Avoid trailing post-replacement delimiters in modern hard cuts

[Eric3-jp]

With algorithm='modern', a hard cut through repeated delimiters introduced by a post replacement can leave a trailing delimiter:

slugify(
    'value',
    algorithm='modern',
    replacements=[('value', 'one--two')],
    replacement_stage='post',
    max_length=5,
)
# Before: 'one-'
# After:  'one'

Keep a run of delimiters with the next nonempty token, and only emit it if at least one character from that token fits. This also handles multi-character separators, leading delimiter runs, and the hard-cut fallback from word-boundary mode. A delimiter that occurs in the word itself is preserved.

This complements #193, which preserves post-replacement output when the whole slug already fits. That fast path, whole-word selection, the default/explicit legacy algorithm, and public smart_truncate are unchanged.

Validation on Windows with CPython 3.14.0:

• The new regressions fail on the original implementation (6 failing subcases and 1 failing fallback test).
• python -m pytest --color=no -q: 125 passed, 117 subtests passed, including the unchanged legacy suite, the 2,688-case legacy differential, and fitting-post-replacement coverage.
• Strict Mypy 1.20.2 and repository-configured pycodestyle/flake8 commands pass.
• git diff --check passes.

The complete supported-interpreter/tox matrix has not been run locally.

OpenAI Codex assisted with the investigation, patch, tests, and this description. The checks listed above were executed locally.

[un33k]

Thank you, @Eric3-jp — verified locally and this is a solid fix. A modern hard cut through repeated post-replacement delimiters was leaving a trailing delimiter (one-); keeping the delimiter run with the next nonempty token and only emitting it when a character fits correctly yields one. I confirmed it handles multi-character separators, leading delimiter runs, and the word-boundary fallback, while a delimiter that occurs inside a word is preserved.

Local verification: full suite 125 passed / 117 subtests, the 2,688-case legacy differential passes (legacy output unchanged), the #193 fitting fast path still holds, mypy clean, and repo pycodestyle/flake8 pass. Confined to _modern_truncate; the legacy algorithm and public smart_truncate are untouched. Merging now.

🚀 Generated with Dojo ⛩️