[guluchen]

Follow-up to #230. This PR renames 19 tier5 stems (four of them now exact) whose lower bounds were raised in 12-36 h runs on a dedicated server. Each value is reproduced in the batch: same matrices, and only the Gx/Gz header comments change, as in #230 / #223.

Rule used: a value goes into a stem only when two runs with different search trees reached it, with the last level EMPTY on both halves (X half only under a verified XZ-dual map). This is the same CC engine family, not an independent second method.

Arms:

• C = main 755e0c2's CC default (gated erasure).
• E = main 1239e9a's CC default plus ungated erasure.

C and E differ in binary and gating. TN_684 uses relabelled E runs instead; COLOR_2791 uses a bfs row order.

Old (main)	New	Arm 1 (s)	Arm 2 (s)
BB_1080_4_32to48	BB_1080_4_34to48	C 40 586	E 44 390
BB_756_16_32to34	BB_756_16_34 (d = 34)	C 59 834	E 65 917
BB_864_4_32to40	BB_864_4_34to40	C 58 134	E 64 158
COLOR_1027_1_31to37	COLOR_1027_1_33to37	C 78 893	E index 77 146
COLOR_2791_1_28to61	COLOR_2791_1_30to61	C 42 701	C bfs 79 340
TN_288_2_22to24	TN_288_2_24 (d = 24)	C 98 733	E 100 460
TN_324_4_23to25	TN_324_4_24to25	C 72 074	E 72 807
TN_360_4_23to24	TN_360_4_24 (d = 24)	C 47 415	E 48 186
TN_396_4_23to26	TN_396_4_24to26	C 47 489	E 47 923
TN_432_8_23to29	TN_432_8_25to29	C 87 063	E 89 519
TN_468_4_23to29	TN_468_4_25to29	C 107 865	E 110 270
TN_496_2_28to30	TN_496_2_30 (d = 30)	C 117 232	E 119 072
TN_504_2_24to34	TN_504_2_25to34	C 54 147	E 56 955
TN_540_4_24to34	TN_540_4_25to34	C 57 564	E 63 505
TN_576_2_24to37	TN_576_2_26to37	C 79 691	E 88 962
TN_612_2_24to40	TN_612_2_26to40	C 97 484	E 107 037
TN_648_10_22to32	TN_648_10_23to32	C 53 620	E 58 436
TN_648_14_22to28	TN_648_14_23to28	C 70 302	E 63 576
TN_684_2_24to29	TN_684_2_25to29	E -seed=1 106 893	E -seed=2 105 414

Checks on the solved stems and on all others:

• verify_levels.txt: in all 71 runs every recorded level is EMPTY. The missing levels are the odd levels of the BB codes (parity rule), and verify_parity.txt confirms that every column of Hx and Hz has odd weight there, in pure Python.
• verify_dual.txt: the XZ-dual maps of the X-half-only codes are re-checked in pure Python: BB_756, BB_864, BB_1080 (halfswap-inverse), COLOR_1027, COLOR_2791 (identity, Hx = Hz) and TN_496 (tanner-swap). The TN_496 map was dumped from a scratch patch of 755e0c2, which is not in any commit.
• Upper bounds are unchanged, and the upper-bound witnesses of the 17 TN/BB stems are re-verified under the new stems in optimization/experiments/GH-146/witnesses/renamed-2026-11-07/VERIFY.txt (17/17 VALID). For the four solved stems the lower bound meets a witness of that weight.
• All four solved stems: both arms printed o d and passed the runner's science check. Their stems were {lb}to{ub} on main.

Caveats:

• C and E both use the erasure rule (gated and ungated), so a soundness bug in shared code would affect both. Runs without erasure reached only the lower values noted in the rename history.
• COLOR_1027's two arms use the same row order and differ only in gating and binary (level 32: 38.0G vs 36.3G nodes), so that pair is the weakest of the set.
• The first batch (12 h, data: rename 19 tier5 stems with reproduced CC lower bounds (GH-225 follow-up) #232's original commit) already reached some of these values with intermediate stems; the rename-history rows here show the final stems only, with the earlier times where useful.

Not renamed: COLOR_2791 lb 31 (only C, 88 551 s; PLAUSIBLE).

Records are in optimization/experiments/GH-225/lb-push/ (REPRODUCTION.md, summary.md, verify_*.txt, push1-push5, repro3).

Changes:

• git mv of the matrix files;
• rename-history rows and the witness note in data/matrices/README.md;
• NOTICE sections 5 and 9;
• stem names in optimization/baselines/SERVER.md, main_server_reference.json and README.md (examples), and in the untiered MANIFEST.csv, validation.json and README.md.

QDistSAT mirror: guluchen/QDistSAT#19 (same branch name); all 80 matrix files are byte-identical.

Hub: #15; follow-up to #225.

🤖 Generated with Claude Code

[guluchen]

All checks pass. Merge is on hold together with the QDistSAT mirror (guluchen/QDistSAT#19), so the two repos keep the same stems. The mirror's test (3.9) job fails in setup: ubuntu-latest is now Ubuntu 26.04, which has no Python 3.9. The fix needs a QDistSAT workflow edit that our token cannot push.