Frozen targeted exploration, V12, 8 November 2026.
Written before executing the new search. Exploratory mathematics, not a statistical
preregistration. No LLM calls. No six-point exhaustive enumeration.

Seeds: the two non-isomorphic four-point pairs certified in C12.
A: {00,01,02,10,13,21,22,30} / {01,02,10,11,13,21,22,30}.
B: {00,01,02,11,13,21,23,30} / {00,02,10,11,13,21,23,30}.
For each seed pair use identical parameters on both targets:
1. One new vertex: all 2^9 choices of its incoming/outgoing base attachments and loop.
   512 per seed; 1024 parameter cases, target n=5.
2. Two new vertices with the same attachment pattern to the four base vertices,
   but arbitrary 2x2 relation among the new vertices: 2^8 * 2^4 choices per seed.
   8192 parameter cases, target n=6. Attachments are directed; new internal loops allowed.
3. Positive full blow-ups of each base vertex, for every ordered weight vector with
   total 4 through 8: 70 per seed, 140 parameter cases.
Total: 9356 parameter cases, not a count of distinct isomorphism types.

For each pair compare exact full labelled sample-pattern distributions at N=2,3,4.
An equal N=4 distribution certifies agreement on EVERY loopless test with <=4
vertices by inclusion-exclusion and adding isolated vertices. Survivors are compared
at N=5. Exact arrays/sorting, not probabilistic hashing, decide equality. Store the
first separating size (when <=5), an actual edge-set witness and its two hom counts.
Extract a witness from a maximum-cardinality differing pattern, then greedily remove
edges while its two counts differ. This is not claimed globally edge-minimal.
If equality persists through N=5, test isomorphism exactly and report separately;
no larger search is launched automatically. Do not infer a global b_Lin(6,2) value.

Controls: original seeds; relabelled identical graphs; independent Python/Numpy
verification of all N=4 survivors and each reported witness; fixed-seed samples of
rejected candidates. Derive algebraically what uniform blow-up and isolated padding
do to the least separating order, so known collision inheritance is not called growth.
