Format HMM probabilities without printf in printprob

[kevinmoran1988]

printprob() writes every emission and transition probability as " %8.5f" of -log(p), and glibc's exact-decimal %f path is 8-10% of bathbuild's profile. A float times 1e5 is exact in double, so rint() of that product rounds exactly as printf does, ties to even included. format_prob() builds the digits directly and printprob() writes them with fwrite(). Zero probabilities still print "*" through fprintf, as does anything non-finite or out of range.

Checked exhaustively: all 2,139,095,039 positive floats format byte-identically to glibc's snprintf(" %*.5f").

bathbuild over 5232 dipteraODB12 alignments, serial, 3 runs each, in seconds:

base 230.97 / 232.23 / 229.27
patch 224.54 / 223.97 / 222.84 (1.03x)

Output byte-identical to base on that corpus.

🤖 Generated with Claude Code

[GenevieveKrause]

Thanks for the profiling work here. I'm going to pass on this one — not because anything looks wrong, but because I don't think the benefit is worth the added complexity.

This replaces a single, well-understood fprintf call with a hand-rolled float formatter — manual digit extraction, sign handling, round-to-even via rint(), and a fallback path for cases it doesn't handle. That's meaningfully more code to read, maintain, and reason about in a file that previously just deferred to the standard library. For a 1.03x speedup on bathbuild, I don't think that trade is worth making. If this were a 1.3-1.5x+ win, I'd weigh it differently.

Appreciate you digging into this and sharing the numbers either way.