Pin the fifth route_mask merge evidence term: loser's re-parented observation headers (#295 follow-up) #324

Follow-up to #295 (#287, merged as de6e83b4), from the round-3 review of #295, finding 1. It did not block the merge.

recomputeMergedRouteMask combines five evidence terms. The fifth is the header bits of the loser's observations after they are re-parented onto the survivor. That term is still not pinned by a test.

Mutant N2. Move recomputeMergedRouteMask above the UPDATE OR IGNORE observations move, so the scan only sees the survivor's own rows. N2 survives the full cmd/ingestor suite.

Probe RZ1:

• the survivor is computed FLOOD, with one FLOOD observation;
• the loser has route_mask NULL and route_type FLOOD, with one surviving DIRECT observation, so the loser's observation header is the only DIRECT evidence.

N2 is not worse than master, but for this shape it silently rolls the fix back.

Acceptance

• A test with the RZ1 shape, where the loser's re-parented observation header is the only evidence for a bit. It is red under N2 and green on master.
• Test-only.