80_gen5x8_25.1.pdi md5 681ca21408bfa5b2b467aed4c7df697f, 99,608,496 B (94.99 MiB), magic 16a5f792, AMC 32/32, 15:02:16Z (re-checked)
rescue FINAL_TRUE +0.043 / +0.022, nfail 0/0 (14:19:27)
route 2,940,040 routable / fully routed / 0 errors
F1 srpol (final dcp b706baa0) tooth 1 PASS mixed 0 / selftest PASS; tooth 2 PASS changed=0 rep_diff=0 root_changed=0 reset_absorbed=77 candidates=77 raw_1to0=77 common=1311478 new=1565 removed=0 unresolved=0 (/mnt/data/eagle4_scratch/qg_buildwatch_20261116/R13A1_ft/GV_R13A1_srpol.txt)

70. 15:52Z 2026-11-17 (date -u): R1399x2 rescue PDI (0xE4B71399 MLPV L256 TREE_P 32, tree SLR2 floorplan, margin rescue, driver v4)

item value
FINAL PDI buildx02:/shared/build/qg_rescue_20261116/R1399x2/out/amd_v80_gen5x8_25.1.pdi md5 37bcfc1666281dc187c89ef5c8f8e0ea, 95,323,872 B (90.91 MiB), magic 16a5f792, AMC 32/32, 15:41:33Z (re-checked)
rescue FINAL_TRUE +0.042 / +0.022 (13:45:38), re-verified on the saved final 14:51Z (AS_LOADED == FINAL_TRUE)
route ^RESCUE_ROUTE_STATUS FINAL routing_errors=0 partial=0 unrouted_or_conflict=0 (14:22:44Z)
F1 srpol (driver v4, final dcp 9a94b3f2) tooth 1 PASS (15:11:16Z); tooth 2 PASS (15:37:52Z); SRPOL_DONE tooth1 PASS tooth2 PASS (waiter.log)

71. 17:26Z 2026-11-17 (date -u): R13A3 rescue PDI (0xE4B713A3 MLPV floorplan-fixed hedge, margin rescue)

item value
FINAL PDI buildx03:/shared/build/qg_rescue_20261116/R13A3/out/amd_v80_gen5x8_25.1.pdi md5 a8cad6cd19cee5044b7d269cb9072e77, 100,159,168 B (95.52 MiB), magic 16a5f792, AMC 32/32, 17:15:16Z (re-checked)
rescue FINAL_TRUE +0.047 / +0.023, nfail 0/0 (16:37:28)
route 2,939,181 routable / fully routed / 0 errors
F1 srpol (final dcp 20b5a5b1) tooth 1 PASS mixed 0 / selftest PASS (17:10:18Z); tooth 2 PASS changed=0 rep_diff=0 root_changed=0 reset_absorbed=14 candidates=14 raw_1to0=14 common=1311469 new=1360 removed=1 unresolved=0 (17:25:19Z) (/mnt/data/eagle4_scratch/qg_buildwatch_20261116/R13A3_ft/GV_R13A3_srpol.txt)

72. FINAL SUMMARY 17:26Z 2026-11-17 (date -u): every image has a PDI or a verdict

Bar = FINAL_TRUE >= +0.030 / >= +0.010 with routing errors 0, F1 srpol both teeth PASS, magic 16a5f792, AMC 32. "admissible by my checks" = all my rows green; ship/fallback class is qwen-gptq2-arch's adjudication.

image role / variant PDI (md5) how FINAL_TRUE route err srpol my checks
0xE4B7139A ATT L256 c537f6ab (build) as built +0.109 / +0.010 0 PASS x2 green (hold at bar)
0xE4B7139C ATT L256 + emb v2 95634f34 (build) as built +0.124 / +0.012 0 PASS x2 green
0xE4B7139E ATT L256 hedge 723ab73b (build) as built +0.057 / +0.012 0 PASS x2 green
0xE4B713A0 ATT L256 + FR_A8 b5db7626 (build) as built +0.112 / +0.011 0 PASS x2 green
0xE4B713A9 ATT L256 + FR_A8 + emb v2 1d269f64 (build) as built +0.072 / +0.012 0 PASS x2 green
0xE4B71397 ATT L256 record_be a5c815e1 (build) as built +0.069 / +0.011 0 PASS x2 green
0xE4B71398 ATT L256 emb non-FR none DEAD at wdi void 35 (on dirty dcp) not admissible (route DRC; R1398v4 did not clear; ECO stopped, resume point eco_placed.dcp)
0xE4B713A6 / A7 / A8 / AE / AF ATT L512 none A7 dead at place (URAM SLR2 682 > 624), A6 / A8 / AE killed by arch, AF route not legal (152,152 overlaps) - - - closed
0xE4B713B0 ATT none killed by arch (id collision) - - - closed
0xE4B7139B MLPV L128 659231ee (R139Bb) rescue +0.041 / +0.019 0 PASS x2 green
0xE4B7139D MLPV L256 1ae08166 (R139D) rescue +0.030 / +0.022 0 PASS x2 green (setup at bar)
0xE4B7139F MLPV L256 hedge 2fc63132 (R139F) rescue +0.041 / +0.022 0 PASS x2 green
0xE4B713A1 MLPV L256 + FR_A8 681ca214 (R13A1) rescue +0.043 / +0.022 0 PASS x2 green
0xE4B713A2 MLPV L128 hedge 9aa6b0d0 (R13A2b) rescue +0.091 / +0.016 0 PASS x2 green
0xE4B713AA MLPV L256 + FR_A8 explore 8c2e6a5a (R13AA) rescue +0.041 / +0.022 0 PASS x2 green
0xE4B71374 MLPV L256 TREE_P 16 7ada84d1 (R1374) rescue +0.044 / +0.022 0 PASS x2 green
0xE4B71399 MLPV L256 TREE_P 32, tree SLR2 37bcfc16 (R1399x2) rescue +0.042 / +0.022 0 PASS x2 (driver v4) green
0xE4B713A3 MLPV floorplan-fixed hedge a8cad6cd (R13A3) rescue +0.047 / +0.023 0 PASS x2 green
0xE4B713E5 MLPV margined FR TREE_P 32 76e0aff5 (build) as built +0.052 / +0.022 0 PASS x2 green (+ preflash PDI-1..6 PASS, PDI-7/8 = 139C classes)
0xE4B71392m MLPV margined L256 TREE_P 16 f88e3619 (build) as built +0.051 / +0.022 0 PASS x2 green
0xE4B71393m MLPV margined L256 TREE_P 16 + FR_A8 33f0932f (build) as built +0.054 / +0.022 0 PASS x2 green

74. ARRAY_FAST_CLK=1 (AFC) MLPV wave: build watch (numbered 74; the launch record is sec. 73 below) (qgptq2-afc-build-watch, opened 18:04:48Z 2026-11-17, date -u)

AFC = MLPV u_array (LUTG A16 lanes) on ap_clk_fast = usr_clk_wiz clk_out2 = 2 x 66 = 132 MHz (derive --array-fast-clk, 4303c0fb6). Bar = the 0x13E5 / 0x1392m / 0x1393m one (in-build margin, no rescue): FINAL_TRUE >= +0.030 / +0.010 at margin 0 with route errors 0, kvc2 margin coverage PASS, srpol t1+t2 PASS, magic 16a5f792, AMC 32. Report-only; nothing here kills a build. New tool: scripts/4c1/qgptq2/watch/final_true_afc.tcl (= final_true.tcl + ^AFC_XING per-direction worst setup/hold + negative-path count, report_clock_interaction, report_cdc 66<->132).

image cfg host root PGID launched (STAGE_TIMES pkg_synth start)
0xE4B7138D 1399 + AFC (non-FR, TREE_P 32, tree SLR2), arm balanced buildy04 /shared/build/s3LCq28D 2608980 18:02:38Z
0xE4B7134B 13A3 + AFC (FR_A8 CFG_FR_A8_V=1'b1), arm default buildz02 /shared/build/s3LCq24B 3124987 18:02:39Z
0xE4B7131F 138D hedge (same cfg, arm default) buildy05 /shared/build/s3LCq21F 2388268 18:02:36Z (pipeline pid start)
0xE4B71380 134B hedge (id-swapped from 13B0 by qwen-gptq2-arch; FR_A8 + AFC), arm balanced buildz04 /shared/build/s3LCq280 1713742 18:05:43Z (pkg_synth end 18:07:19Z)
Straps read from each stage's QG_STRAPS_EFFECTIVE.env: CFG_ARRAY_FAST_CLK_V=1'b1, CFG_LANES_V=256, LUTG 1 / LUTG_AW 16 / A8Q 0 on all four; 134B and 1380 add CFG_FR_A8_V=1'b1. Auditor rows added per build (qwen-gptq2-arch 18:1xZ): (a) wrapper .ARRAY_FAST_CLK(1'D1) + top.v ap_clk_fast = clock_reset_clk_usr_1 = usr_clk_wiz clk_out2, XCI MULT 205.921875 / DIVCLK 2 / O1 52 / O2 26; (b) sim netlist bridge + g_frst + mem_[eo]_reg PER g_lc_grp[0..2] (per-group, never a total); (c) routed intra usr_1 > 100K endpoints @ ~7.57 ns; (d) both inter-clock directions Timed + met, bands from M1361 (usr_0->usr_1 1,964 ep / +0.175; usr_1->usr_0 12,360 / +0.256; usr_0->usr_1 >> 3K = raw a_chunk still on the cluster); (e) bar rows; impl [loadpath-mcp] applied (info). Probe scratchpad/afc/afc_probe.sh (auditor's, read-only, reviewed). AFC=0 control: buildz01 s3LCq299.

73. 18:06Z 2026-11-17 (date -u): ARRAY_FAST_CLK 1 MLPV wave LAUNCHED (STAGING sec.24)

id host root PGID start config / arm
0xE4B7138D buildy04 /shared/build/s3LCq28D 2608980 18:02:34Z 1399 cfg + AFC + margin, balanced
0xE4B7134B buildz02 /shared/build/s3LCq24B 3124987 18:02:35Z 13A3 cfg (FR_A8) + AFC + margin, default
0xE4B7131F buildy05 /shared/build/s3LCq21F 2388268 18:02:36Z = 138D, default
0xE4B71380 buildz04 /shared/build/s3LCq280 1713742 18:05:39Z = 134B, balanced (replaces the coordinator-held 0x13B0 label)

74.1 20:08:45Z 2026-11-17 (date -u): STEP1 package + BD wrapper rows (a), all four AFC builds

pkg_synth (STAGE_TIMES): 138D 18:02:38-19:45:40Z (6,182 s, pgroup peak 52,214 MB), 131F 18:02:40-19:43:59Z (6,079 s, 52,915 MB); 134B / 1380 component.xml present by 19:51Z. STEP3 [FAIL] rows on 131F (csa_tree16_nw / gemv_lane_scaled_stream_dual(_ext)_opt NOT in staged src) = the LUTG-replaced modules, BY DESIGN (same rows on 139D, sec. above).

image wrapper md5 .ARRAY_FAST_CLK other wrapper literals (x1 each) ingested ipshared/*/src/eagle4_config.svh top.v usr_clk_wiz XCI (a)
0xE4B7138D 7716fefb 1'D1 LANES 256, TREE_P 32, CTX_WIN 768, TREE_V_MAX 248320, GEMV_SCALE_FRAC 16, MLPV_G8_OWN 1 8564/ d71b1bd2: ARRAY_FAST_CLK_V 1'b1, ARRAY_FAST_RST_SYNC_V 1'b1, LUTG 1, LUTG_AW 16, A8Q 0, FR_A8_V 1'b0 .ap_clk(clock_reset_clk_usr_0) x1, .ap_clk_fast(clock_reset_clk_usr_1), .clk_out2(clk_usr_1) PRIM_IN 33.333, MULT 205.921875, DIVCLK 2, O1 52, O2 26 (= 66.0006 / 132.0012 MHz, exactly 2:1), REQUESTED 66,132, USED true,true PASS
0xE4B7134B 7716fefb 1'D1 same 50fb/ e62ab05e: same + FR_A8_V 1'b1 same same PASS
0xE4B7131F 7716fefb 1'D1 same 17bc/ d71b1bd2: = 138D, FR_A8_V 1'b0 same same PASS
0xE4B71380 7716fefb 1'D1 same e48a/ e62ab05e: = 134B, FR_A8_V 1'b1 same same PASS
control 0xE4B71399 (buildz01 s3LCq299) 9f4a8e0e 1'D0 identical set 05fa/ 8a6dc013: ARRAY_FAST_CLK_V 1'b0, rest = 138D - - (AFC=0 reference)
Wrapper port annotation: ap_clk_fast FREQ_HZ 131999881 (INFO: its CLK_DOMAIN string reads top_usr_clk_wiz_0_clk_out1, a BD domain label; the routed clock-interaction row (d) is the real check). Rows (b) sim-netlist per-group LCACHE, (c)/(d) routed clock rows, (e) bar: owed at OOC synth end (~21:30Z) / route (~05:00Z 11-18).

74.2 22:08:02Z 2026-11-17 (date -u): drafter OOC sim netlist rows (b), all four AFC builds + the AFC=0 control

Instrument scratchpad/afc/lcgrp.sh (read-only): counts INSTANCES by primitive (the line naming the primitive precedes the \g_lcache.g_lc_grp[g].g_bram.mem_[eo]_reg* instance line), per group, never a total. (The auditor probe's RAMB_lines column greps the instance-name line for "RAMB" and reads 0 by construction, and its lc_bram_lines counts wires: both annotated, not used.) File complete = endmodule in the last 200 B and size stable over 3 min.

image sim netlist (prj.gen, size / mtime) g_lc_grp[0] e / o [1] e / o [2] e / o gemv_dualclk_bridge module / g_dc.u_dc_bridge inst g_frst.frst_q pipeline p2-netlist AFC teeth (b)
0xE4B7138D 1,127,364,746 B / 21:57:50Z 4 / 4 RAMB36E5_INT 4 / 4 4 / 4 1 / 1 8 bridge, g_lc_grp[0..2], FRST PASS (21:58:25) PASS
0xE4B7134B 1,126,738,203 B / 21:50:35Z 4 / 4 4 / 4 4 / 4 1 / 1 8 PASS (21:50:12) PASS
0xE4B7131F 1,127,362,776 B / 21:31:12Z 4 / 4 4 / 4 4 / 4 1 / 1 8 PASS (21:30:59) PASS
0xE4B71380 1,126,730,811 B / 21:37:00Z 4 / 4 4 / 4 4 / 4 1 / 1 8 PASS (21:37:13) PASS
control 0x1399 s3LCq299 1,168,524,420 B / 20:13:26Z none none none 0 / 0 0 - (AFC=0 reference)
Pipeline GATE_VERDICTS at 22:07Z: pass 101 / fail 16 on each of the four, and the 16 FAIL rows (slot-normalised) are a SUBSET of the control's 39 FAIL rows on its finished build (no new FAIL class; the 23 control-only rows are later-stage / link rows not yet run). Also DCCHECK-ARRAY_FAST_CLK-present, QG-DCCHECK-ARRAY_FAST_CLK-1, DCCHECK-ap_clk_fast-FREQ_HZ, BRIDGEPIPECHECK-census-present PASS on all four. Rows (c) intra usr_1 endpoints, (d) inter-clock bands, (e) bar and [loadpath-mcp] owed at route.

75. ATT_ARRAY_FAST_CLK (ATT AFC) wave: 0xE4B7138A / 0xE4B7138C / 0xE4B7138F (qwen-gptq2-arch build agent; STAGING sec.25; commit 84b853dbc)

AFC = the Q5120 ATT u_array (LUTG A16 lanes, G4 / G6 / G7g / G7u) on ap_clk_fast = usr_clk_wiz clk_out2 = 132 MHz (derive --att-array-fast-clk), in-build margin 0.050:0.012 on the margin_att_d1 clock set (incl. clock_reset_clk_usr_1). Report-only watcher: scripts/4c1/qgptq2/watch/watch_att_afc.sh <SLOT> [kv_mrow] (pkg A1/A3 + PARAM_VALUE, ingested header straps, top.v ap_clk_fast / clk_out2, clk_wiz XCI M/D/O + 2:1 verdict, sim-netlist bridge / g_frst / g_lc_grp[0..2], routed usr_1 rows). Bar = the ATT L256 one: FINAL_TRUE >= +0.030 / +0.010 at margin 0, route errors 0, srpol t1+t2 PASS, magic 16a5f792, AMC 32.

image cfg host root PGID launched
0xE4B7138A 139C + ATT AFC (non-FR, emb v2), arm default buildz01 /shared/build/s3LAq2FC8A 2913217 23:21:34Z
0xE4B7138C 131C (c5 KV_MROW) + ATT AFC, arm default buildy01 /shared/build/s3LAq2FC8C 3743166 23:21:35Z
0xE4B7138F 13A9 (FR_A8, emb v2) + ATT AFC, arm default (FR hedge) buildz05 /shared/build/s3LAq2FC8F 2677263 23:21:37Z

74.3 23:25:08Z 2026-11-17 (date -u): ATT AFC wave added to this watch (qwen-gptq2-arch 23:1xZ) + MLPV wave progress

image recipe host root (clone aved_oneflip_p2s_head_asgard/hw/amd_v80_gen5x8_25.1_pdi23) PGID pipeline start (ps lstart) / pkg_synth start
0xE4B7138A 139C + ATT AFC buildz01 /shared/build/s3LAq2FC8A (stage qgptq2_att_stage_138A_afc) 2913217 23:21:34Z / 23:21:38Z
0xE4B7138C 138A + KV_MROW_V 1 (c5) buildy01 /shared/build/s3LAq2FC8C (qgptq2_att_stage_138C_afc) 3743166 23:21:35Z / 23:21:40Z
0xE4B7138F 13A9 FR_A8 + ATT AFC buildz05 /shared/build/s3LAq2FC8F (qgptq2_att_stage_138F_afc) 2677263 23:21:37Z / 23:21:41Z
All three carry in-build margin 0.050:0.012. Auditor rows for ATT (a)-(m) per qwen-gptq2-arch (STEP1 must_pass --kind ATT, ingested ATT_ARRAY_FAST_CLK_V 1'b1 / ATT_FASTCLK_RED_V 0 / ARRAY_FAST_CLK_V 1'b0, bridge md5 0395872d / array c6c4caf9, clk binding + XCI O3 12 REQ {66,132,286}, per-group g_lc_grp RAMB + LUTG a_r fan-in, routed periods, intra/inter usr_1 rows + report_cdc, SLR census vs 139C, margin xdc d37f7c01 non-vacuous + cancelled, route/FINAL_TRUE/srpol, PDI).
MLPV wave at 23:24Z: all four past opt_design (138D / 134B Phase 2 Floorplanning, 131F / 1380 Post FP PSIP Lite).
PGID note (23:27:46Z): the build process group is the setsid'd run_* shell (138A 2913217 = the qg_stage_timer pgid arg); the build agent's watcher's pgid 2913213 is the outer launcher subshell (ppid 1). Both alive.
xsim flash-condition arms (buildx01, handed to this watch by qwen-gptq2-arch 23:2xZ): running at 23:27Z = AFC1_S5 / AFC1_G1 / AFC1_FA3 (12 item-ENDs each), CTL_RR / FCTL_RR / C5CTL_RR (4), AFC1_H_A (4), AFC1_RR / AFC1_FRR / AFC1_C5RR / AFCRL_ROOT (in boot, 0). Harvest = afc_pull_check.sh + qg_afc_ab.py; flash condition = all items GREEN and LCCHASE END 0 violations. SLR census v3 (buildz03, 139C placed dcp) running since 22:27Z, census3.out 22:33:52Z.

74.4 00:15:52Z 2026-11-18 (date -u): ATT AFC rows (a) STEP1, (b) ingested config, (c) clock binding: 3/3 PASS

image (a) must_pass --pkg (build agent watcher, watch/<id>.log 00:0xZ) ingested ipshared/*/src/eagle4_config.svh (md5) bridge / array in ipshared top.v usr_clk_wiz XCI verdict
0xE4B7138A MUSTPASS VERDICT PASS (component.xml 23:58:17Z) c00e/ c6747209: ATT_ARRAY_FAST_CLK_V 1'b1 x1, ATT_FASTCLK_RED_V 0 x1, ARRAY_FAST_CLK_V 1'b0 x1, KV_MROW_V absent, FR_A8_V 1'b0, LUTG 1 / LUTG_AW 16 / A8Q 0, LANES_V 256 gemv_array_dualclk.sv 0395872d, gemv_stage_array_scaled_slr.sv c6c4caf9 .ap_clk_fast(clock_reset_clk_usr_1) x1; clock_reset_imp: usr_clk_wiz .clk_out2(clk_usr_1) -> port clk_usr_1 -> .clk_usr_1(clock_reset_clk_usr_1) (:4645) M 205.921875 / D 2 / O 52 / 26 / 12, REQ 66,132,286, USED true x3 PASS
0xE4B7138C PASS (00:02:18Z) 7c23/ 1fc48241: same + KV_MROW_V 1'b1 x1 0395872d / c6c4caf9 same same PASS
0xE4B7138F PASS (00:01:27Z) 1030/ 5de792f4: same, KV_MROW absent, FR_A8_V 1'b1 0395872d / c6c4caf9 same same PASS
INFO: component.xml PARAM_VALUE.ATT_ARRAY_FAST_CLK = ABSENT on all three (header-localparam class by design, STAGING sec.25: no wrapper literal; the shipped-layer evidence is the ingested header above + netlist + routed clocks). Pipeline gate rows at 00:0xZ: QGAFC-clkwiz-request PASS x3.

74.5 00:29:07Z 2026-11-18 (date -u): xsim F10 RED AFCRL_ROOT did NOT fail (vacuous RED: redundant reload trigger)

74.6 00:40:25Z 2026-11-18 (date -u): xsim re-arm A/B pairs AFC1_RR vs CTL_RR, AFC1_FRR vs FCTL_RR: PASS (with a stated limit)

Snapshots box1 afc_att_20261117/snap/{AFC1_RR,CTL_RR,AFC1_FRR,FCTL_RR}_003947 (log md5 0ebb4fb1 / 8a05fa9d / d93a6b4e / 1aac3671; 12 item-ENDs + $finish each). A/B qg_afc_ab.py -> snap/AB_<a>_vs_<b>_003947.txt.

pair D9R / card KVWR TAPG4 / G5 / G6 TAPR4 seams VERDICT LCCHASE END (AFC arm)
AFC1_RR vs CTL_RR (qgafc1 vs qgl256, stim_root_rr3) 134 x4 identical 510 x4 30 / 2 / 40 x4 1,024 x4 2 / 2 identical x4 PASS (D9R 536) 12 x violations=0
AFC1_FRR vs FCTL_RR (qgfafc1 vs qgfra8, stim_fa3root_rr3) same counts, identical 510 30 / 2 / 40 1,024 2 / 2 identical PASS (D9R 536) 12 x violations=0
AFC1_RR ITEMMON (66 MHz design cycles, re-armed item): G4 2,584 / G6 883 / G7G 2,736 / G7U 1,384 (= the root AFC numbers in STAGING sec.25).
Limit (annotated, not bent): both re-arm items replay the SAME root, so item_01 and item_03 seams are md5-identical (0a9156c4 on c0/c1 of both arms) and the per-item equality cannot distinguish "re-armed item computed fresh" from "re-armed item reproduced the prior item's state"; it does prove the re-arm path (W 154 1, doorbell seq continuation, POLL 184, output slot) completes and matches the shipped RTL on the 2nd request. INFO: the FR_A8 pair's root seam md5 equals the non-FR pair's (FR_A8 only changes the child a-bank grid; this root-only stimulus has no children).

74.7 00:51:31Z 2026-11-18 (date -u): xsim AFC1_G1 (GPU req00 ctx 112, root + L1 0-3) FINISHED: GREEN

Snapshot snap/AFC1_G1_005101 (log md5 464c550d, 24 item-ENDs = 6 items x 4 ranks, $finish 1). Checker vs ref/gpu_round0/vectors_gpu_req00_kc512_md2 (CHK_AFC1_G1_005101.txt): AFC_VERDICT PASS (forward items 5, 19 stream rows); every checked stream 0 on c0..c3 for items 1:L0[0], 2:L1[0], 3:L1[1], 4:L1[2], 5:L1[3] (g4_y48 / D9R 19,200, kv_k / kv_v / g5_out 6,400, g6_part / g6_sum48 / h1 25,600, g7g / g7u / r4g / r4u / r5 20,480; kvload 142,080); g1_x / g6_in / pn(g7*_in) mismatches = the D9V array-port INFO class (lanes read LCACHE at AFC 1). A/B vs shipped L256_GPU00_MD2 (prefix): VERDICT PASS, D9R 1,340, 5 / 5 seams identical per card (snap/AB_AFC1_G1_vs_L256_GPU00_MD2_005101.txt). LCCHASE END 12 x violations=0. => flash-condition arm AFC1_G1 MET.

74.8 01:02:38Z 2026-11-18 (date -u): xsim c5 AFC1_H_A (8-row + 1-row KV-only items with on-card MROW reload, then root) FINISHED: GREEN

Snapshot snap/AFC1_H_A_010215 (log md5 70832305, 20 item-ENDs = 5 items x 4 ranks, $finish 1). c5_check.py vs ref a16_kc512 + the ordered KVWR stream of the shipped serial control SER_SA (CHK_AFC1_H_A_010215.txt): VERDICT PASS -- kv_final_K / kv_final_V 0/62,720, kv_root 0/2,560, h1 0/5,120, r5 0/4,096, stream_vs_ctl 0/500 on c0..c3 (KVWR 500 per card both). A/B vs shipped H_A (snap/AB_AFC1_H_A_vs_H_A_010215.txt): VERDICT PASS D9R 808, every class + seams identical. LCCHASE END 12 x violations=0. ITEMMON: 8-row KV-only item G4 20,672 + MROWLD 37,037; 1-row G4 2,584; root G4 2,584 / G6 883 / G7G 2,736 / G7U 1,384. => flash-condition arm AFC1_H_A MET.

74.9 01:45:12Z 2026-11-18 (date -u): ATT AFC rows (d) per-group LCACHE RAMB, (e) LUTG a_r fan-in, (f) bridge / g_frst / stubs: 3/3 PASS

(d)/(f) instrument scratchpad/afc/lcgrp.sh on the drafter OOC sim netlist (prj.gen; 138A 593,575,629 B 00:56:54Z / 138C 588,531,834 B 01:00:34Z / 138F 589,612,028 B 01:09:17Z, endmodule tail, size stable 3 min). (e) scripts/4c1/qgptq2/watch/afc_fanin.tcl on md5-checked COPIES of the OOC drafter dcp (138A 71524ce6 / 138C 1e0a4cfe / 138F 409b8b1c) on buildx04 /shared/scratch/qg_afc_fanin_20261118 (RC 0 x3; a first launch failed on a copy-name typo, kept in try1/).

image g_lc_grp[0] mem_e / mem_o [1] [2] gemv_dualclk_bridge module / u_dc_bridge inst g_frst.frst_q ERROR_ stubs u_a8q / lutg_a8q (e) a_r fan-in RAMB by group (2 samples / group: g_cl[0].a_r_reg[0], [100]) verdict
0xE4B7138A 4 / 4 RAMB36E5_INT 4 / 4 4 / 4 1 / 1 8 lines 0 0 grp0: lc0=1 lc1=0 lc2=0 (1 startpoint); grp1: lc1=1 only (+ g_rst_pipe rq); grp2: lc2=1 only (+ rq) PASS
0xE4B7138C 4 / 4 4 / 4 4 / 4 1 / 1 8 0 0 identical shape PASS
0xE4B7138F 4 / 4 4 / 4 4 / 4 1 / 1 8 0 0 identical shape PASS
Vivado per-group BLOCKRAM cells under g_lc_grp[g].g_bram.mem_*: n=8 per group on all three (= 4 e + 4 o). Each sampled cluster register's ONLY data startpoint is its own group's cache RAMB: the raw a_chunk / slow seam is not in the cone (the MLPV 4303c0fb6 / RED AFCRC defect class is absent in the built netlist). Limit: 2 of the cluster bits per group sampled, not every a_r bit.

75.3 ATT AFC cache-refill coverage: F10 follow-up (the RED was masked), RED2 / RED3, per-transition table (qwen-gptq2-arch build agent, 2026-11-18)

Why: build-watch (74.5, 208890787) found that F10 AFCRL_ROOT (bk_stage_start dropped from act_reload) PASSES. Confirmed [M]: AB_AFCRL_ROOT_vs_AFC0_ROOT.txt 28/28 PASS, LCCHASE 0. The mechanism [R]: the bridge's need_ld = reload_pend || act_reload || !cache_ok || (act_grp(stage_sel) != cache_grp) (gemv_array_dualclk.sv 0395872d :221) refills on every GROUP change, and G4 -> G6 -> G7g changes the group at every boundary. So F10 was a redundant-trigger RED (memory redundant-trigger-masks-red-arm): it says nothing about the GREEN arm, and it proves nothing about the instrument.

Transition table [R] (staged 138C sources: drafter 7f9adce9, top_fsm e252b32b, bridge 0395872d; the 138A / 138F drafters carry the same lines).

transition (ATT item / tree / re-arm) stage_sel -> act_grp group term fires? act_reload term other same activation? sole guard
reset -> first G4 0 -> 0 (cache_grp F) yes (F != 0) g4_start !cache_ok - (3 terms)
G1 -> G4 (root / child item) after a full item (prev G7u) 0 -> 0 (cache_grp 2) yes g4_start new redundant
G4 -> G5 -> G6 1 -> 1 (cache_grp 0) yes bk_stage_start (g6_start) new (G5 out) redundant
G6 tile re-issue (G6_REISSUE, att_g6_tp replay) 1 -> 1 no no new stage entry same G6 seam none needed
G6 -> (R3, premlp) -> G7g 2 -> 2 (cache_grp 1) yes bk_stage_start (g7g_start) new (premlp) redundant
G7g -> G7u 3 -> 2 (cache_grp 2) no bk_stage_start (g7u_start) same premlp seam (act_grp 3 -> 2 by design) bk_stage_start, benign: the refill rewrites identical data
KV-only item: G1 -> G4 -> G5_ROPE -> DONE; the next item's G4 (KV-only or root) 0 -> 0 (cache_grp 0) no g4_start new row g4_start ALONE
c5 KV_MROW sub-row j -> j+1: G4 -> G5_ROPE -> S_MROW -> G1 -> G4 (top_fsm.sv:528, 531, 536, 527) 0 -> 0 throughout (default decode) no g4_start (re-pulses at :756 on S_G1 -> S_G4) new row (FASTIN row j+1) g4_start ALONE
re-arm: request 1 last item -> W 154 1 -> request 2 first G4 0 -> 0 (cache_grp 2 after a full item, 0 after a KV-only item) after a full item only g4_start new g4_start alone if request 1 ended on a KV-only item
G9 / G12 / TREE not on the ATT (g12_start constant 0; no S_G9 on the head)
G1 (RMSNorm) / G5 (attention) not array stages: stage_sel decodes 0 and no array start is issued

Answer: no transition has bk_stage_start as the sole guard of a DIFFERENT activation. Its only sole-guard case is G7g -> G7u, where the activation is the same, by the act_grp design. The cases with a single guard and NEW activations are all G4 -> G4 with g4_start as the only trigger: consecutive KV-only items, KV-only -> root, the c5 multi-row sub-rows on 138C, and re-arm after a KV-only item.

REDs [M] (buildx01; xsim-only source copies in new dirs; afc_att_evidence/xsim/afc_red2.sh, PATCH_RED*.log):

RED source edit stim vs result
F10 AFCRL_ROOT act_reload = g4_start | g12_start (bk_stage_start dropped) stim_root AFC0_ROOT PASS 28/28, LCCHASE 0: masked by the group-change term (the table above)
RED2 AFCRED2_ROOT + bridge need_ld = reload_pend || act_reload || !cache_ok (group term forced false) stim_root AFC0_ROOT FAIL as required. TAPG4 15/15, TAPG5 1/1 and KVWR 500/500 are identical (G4 refilled by g4_start, G5 not an array stage). The first divergence is D9R line 15 = p=1 st=1 t=0 (G6 tile 0) on all 4 ranks, then TAPG6 20/20, D9R G7g/G7u, TAPR4 512/512, seams. LCCHASE 0: the chase checker sees ORDER, not staleness, so the A/B is the stale-cache instrument (AB_AFCRED2_ROOT_vs_AFC0_ROOT.txt).
RED_C5BK AFCRL_C5_H_A act_reload = g4_start | g12_start (bk dropped), c5 source c5 stim A (KV_LOAD, 8-row + 1-row KV-only, KV_LOAD, root) H_A (shipped c5h) PASS so far (prefix: 8-row sub-rows 0..2+ identical, TAPG4 / D9R / KVWR). Expected, because the sub-row refill term is g4_start (group term does not fire 0 -> 0; cache_ok stays 1; bk_stage_start is 0 at stage_sel 0).
RED3 AFCRED3_C5_H_A act_reload = g12_start (g4_start AND bk dropped) + group term forced false c5 stim A H_A FAIL as required, localized to row 1's G4. Row 0 identical (filled by !cache_ok after reset). First divergence TAPG4 / D9R index 15 = p=1 st=0 t=0 (sub-row j=1, G4 tile 0) on all 4 ranks, and KVWR n=140 = sub-row 1's K/V at pos 14, 10/10 heads (AB_AFCRED3_C5_H_A_vs_H_A_prefix.txt). This proves the per-row compare sees a stale cache.
Conclusion: g4_start is the sole refill term for every G4 -> G4 transition (c5 sub-rows, consecutive KV-only items, KV-only -> root); RED3 removes it and the per-row K/V compare fails at the first such transition.
GREEN, c5 full [M]: AFC1_H_A (138C recipe, AFC 1) vs shipped H_A: 28/28 classes, every KVWR (500 per card = 13 KV_LOAD rows + 8 + 1 KV-only sub-rows + 27 KV_LOAD rows + the root row, x10 heads), seams, LCCHASE 0 on 12 groups. c5_check vs the serial control SER_SA: kv_final K/V 0/62,720, kv_root 0/2,560, h1 0/5,120, r5 0/4,096, ordered stream 0/500 per card: VERDICT PASS (CHK_AFC1_H_A_005835.txt).
Cycles: 8-row item G4 38,600 -> 20,672.
Re-arm with a DIFFERENT second request [M] (mk_rrd.py; auditor note: the same-root re-arm pairs AFC1_RR / AFC1_FRR produced md5-identical item_01 / item_03
seams, so on their own they cannot tell a fresh computation from a stale reproduction):

74.10 02:33:35Z 2026-11-18 (date -u): xsim AFC1_FA3 (FR_A8, root + L1 0-3 + L2 0-3) FINISHED: GREEN

Snapshot snap/AFC1_FA3_023322 (log md5 945b73aa, 40 item-ENDs = 10 items x 4 ranks). Checker vs ref/a16fra8_kc512/vectors_cap0_kc512 (CHK_AFC1_FA3_023322.txt): AFC_VERDICT PASS (forward items 9, 19 stream rows) -- every checked stream (g4_y48 / D9R, kv_k / kv_v, g5_out, g6_part / g6_sum48, h1, g7g / g7u, r4g / r4u, r5_quarter, kvload) 0 on c0..c3 for each of 1:L0[0], 2-5:L1[0..3], 6-9:L2[0..3] (FR_A8 child a-bank items included); g1_x / g6_in / pn = D9V INFO class. LCCHASE END 12 x violations=0. => flash-condition arm AFC1_FA3 MET. C5CTL_RR (control) done 02:3xZ; its pair AFC1_C5RR still running.

74.11 02:44:25Z 2026-11-18 (date -u): xsim AFC1_S5 (cap0 root + L1 0-3 + L2 0-3, cross-item KV chain + doorbell re-arm) FINISHED: GREEN

Snapshot snap/AFC1_S5_024411 (log md5 adb7b059, 40 item-ENDs). Checker vs ref/a16_kc512/vectors_cap0_kc512 (CHK_AFC1_S5_024411.txt): AFC_VERDICT PASS (forward items 9, 19 stream rows), every checked stream 0 on c0..c3 for 1:L0[0], 2-5:L1[0..3], 6-9:L2[0..3]. LCCHASE END 12 x violations=0. => flash-condition arm AFC1_S5 MET. xsim flash-condition tally at 02:44:25Z: GREEN = AFC1_G1, AFC1_S5, AFC1_FA3, AFC1_H_A (c5), AFC1_RR, AFC1_FRR (+ earlier AFC0_ROOT, RED AFCRC / AFCES fired). NOT MET: F10 RED AFCRL_ROOT (did not fire, 74.5). Pending: AFC1_C5RR (vs C5CTL_RR, done).

74.12 02:45:13Z 2026-11-18 (date -u): xsim tally CORRECTED (qwen-gptq2-arch, BUILDS sec.75.3 / commit bbbebd4de, read and checked)

F10 is CLOSED (not "NOT MET"): AFCRL_ROOT was a redundant-trigger RED (74.5 mechanism, confirmed in 75.3's transition table). The biting REDs are in 75.3: RED2 (bk_stage_start dropped + the group term forced false) FAILs at the first G6 on 4 ranks; RED3 (g4_start dropped too, c5 8-row stim) FAILs at sub-row 1's G4; RED_C5BK (bk_stage_start only, c5) is predicted PASS (g4_start refills alone). Tally: GREEN = AFC1_G1, AFC1_S5, AFC1_FA3, AFC1_H_A, AFC1_RR, AFC1_FRR + AFC1_RRD / AFC1_FRRD (distinct 2nd request, 28/28; closes the 74.6 same-root limit) per 75.3; REDs fired = AFCRC, AFCES, RED2, RED3. Pending: AFC1_C5RR, RED_C5BK.

74.13 02:56:13Z 2026-11-18 (date -u): xsim c5 re-arm AFC1_C5RR vs C5CTL_RR: PASS

Snapshots snap/{AFC1_C5RR,C5CTL_RR}_025556 (log md5 67fc1e5b / 46a34a81, 24 item-ENDs + $finish each; c5 stim A_rr3 = 8-row + 1-row KV-only, KV_LOAD, root, root again). qg_afc_ab.py (snap/AB_AFC1_C5RR_vs_C5CTL_RR_025556.txt): VERDICT PASS, D9R 269 / KVWR 510 / TAPG4 165 / TAPG5 11 / TAPG6 40 / TAPR4 1,024 per card identical, seams 2 / 2 identical (D9R total 1,076). LCCHASE END 12 x violations=0. c5_check not run on this arm (its A_rr3 stim dir is not on box1 c5stim/; the ordered KVWR stream equality vs the shipped control is inside the A/B). Same-root limit as 74.6 applies (both root seams 0a9156c4); the distinct-request closure is AFC1_RRD / AFC1_FRRD (75.3). Pending xsim: RED_C5BK (run AFCRL_C5_H_A, predicted PASS).

74.14 03:19:52Z 2026-11-18 (date -u): ATT AFC row (j) SLR census at place -- SLICE / SLL PASS-shape, but the locality pblocks are EMPTY (inert xdc)

Placed util impl_1/place_report_utilization_0.rpt (md5 138A 6235ce0b / 138C a288c2dc / 138F 3c6df531; base 139C buildy00 s3LAq2BE9C bf082514), lines :61 device SLICE, :393 SLICE per SLR, :364 / :367 SLL:

image SLICE device SLICE SLR0 / SLR1 / SLR2 SLL 2<->1 SLL 1<->0 URAM SLR0/1/2 BRAM tile SLR0/1/2
139C base 178,422 (55.45%) 87.13 / 73.06 / 5.12% 6,583 (34.89%) 13,149 (69.68%) 301 / 519 / 93 1,035.5 / 645 / 87
0xE4B7138A 182,829 (56.82%) 79.37 / 65.28 / 24.91% 1,412 (7.48%) 10,810 (57.29%) 403 / 258 / 252 1,026.5 / 505 / 260
0xE4B7138C 179,831 (55.89%) 76.28 / 65.64 / 25.01% 1,519 (8.05%) 11,155 (59.11%) 169 / 492 / 252 866.5 / 665 / 260
0xE4B7138F 180,694 (56.16%) 76.51 / 66.31 / 24.94% 1,413 (7.49%) 11,662 (61.80%) 172 / 489 / 252 871.5 / 660 / 260
Placed-dcp audit (watch/afc_pblock_slr.tcl on md5-checked copies, buildx04, 138A ad9496cb / 138C 23349e2d / 138F 996cf2f1, RC 0 x3):

74.15 03:46:05Z 2026-11-18 (date -u): xsim RED_C5BK (AFCRL_C5_H_A vs H_A) PASS as predicted; xsim flash-condition side COMPLETE for 138A / 138C / 138F

RED_C5BK = the c5 source with bk_stage_start dropped from act_reload; run buildx01 ~/box1_scratch/c5_mrow_20261117/runs/AFCRL_C5_H_A (log md5 9ab5a647, mtime 03:42:50Z, $finish 1, ERROR/FATAL 0, LCCHASE END 12/12 violations=0; verified by this watch) vs shipped H_A (ed2af8bf). qg_afc_ab (per qwen-gptq2-arch): every class identical per card (D9R 202, TAPG4 150, TAPG5 10, TAPG6 20, KVWR 500, TAPR4 512, seams 1/1) => PASS, as predicted (g4_start refills every sub-row; RED3 is the biting one, 75.3). Final xsim tally: GREEN = AFC1_G1, AFC1_S5, AFC1_FA3, AFC1_H_A, AFC1_RR, AFC1_FRR, AFC1_RRD, AFC1_FRRD, AFC1_C5RR. REDs fired = AFCRC_ROOT, AFCES_ROOT, RED2, RED3. REDs masked by a redundant trigger (expected, mechanism 74.5 / 75.3) = AFCRL_ROOT, RED_C5BK.

76. K64 + ATT AFC wave 0xE4B71383 / 1384 / 138E / 1395: pre-launch xsim gate (qwen-gptq2-arch build agent + its xsim agent; STAGING sec.26), 03:45Z 2026-11-18 (date -u)

DUT sources (buildx01): K64 patch (config context adapted to the staged LUTG_ZSRC_V 1'b0: a154cc08, added block byte-identical) -> patch_qg_att_fastclk.py -> straps ATT_ARRAY_FAST_CLK_V 1'b1 + K64_V 1. Array 5ab0f4fe / bank 7fc1460b / multistage a0358b70 / lane dacfcd99 / bridge 0395872d = the derived stages' staged files. Evidence k64afc_xsim/checks/, box1 k64afc_20261118/snap/, harvest k64afc_xsim/tools/k64_harvest.sh (A/B = qg_afc_ab.py; checker = qg_att_check_afc.py).

arm DUT stimulus parent / ref verdict LCCHASE / K64CHASE viol ITEMMON busy [M]
K64B_ROOT (i) 139C stim_root shipped L256_ROOT_GREEN2 / a16_kc512 A/B PASS (28 classes + seams, D9R 268); AFC_VERDICT PASS 0 / 0 (60,192 quad reads) 8,102 (G4 1,464 G6 883 G7G 2,136 G7U 784)
FK64B_ROOT (iii) 13A9 stim_root L256_ROOT_GREEN2 A/B PASS; AFC_VERDICT PASS 0 / 0 8,102
K64B0_ROOT strap 0 patched, K64 0 / AFC 0 stim_root L256_ROOT_GREEN2 --keep-time PASS cycle-identical (TB_ACCPAR_DONE @395,462,000 both) n/a 15,085
K64B_RRD (i) stim_s5_rrd (root, re-arm W 154 1, L1 child) shipped CTL_RRD A/B PASS, 2/2 seams nonzero x4; item_01 != item_02 in the control 0 / 0 (120,384) 8,102 / 8,774
FK64B_RRD (iii) stim_fa3_rrd shipped FCTL_RRD A/B PASS, same counts 0 / 0 8,102 / 8,774
FK64B_FA3 (partial) (iii) stim_fa3x root + L1[0] so far a16fra8_kc512 every non-D9V stream 0 on 4 ranks (FR_A8 child a-bank exercised) running L1 8,774
RED K64_RED 2 (i) stim_root L256_ROOT_GREEN2 FAIL as required (G4 3,584/3,840, tiles 1..14) 0 / 0
RED K64_RED 3 (i) stim_root FAIL as required (X; D9R 53/67) 1,272 / 1,284
RED R1L (LCACHE->K64 cache write stale-by-1) (i) stim_root FAIL as required (67/67 tiles) 0 / 0
RED RCB (AFC cluster-source revert + K64 || LCACHE dropped) (i) stim_root FAIL as required (G4 tile 0, X) 0 / 0
RED RF (all refill terms: bridge act_reload w/o bk_stage_start + need_ld act_grp compare off + k64_act_new w/o bk_stage_start) (i) stim_root FAIL as required (G6 5,120/5,120, G7g/u all) 0 / 0 (chase blind to RF: its counter resets on k64_act_new)
RED AN (k64_act_new w/o bk_stage_start only) (i) stim_root FAIL (bites) 660 / 660
RED K64_RED 1 / K64_RED 4 / RC (--red-cluster alone) (i) stim_root PASS = VACUOUS by construction: snoop g_wsn not elaborated at LCACHE 1; weights by snapshot; at LCACHE 1 kc_use01 = 1 so the AFC src_a edit is dead code under K64. Biting forms: R1L, unit RED 4 (67/67), RCB
Unit gate (tb_k64_array LCACHE harness, 67 tiles): GREEN 0 mismatch; RED2 63, RED4 67, R1L 67 FAIL; RED1 / RED3 vacuous in the unit (no snoop; TB pre-fills). UNIT_K64AFC_LCRED.txt.
Found [M]: every mk_rr3 re-arm stimulus (stim_root_rr3, stim_fa3root_rr3, c5 A_rr3) has 0-byte pre.ops / final.ops (open(p,'w').write(open(p).read()) truncates first):
no out-base CSRs, no PIPEOUT, every seam = 98,304 zero bytes (md5 0a9156c4 = /dev/zero) -> the seam compares of sec.74.6 (AFC1_RR / AFC1_FRR) and 74.13 (AFC1_C5RR) are VACUOUS;
their D9R/KVWR/TAP rows ran without the pre.ops setup. RRD stims are fine (pre.ops 315 B). Re-run owed for the AFC wave's re-arm flash condition.
Still running at 03:45Z (buildx01): K64B_G1 (GPU req00 ctx 112 root + L1 0-3 vs L256_GPU00_MD2 + gpu_round0), FK64B_FA3 (remaining L1/L2 items), C5K64A_H_A (c5 8-row + 1-row
KV-only + root; c5_check.py per-sub-row K/V vs SER_SA), and the different-root re-arm K64B_RRRB / CTL_K64RRRB / FK64B_RRRB / FCTL_K64RRRB (stim_k64rr_root2b: cap0 root, re-arm,
GPU req00 KV_LOAD + root ctx 112; the v1 stim_k64rr_root2 had the same truncation bug and is marked INVALID).