[guluchen]

Closes #217 once merged (hub #15). Not merged by the agent: the Tier 1 start-up floor is outside the rules and needs the PI's decision (below).

Change

CC engine defaults (no search code changed):

• rules excl,pack,orbf,parity,leaf,overlap,charge -> excl,pack,orbf,parity,leaf,overlap,charge,tiebreak,sparse,sparseov,sparsemrv (candidate A);
• row order -cc-roworder=index -> -cc-roworder=auto.

The previous default stays available: -cc-rules=excl,pack,orbf,parity,leaf,overlap,charge -cc-roworder=index (byte-identical to main). README, -h, MODIFICATIONS and the tests are updated; the -cc-trace header names auto for the auto row order. The MaxSAT default path is unchanged.

Credits as for the components: Qubitserf 04b9349 (S. Cercelescu, MIT) for the first(g) tie-break, reversed-BFS row order and redundant pair-XOR checks; Puchert and Tillmann.

Tier 0

• make test-cc: PASS. make test-cc-flags: 551 checks, 0 failed.
• Identity vs main 02879a9 on tier0+tier1 (15 codes): every CC configuration with the old-default flags is IDENTICAL. MaxSAT with -ub-pre=off is IDENTICAL 30/30; with the default -ub-pre=on, 3 of 30 runs differ only in which half's finder converges first (timing; main-vs-main differs the same way).
• 253 answer runs (tier0-3 x 11 candidate variants): every answer equals its stem. The GF(2) witness check passed: 152 witnesses, 0 bad.

Tiers 1-4

5 seeds; main, A and B ran concurrently on pinned yfclab2 cores. Ratio = median(arm) / median(main).

Tier	A geomean	A worst cell	B geomean	B worst cell
1	0.969	0.974 (LP_442)	0.918	0.949
2	0.727	1.013 (GB_144_12_12)	0.869	1.000
3	0.769	1.023 (LP_714)	0.778	1.046
4	0.081	0.081 (TN_250: 60.9 s -> 4.9 s)	0.444	0.444

Tier 1 start-up floor (PI decision). Tier 1 cells take 24-39 ms. I also ran a quiet interleaved panel: 9 repetitions x 5 seeds, arms round-robin on one core. There the Tier 1 geomean is A 1.031 (LP_340 0.988, LP_442 1.076) and B 1.035. LP_442 costs about 2.5 ms more in both A and B, so the cost comes from -cc-roworder=auto (it builds the second half search), not from the sparse tokens. The tier0 codes measured the same way give a geomean of 1.01 (max 1.050).

Tier 5

6 h, default seed, main and A concurrent; all 27 codes. Every code that main's CC default solves:

Code	main	A	A/main	interleaved re-run
LP_1020	0.23 s	0.26 s	1.16	1.044
LP_1054	1.77 s	2.08 s	1.17	1.119
LP_1122	4.81 s	5.63 s	1.17	1.137
LP_1428	127.4 s	130.4 s	1.02
LP_1768	16.9 s	17.7 s	1.04	1.021
LP_2610	2.96 s	3.39 s	1.14	1.078
BB_360	181 s	31 s	0.17
TN_180	142 s	145 s	1.02
TN_216	466 s	393 s	0.84
TN_252	8490 s	8581 s	1.01

No solved code is more than 1.2x slower. On the LP codes the extra time is the auto probe: one aborted bfs + mrv rerun of the first refuted level that takes at least 0.1 s.

Codes main does not solve:

• BB_756, BB_864 and BB_1080 reach lb 30 about 7-10x sooner, and A reaches lb 32 (main reaches 30).
• On TN_496, A reaches lb 27 (main reaches 23).
• On the other TN codes, lb times are within 2 %, except TN_684, where A is 0.74x.

Lower bounds above the stem names, reached by both arms:

• TN_468 23;
• TN_612 24;
• TN_648_14 21;
• TN_684 22.

Relabelled reproductions are running; the stems will be renamed in a separate PR.

Untiered codes are not judged here. A's worst cell is 1.18x (concurrent) and 1.051 (interleaved). A is much faster on HGP_1105 (187 s -> 3.7 s) and COLOR_331 (121 s -> 15 s).

Evidence: optimization/experiments/GH-217/ (runners, S0/, T5/, analysis.md).

Update 2026-11-05: merged main 972b28c (#219 erasure opt-in, #221 re-tier) and fecee8c (#223 renames)

Tier 0 on the merged binary (b7b4126):

• make test-cc-flags: 572/572.
• Identity vs main 972b28c on the new tier0+tier1 (35 codes): every CC configuration with the old-default flags is IDENTICAL. MaxSAT -ub-pre=off is IDENTICAL 70/70; with the default -ub-pre=on, 1 of 70 differs by finder timing only.
• Answers: 550 runs on new tier0-3 x 11 variants, all equal to their stems; 290 GF(2) witnesses OK.
• make test-cc: see the issue comment.

Re-judged under the new tier layout: 5 seeds, merged candidate vs main 972b28c, concurrent on pinned cores with idle SMT siblings.

Tier	A geomean	A worst cell	B geomean	B worst cell
1 (9 codes)	0.855	1.111 (HGP_265)	0.873	1.269 (BB_186)
2 (8)	0.648	1.222 (LP_544, a 36 ms cell)	0.745	1.194
3 (7)	0.573	1.024 (LP_714)	0.743	1.224 (HGP_2866)

Quiet interleaved Tier 1 (new layout, merged binaries, 9 repetitions x 5 seeds on one core): A geomean 0.898. LP_442 is 1.467, because the cell's wall time is quantized: the parent polls the CC child every 10 ms (usleep(10000) in src/core/distqldpc.cc). LP_442 runs land at about 23 ms or about 33 ms. A's extra user time (+9 %) pushes every A run into the 33 ms bucket, while main lands mostly at 23 ms. This one start-up-floor cell is the only item outside the rules, so the PR stays open for the PI.

Tier 5:

• New tier5 members: COLOR_331 (main 121 s -> A 15 s), HGP_1105 (70 s -> 4 s); COLOR_1027 and COLOR_2791 are unsolved by both, and A reaches higher lbs.
• Interleaved re-run of the short LP cells with the merged binaries: LP_1054 1.102, LP_2610 1.101, LP_1122 1.112, LP_1768 1.009. All are within 1.2x.

TN lb shortfall vs stems (coordinator question):

• Not a code regression. Same-session TN_360 to level 20: main 02879a9 runs within 1 % of the [Mac agent] CC engine: fractional-charge / edge-cover node lower bounds #157 binary c3c315d, with identical node counts. In the Tier 5 batch, levels took about 10-20 % longer than on the [Mac agent] CC engine: fractional-charge / edge-cover node lower bounds #157 day, which pushes lbs that [Mac agent] CC engine: fractional-charge / edge-cover node lower bounds #157 reached at 21 100-21 560 s past the 6 h limit.

• -cc-lb-auto is not the cause on TN_360/396/504. Charge stays on there (about 75 % prune rate).

• Forcing charge on does not recover the stem lbs (6 h, concurrent A vs A with -cc-lb-auto=0):

  Code	A	A with -cc-lb-auto=0
  TN_648_10	lb 21 at 8370 s	lb 21 at 8219 s
  TN_648_14	lb 21 at 9003 s	8747 s
  TN_468	lb 23 at 17 381 s	17 360 s
  TN_496	lb 27 at 10 028 s	13 608 s (X levels about 1.35x slower)

• The flip therefore does not include an lb-auto change.

• The BB lbs (32) come with the flip itself (tiebreak).

Renames done (#223 in DistQLDPC, guluchen/QDistSAT#17), each reproduced with a relabelled run:

• TN_468_4_23to29;
• TN_612_2_24to40;
• TN_648_14_21to28;
• TN_684_2_22to29.

🤖 Generated with Claude Code

[guluchen]

Tier 1 re-check on top of the GH-224 fix (#224, PR #226, main 82ca41d)

#226 removes the 10 ms polling floor in the parent's wait loop (poll() on a SIGCHLD self-pipe instead of usleep(10000)).

Setup

• Candidate A = this PR's head 42d7d1d with main 82ca41d merged in locally (d8cf2b39; not pushed). Compared with this PR, its src/ differs only by GH-224: parent waits for the solver child in poll() (removes the 10 ms timing floor) #226.
• Reference = main 82ca41d.
• CC defaults of each binary; 5 seeds x 9 repetitions.
• Arms interleaved with rotating order on one pinned core per (code, seed), on yfclab2 cores 72-103. Each core and its SMT sibling were checked idle before every unit.
• Ratio = median over seeds of the repetition medians, A / main.

Tier 1 (9 codes): geomean 0.870, worst cell 1.082 (LP_442), no cell > 1.25.

Code	main wall (s)	A wall (s)	A/main	A/main CPU
AMC_438_6_17	0.0746	0.0595	0.798	0.769
BB_186_10_14	0.0379	0.0332	0.875	0.988
BB_288_12_18	0.3464	0.2628	0.759	0.760
COLOR_127_1_13	0.0672	0.0332	0.494	0.450
GB_144_12_12	0.0744	0.0721	0.969	0.995
HGP_265_1_12	0.0331	0.0343	1.037	0.978
HGP_610_64_6	0.1482	0.1437	0.970	1.000
LP_340_56_8	0.0244	0.0252	1.033	1.058
LP_442_68_10	0.0271	0.0293	1.082	1.064

LP_442

• The 1.467 in the earlier quiet panel was the 10 ms bucket.
• With the fix, both arms' wall times are contiguous: main 24-37 ms, A 25-39 ms, 1 ms bins.
• What remains is A's real extra CPU, about 6 % (the -cc-roworder=auto set-up noted earlier). It gives a 1.08x wall ratio.

Tier 0 (26 codes, information only): geomean 0.991, worst cell 1.186 (BB_144_12_12, 15.6 ms vs 18.5 ms), no cell > 1.25.

Correctness: 3150 runs, all rc 0, all answers equal to the stem. stdout differs only in timing-dependent line order on GB_144_12_12, and main varies the same way across repetitions.

The PR is not merged. The merge decision stays with the PI.

[guluchen]

Merged under the standing progress rules:

• Tier 0 passes.
• After the Parent wait loop: remove the 10 ms usleep polling floor on fast runs #224 timing fix (GH-224: parent waits for the solver child in poll() (removes the 10 ms timing floor) #226), Tier 1 has geomean 0.870 and worst cell LP_442 at 1.082. Tier 2 is 0.648 and Tier 3 is 0.573, with no cell above 1.25.
• Tier 5: every solved code stays at or below 1.17x.
• The new lower bounds were reproduced and renamed in data: rename TN_468/TN_612/TN_648_14/TN_684 with GH-217 CC lower bounds #223.

🤖 Generated with Claude Code