[TheNetherWatcher]

coverage_mask located each peptide with re.finditer(f'(?={peptide})', protein). A lookahead pattern can't use the regex engine's literal prefix search, so every peptide costs a full regex scan of the protein, and each occurrence then adds 1 over its slice of the mask.

This finds the same overlapping occurrences with protein.find(peptide, i + 1), records +1 where an occurrence starts and -1 where it ends, and returns the running sum. Empty peptides are skipped (they matched everywhere before but added nothing). The result is the same array with the same dtype.

On a random 35,000-residue protein with tryptic peptides:

Peptides	Before	After
2,000	0.42 s	0.023 s
4,606	0.98 s	0.054 s

coverage() calls coverage_mask(), so it gets the same speedup.

Checks:

• New tests for overlapping and repeated occurrences, empty and missing peptides and an empty protein. They pass on the old code too.
• 200,000 random cases (2-, 3- and 20-letter alphabets) give identical masks and dtype to the old implementation.
• A Lean 4 proof that the find loop yields exactly the lookahead matches and that the running sum equals the old mask at every residue: https://github.com/TheNetherWatcher/lean-perf-proofs/tree/main/pyteomics-coverage-mask

One difference, outside the documented input: a peptide containing regex metacharacters was matched as a pattern before and is now matched literally. Peptides are documented as plain one-letter sequences, and fast=False strips them to that.