[guluchen]

Closes nothing yet: PI decision needed. Do not merge without it. Issue #218, hub #15. Follow-up of #148 / PR #158.

Problem

With GH-148's -stab-overlap=auto (main default), TN_216_4_18 is slower on 5-seed medians: 1.53x (no-card) and 1.39x (card-mto). That violates the Tier 5 1.2x rule.

Change

• New rule, -stab-overlap-c4=N|off (default 11). A verified overlap row g is dense if it lies on more than N Tanner 4-cycles of the stabilizer matrix, c4(g) = sum over the other rows g' of C(|g & g'|, 2).
  • Every weight-9 TN row is dense; BB/LP/GB rows have c4 <= 4, so they never are.
  • Dense rows are used as in [Mac agent] MaxSAT path: stabilizer-overlap dominance constraints in every CSS-half probe #148 until a half's first probe with cap >= P' + 4, where P' = max(12, lb, activation cap). On the TN codes that is cap 16. They are dropped from there on.
  • Sparse rows are always used.
  • -stab-overlap-c4=off gives [Mac agent] MaxSAT path: stabilizer-overlap dominance constraints in every CSS-half probe #148's search byte for byte.
• Opt-in -stab-overlap-pilot=N (default 0). A per-half pilot: "weight <= P'?" with and without the dense rows, N seeds; keep the rows iff the median conflict ratio is <= 1. It is not the default because a cap-12 pilot does not predict the cap-16/17 effect (below).
• Output. -v prints the dense-row count per half. -frontier-stats prints c OVERLAP-DROP / c OVERLAP-PILOT lines.
• Solver. The two ovPropagate call sites first test inline whether the literal sets a qubit on a row true. Same propagation; this only saves calls.
• Soundness. Any subset of [Mac agent] MaxSAT path: stabilizer-overlap dominance constraints in every CSS-half probe #148's verified rows keeps every minimum-weight element of the half instance, so no probe answer changes. See the [Mac agent] MaxSAT: gate -stab-overlap to remove the TN_216 regression (GH-148 follow-up) #218 paragraph in MODIFICATIONS.md.

Mechanism (diagnosis; optimization/experiments/GH-218/diag/)

• Not overhead. On the same core, with the rows set but inert (same search as off), the cost per propagation equals off within 1%. ovPropagate takes about 3% of a TN_216 Z16 probe.
• A search effect at caps 16-17 of the TN Z half.
  • TN_216 seed 0 no-card, conflicts rows/off: X16 0.94, Z16 1.33, X17 1.02, Z17 1.38.
  • Over [Mac agent] MaxSAT path: stabilizer-overlap dominance constraints in every CSS-half probe #148's 10 seed/mode pairs: Z half 1.03-1.54x, X half 0.85-1.07x.
  • The rows rarely fire: about 0.1 implications per conflict, against 0.45 on BB_288.
• The same rows help TN_180's X half. Card-mto X16 is 0.48-0.66x the conflicts on 3 seeds (seed 0: 9062 -> 3951 s).
• No cheap separator found. Checked:
  • structural statistics: 4-cycles, overlap neighbours, local parity rank, logical weights;
  • usage variants: no rows in the lookahead (Z16 1.18x but X16 1.05x), lookahead only, no VSIDS bumps from overlap reasons;
  • symmetry breaking: without it, TN_216 Z16 is 1.09x;
  • pilots at caps 10-12. TN_216 Z at cap 12 over 10 seeds has median 0.92-0.97, and the live 3-seed pilot kept Z on seeds 3 and 4.
• Below cap 16 the TN rows are neutral to helpful. TN_108 is 0.92-0.95x in conflicts. TN_250's gain comes entirely from its weight-6 rows (c4 7 and 10), which stay sparse.

Gates

Tier 0: PASS.

• make test-cc PASS. Part G: 4600 driver runs, 0 bad, including 5 new dense/drop/pilot configurations.
• Part G (c):
  • the sparse-row selection equals a brute-force 4-cycle count in 224 selections;
  • the dense flags partition the rows;
  • forced on/off probe answers agree;
  • pilot decisions are reproducible;
  • mutant 3 (overlaps counted instead of pairs) changes 78 selections and is caught.
• make test-cc-flags: 579 checks, 0 failed.
• On every tier0-4 code, in both modes, the default and -stab-overlap-c4=off match main exactly: byte-identical stdout and identical per-probe FRONTIER/OVERLAP lines.

Tiers 1-4: PASS by identity (ratio 1.000).

• No Tier 1-4 probe reaches cap 16: -ub-pre gives U0 = d on 5 seeds x 2 modes for TN_108, TN_200, TN_144 and TN_250.
• BB/LP/GB have no dense rows.
• So the search is main's for every seed.

Tier 5 (yfclab2, 6 h, 2026-11-04). Data in optimization/experiments/GH-218/T5/.

TN_216_4_18, 5 seeds. The candidate's per-probe conflict counts equal 12b3e05's on every probe of all 10 seed/mode pairs, so the search is the same. The 12b3e05 times are GH-148's 2026-10-28 runs; a seed-0 drift control in the same session measured +11.6% / +11.7%.

Mode	cand s0..s4 (s)	cand median	12b3e05 median	raw	drift-adjusted	vs current main
no-card	10485, 9515, 11596, 13349, 9945	10485	9282	1.13	1.01	s0 same session 0.71 (14697 s)
card-mto	11284, 12018, 11577, 13210, 9272	11577	10217	1.13	1.01	s0 same session 1.005 (11225 s)

1.2x check against current main (default seed, same session):

Code	no-card	card-mto
TN_216	0.71	1.005
TN_180	1.07 (5919 -> 6314 s)	1.57 (6703 -> 10554 s): FAIL
BB_360	still solved, identical search (7775 s, d = 24)	still solved, identical search (8075 s)
LP_1122	still solved, identical search (13590 s, d = 20)	still solved, identical search (14023 s)
LP_1020	identical search (664 s)	identical search (466 s)

The identical-search rows have the same per-probe conflicts as GH-148's Tier 5.

Decision for the PI

1. Merge this rule. TN_216 returns to the 12b3e05 time. TN_180 card-mto gives back its [Mac agent] MaxSAT path: stabilizer-overlap dominance constraints in every CSS-half probe #148 gain: 11634 s before [Mac agent] MaxSAT path: stabilizer-overlap dominance constraints in every CSS-half probe #148, 6703 s on current main, 10554 s with this rule. TN_144 keeps [Mac agent] MaxSAT path: stabilizer-overlap dominance constraints in every CSS-half probe #148's 1.07-1.10x (cap 12 < 16).
2. Keep [Mac agent] MaxSAT path: stabilizer-overlap dominance constraints in every CSS-half probe #148 as is. TN_216 stays 1.4-1.5x slower.

No rule found in this round satisfies both. Possible follow-up: restrict dense rows to the main search (no lookahead). On TN_216 seed 0 no-card that gave 1.10x conflicts vs off, and it kept TN_180's X gain, but only one seed was measured.

🤖 Generated with Claude Code

[guluchen]

Follow-up (branch experiment/gh-218-nolk; no PR): I tested keeping the dense rows in the search's propagation but out of the lookahead. It fails at the time level, so I stopped; no Tier 5 was run.

Conflict-level panel at caps 16-17, 5 seeds x 2 modes (single-probe diagnostic runs):

	no-card	card-mto
TN_216 conflicts vs 12b3e05, median	1.10	1.15
TN_216 conflicts vs 12b3e05, GH-148 auto, median	1.20	1.10
TN_216 probe-CPU sums vs 12b3e05 medians (raw)	1.47	1.40
TN_216 probe-CPU sums vs 12b3e05 medians (drift-adjusted)	1.32	1.26
TN_180 X16 conflicts, variant/auto, median (gain kept)	1.00	0.945

• The conflict medians alone are within 1.2x, but they are no better than [Mac agent] MaxSAT path: stabilizer-overlap dominance constraints in every CSS-half probe #148's. The probe CPU time per conflict stays about 20-30% above off. As a result the variant's time is close to [Mac agent] MaxSAT path: stabilizer-overlap dominance constraints in every CSS-half probe #148's and above 1.2x of 12b3e05.
• TN_180's X16 gain is kept.
• Data is in optimization/experiments/GH-218/nolk/ on that branch. PR GH-218: drop dense (4-cycle-rich) overlap rows from cap 16 on: fixes TN_216, costs TN_180 card-mto (PI decision) #222's choice is unchanged.

[guluchen]

Closed without merging, by PI decision (2026-11-05).

• The cap-16 dense-row drop fixes TN_216 but makes TN_180 card-mto 1.57x slower, which fails the Tier 5 1.2x rule.
• The no-lookahead follow-up (experiment/[Mac agent] MaxSAT: gate -stab-overlap to remove the TN_216 regression (GH-148 follow-up) #218-nolk) also failed: TN_216 came out at an estimated 1.26–1.32x after drift adjustment.
• [Mac agent] MaxSAT path: stabilizer-overlap dominance constraints in every CSS-half probe #148 behaviour stays on main; TN_216 MaxSAT remains about 1.4–1.5x slower than 12b3e05.
• The branches are kept for reference.

🤖 Generated with Claude Code