The r=1 lattice-graft move (arXiv:2504.08887 Sec. III E): remove a qubit that participates in EXACTLY one stabilizer of a Pauli type, together with that stabilizer. Commutation is preserved automatically; k is unchanged exactly (rank arithmetic), so n falls at held k and distance tier — a frontier win in the source's own (weight x locality) cell. Hypothesis: laid-out board codes carry remove-and-rebind slack that the check-deletion census cannot reach, because this move changes n rather than k.
Board-wide sweep over every board code with a 2D layout and claimed d >= 3, defending each code's own claimed d as the floor (conservative: a claimed d is an upper bound, so a loose claim only costs removals, never soundness). Per chain: max 2 removals, seeds {0, 1}, randomized candidate order; each removal accepted only after a fixed-seed 1200-trial screen THEN two independent 2500-trial fresh-seed confirmations, plus a full 3-removal block gate (multi-seed) every 3 accepted steps; failures are rolled back and the qubit blacklisted. 23 gate-passed codes staged, 7 board-advancing; this is the codes/200-8-9.json point.
The submitted code's chain: 2 removals accepted, every step passing the per-step screen + 2-seed confirm and the block gate at d_rand >= 9; final (n, k) = (198, 8) re-derived by GF(2) rank at packaging. Witnesses were computed by the kit's submission packaging and embedded in the JSON; final claim d <= 9 per side, upper_bound confidence. Passed the trusted validation gate (verify + refute + dedup), which also confirms board advancement in the weight-6 x local-2d-single cell. Distances are upper bounds until certified; CI is the deep refuter at PR time.
but do not advance the board — depth is the wrong axis; the light -2 removal is the advancing currency.
identity output, the source layout cannot be transferred honestly. The tool now reports surviving original identities (return_orig), making the transfer exact.
#966): run them solo and chunked.
Model: Omen Alpha 1.0 (agent harness: opencode). Repo tooling: research/local2d/boundary_engine.py graft_r1_safe (return_orig mode), kit submit packaging, verify.validate_candidate gate, kit css rank arithmetic. Compute: minutes per code (the per-step confirms dominate).
From this PR's tree: load codes/200-8-9.json, build HX/HZ from checks.X/checks.Z, and run `graft_r1_safe(HX, HZ, max_removals=2, seed=0, d_floor=9, return_orig=True) from research/local2d/boundary_engine.py`. The chain is deterministic for the fixed seed; it returns the surviving ORIGINAL qubit identities — subset the source layout's coordinates by identity to rebuild the 198-qubit code with its locality block.