feat: add speech normalization for md files

[StuBehan]

Summary

Adds stackvox.text — a reusable speech-text normalization mechanism that turns real-world text (Markdown, Claude responses, prose) into something that sounds right when synthesized. Two consumers (behan.codes read-aloud and stackone-speaklast) had grown separate, diverging copies of this pre-step; this extracts the shared mechanism into StackVox while leaving each caller's pronunciation dictionary as per-consumer config. See docs/speech-normalization.md for the full proposal and API contract.

Changes

• New stackvox/text.py module with normalize_for_speech() as the primary entry point, plus each stage exposed for composition/testing: markdown_to_paragraphs, strip_emoji, strip_thousands_separators, decimals_to_words, expand_units, apply_pronunciations, shape_pauses, ensure_terminal_stop.
• Markdown → prose reduction (headings/list items become their own paragraphs), locale-keyed unit/symbol expansion (£1.63 → "1.63 pounds"), decimals → words (1198.9 → "1198 point 9"), pause shaping, emoji stripping, and caller-supplied pronunciation dictionaries.
• Backward compatible — existing speak/synthesize behaviour is unchanged; normalization is opt-in.
• Export normalize_for_speech from the package root (stackvox/__init__.py).
• Add docs/speech-normalization.md proposal.

Testing

Ran the full CI-equivalent suite locally, all green:

ruff check .            # All checks passed!
ruff format --check .   # 20 files already formatted
mypy                    # Success: no issues found in 18 source files
pytest -q               # 141 passed

Related issues