fix(ssml): spell_out splits combining marks and regional-indicator pairs #2825

Summary

backend/services/ssml_lite.py:spell_out iterates by Python str codepoints, so a single visible character built from a base plus one or more combining marks (NFD-decomposed) is split into multiple space-separated units. The same happens to paired regional-indicator sequences (flag emoji), even though the JS-side SSML-LITE path was already fixed by 55f3d58 fix(ssml): spell out by grapheme.

Repro

import unicodedata
from services.ssml_lite import spell_out

# NFD: 'é' = e + U+0301
print(spell_out(unicodedata.normalize('NFD', 'café')))
# c a f e  ́   (the combining acute stands alone; TTS will drop it)

# regional indicators: '🇺🇸' = U+1F1FA U+1F1F8
print(spell_out('🇺🇸'))
# 🇺 🇸          (a single character announced as two)

The correct output is the same as the composed / precomposed form (c a f é, 🇺🇸). The same bug does not exist for NFC text because precomposed characters are a single codepoint, but the function is reachable from longform_parser.py for any audiobook / longform line that originated in a Mac keyboard buffer or an editor that emits NFD (Pages, TextEdit, Vim with decompose on, copy-paste from a web page).

Suggested fix

Segment by grapheme cluster using only stdlib: NFC-normalize first (handles decomposed Latin/Hangul), then group trailing unicodedata.combining(c) characters onto the previous base. The narrow fix covers the common audiobook cases without adding a new dependency. Regional-indicator / ZWJ emoji remain a separate, smaller follow-up (a real grapheme library would be the right tool, and would also have to land in the locked deps).