rbridge 1.0.3 completed benchmark evaluation

All 96 shards and 48 comparisons pass checksum, shape and numerical validation.
Ratios divide the baseline median time by the candidate median time.
Each case uses 15 timed fits per engine, after two warmups; process startup is excluded.
q is the bridge-penalty exponent; n is observations and p is predictors.
The four synthetic designs use seed 2019, q=0.5/1/2, 24 penalties and five CV folds.
CV and BLAS each use one thread; this measures serial engine optimization.

group         median_speedup   min_speedup   max_speedup   cases
-----------  ---------------  ------------  ------------  ------
bridge                  3.14          1.52         14.62      12
rbridge                 2.59          1.18          3.93      12
cv.bridge               2.99          1.31         13.00      12
cv.rbridge              2.57          1.41          3.88      12

Overall casewise median speedup: 2.58x (range 1.18-14.62x).
Maximum coefficient/CV/count relative difference: 5.5105157e-14.
Maximum prediction relative difference: 3.310494e-15.
Selected penalties, support at 1e-7 and variable counts are exactly preserved.

Limits: one machine, one data seed, fixed engine order, no randomized repetitions across sessions.
The reference is the first repaired 1.0.3 candidate, not the original 1.0.2 archive.
No accuracy ranking, convergence certificate, universal speedup or paper reproduction is claimed.
The timer precision notice is retained in provenance.json; solver warnings were errors.
The historical frozen driver and completed source outputs were not modified.