pow is a difficult routine to implement, and not all implementations give good results. Roughly speaking, the core algorithm for pow(x, y) computes a logarithm from (a part of) x, multiplies it by y, and computes an exponential function on the product. Doing this in floating-point introduces rounding errors that are hard to control.
The result is that the computed result for pow(10, 4) may be something near 10,000 but slightly less or greater. If it is less, than converting it to an integer yields 9999.
When you use arguments hard-coded in source code, the compiler may compute the answer during compilation, possibly using a different algorithm. For example, when y is three, it may simply multiply the first argument by itself, as in x*x*x, rather than using the logarithm-exponent algorithm.
As for why the low result happens with the odd numbers you have tested, consider what happens when we multiply 5.45454545 by various powers of 10 and round to an integer. 5.45454545 rounds down to 5. 54.5454545 rounds up to 55. 545.454545 rounds down to 545. The rounding up or down is a consequence of what fraction happens to land beyond the decimal point. For your cases with pow(10, loop), the bits of the logarithm of 10 may just happen to give this pattern with the few odd numbers you tried.
与恶龙缠斗过久,自身亦成为恶龙；凝视深渊过久,深渊将回以凝视…
