[naseweisssss]

Current output

benchmarks.Pmt.time_pmt    ok

======== ====== ============ =============
                         when
--------------- --------------------------
  size    rate      end          begin
======== ====== ============ =============
 scalar   0.0    19.3~0.4us    19.5~0.4us
 scalar   0.05   19.4~0.3us    19.6~0.3us
   1      0.0    20.4~0.3us    20.4~0.3us
   1      0.05    26.1~8us     20.7~0.3us
   10     0.0    20.8~0.3us   21.0~0.09us
   10     0.05   20.9~0.2us    20.7~0.2us
  1000    0.0    34.8~0.4us    34.7~0.3us
  1000    0.05   46.6~0.5us    47.2~0.4us
======== ====== ============ =============

For nonzero interest and end-of-period payments, a scalar call took 19.4 µs, compared with 20.9 µs for 10 loans and 46.6 µs for 1,000 loans.

These results support the hypothesis that fixed per-call overhead dominates small inputs. This PR does not discuss about the causes of this but just to show the results of the benchmark.

[naseweisssss]

Added a scalar fast path in 157712e: ordinary floating-point scalar calls use NumPy scalar arithmetic without allocating arrays. Arrays, Decimal and other types retain the existing path. Paired timings show 6.9–7.9× faster scalar calls; 196 tests and lint/type checks passed on the implementation branch. Details: #36 (comment)