Un-decoded **h leaks into this corpus's own text (needs an em dash), e.g. br-k07.yaml

Found while investigating #70 (stripping \`**h\` from the reference comparison): \`**h\` also appears literally in this corpus's own \`data/*.yaml\` \`text\` fields, in a handful of files -- most heavily \`data/fiction_general/br-k07.yaml\`, also \`br-k13.yaml\`, \`br-k24.yaml\`, and \`data/miscellaneous/br-h05.yaml\` (18 instances total).

Unlike #70's reference-side instances (which correspond to nothing at all in this corpus's real text), every one of these reads naturally as an em dash marking a pause or interruption, e.g.:

- \`"the dumb jerk **h Coughlin grinned, feeling supremely on top of things **h He watched the snow once again."\`
- \`"There it is **h and there is what was in it."\`
- \`"\`Lift, don't shove **h lift\`".\`

This corpus's own escape convention for this transcription (\`&\`/\`+\` for periods/ampersands, \`^\` for diaeresis, etc.) apparently never got applied to this specific marker in these few sentences, leaving the raw \`**h\` escape sitting in the text unconverted -- the same class of leftover-transcription-artifact bug as #9 (em dashes corrupted to a spaced hyphen) and #60 (a literal \`^\` diaeresis marker leaking through), just for a different original escape code.

Needs its own fix (replacing \`**h\` with this corpus's own already-established em-dash rendering) and verification against \`brown-nolines.txt\` before touching \`data/*.yaml\` -- not folded into #70, which only changes the comparison tool.

Inspect the 18 occurrences in data/fiction_general/br-k07.yaml, br-k13.yaml, br-k24.yaml, and data/miscellaneous/br-h05.yaml, comparing them with brown-nolines.txt first. Replace only the transcription artifacts with the corpus's established em-dash rendering, then verify all affected text against brown-nolines.txt.

Written by the indexing model from the issue text.