Strip **h/**H typesetting marker from the reference comparison

\`**h\`/\`**H\` has no corresponding character anywhere in this corpus's own data -- confirmed against several instances in \`data/press_reportage/br-a12.yaml\`: \`"40 per cent **h their total passing yardage"\` and \`"86 tries **h. Tailback"\` in \`brown-nolines.txt\` both have nothing at all where \`**h\` sits in this corpus's own \`text\` (\`"per cent their"\`, \`"tries. Tailback"\`). Unlike \`**f\`, which is tangled up with the already-tracked formula-placeholder gap (#16/#34), \`**h\` is clean reference-side noise.

Dropped in \`decode_reference_text\`, the same as the paragraph-break tokens, swallowing an optional leading space so two real words end up with exactly the single space that was already between them, while a trailing character with no space of its own (\`tries **h.\` -> \`tries.\`) is left in place.

Divergent word-line count drops by 508 with no \`data/*.yaml\` changes.

Start in decode_reference_text and inspect how paragraph-break tokens are handled, using the examples in data/press_reportage/br-a12.yaml and brown-nolines.txt as reference cases. Strip the optional leading space with **h/**H while preserving the character after a token with no space of its own; done is a 508-line drop in divergent word-line counts with no data/*.yaml changes.

Written by the indexing model from the issue text.