Scope and reproduction
----------------------
Run the exact command in run.json using the existing Python environment. This
script uses only the standard library and derives all transfer links from
literal physical coordinates. Production files are hashed for context only;
no production solver, exchange map, boundary helper or CUDA code is imported.
Only this newly created evidence directory was written by this analysis.

The inputs are unit basis coefficients, NOT finite-density physical LBM states.
Fine active populations pass unchanged between hops. F1/F2 therefore denote
transport/BB stages of an identity-transfer probe, not an SRT validation. These
results do not establish a bug in the shipped backend (which does not implement
this wall-crossing mode) or in any external library.

Fixed geometry
--------------
Coarse 16x16, centers (x+.5,y+.5), dx=dt=1. Patch [4,12)x[4,12).
Fine 24x24 with halo 4, origin (2,2), centers origin+.5*(index+.5), dx=dt=.5.
Halfway planes Y=8 and Y=15; fluid occupies 8<Y<15, periodic x.
Parent/child fluid classification is checked over the entire fine footprint.
Only fluid active cells count as fine leaves; solid targets are absent.
Every valid fine ghost is exploded once. Inner fluid ghosts stream/reflect but
never collide; outer halo is absent after each hop. Invalid is distinct from 0.
Coalescence uses independently derived closed link midpoints and four children
with weight 1/4, and excludes solid TARGETS, not solid-source directions.

Three proposed extensions, none generally conservative in these probes
---------------------------------------------------------------------
1. Coarse BB then coalescence: final writer for intersecting slots is F->C.
2. Coalescence then coarse BB: final writer for solid-source slots is coarse BB.
3. Omit solid-source coalescence slots: coincides numerically with option 2 here.
All alternatives assign each final destination once; source ownership still
duplicates or loses packets, as the numeric totals in stdout/results show.

Measured packet path for Gc(q7,x2,y8)
-----------------------------------
Gc0(7,2,8) --explosion--> Gf0(7,1,13)
 --hop1 pull--> F1(7,2,12) --hop2 local BB--> F2(8,2,12)
 --weight .25 coalescence--> Fc1(8,3,8).
Option 1 also retains the full independent coarse BB at Fc1(8,2,8), producing
total leaf weight 1.25 from 1. Options 2/3 avoid this particular duplication,
but the adjacent coarse pulse produces 1.25 from 1, and the fine-active pulse
loses its entire initial leaf weight .25. Thus neither blanket precedence nor
blanket solid-source filtering resolves the seam.

Both local postpull BB and an independently implemented prestream solid-slot
write yield exactly equal values AND validity sets for both fine hops on all
three probes. This establishes equivalence only for the explicit footprint and
operations in this script; it is not a reproduction of an external library.

Next inference, not implemented
------------------------------
A correct wall/interface transfer needs per-packet geometric ownership beyond
these unmodified straight-link four-child rules. Preserve this counterexample
and both controls when proposing that ownership. No scaling fix, interpolation,
wall-shear solver run, GPU experiment, or production change was attempted.
sources.csv records all nonzero stages; results.json includes recursive final
coarse ancestry, weights, parameter scope, and hashes. stdout.txt is the raw
single-run output. run.json records the observed exit and resource limits.