feat: extractive transcript summarizer (summarize.py)

[seonghobae]

Why

Transcripts produced by the pipeline are long walls of text — great for archival, poor for skimming. This adds a summarization step that turns any transcript into concise notes with key points, giving real archival value without adding a single dependency.

What

New module summarize.py — an honest extractive summarizer (it selects existing sentences verbatim; it never generates or paraphrases text, and makes no LLM or network calls):

• summarize_text(text, max_sentences=5) -> Summary — dataclass with summary_text, key_sentences (list, original document order), and word_count of the input.
• Algorithm: regex sentence tokenization → word-frequency table over lowercased whitespace tokens with a small built-in English stopword list → sentences scored by average content-word frequency (so long sentences don't win on length alone) → top-N returned in original order so the summary reads chronologically.
• summarize_segments(segments, max_sentences=5) — accepts any iterable of duck-typed objects exposing .text (e.g. timestamped transcription segments); joins them so sentences spanning segment boundaries are reassembled.
• Edge cases: empty/whitespace input → empty Summary; input with ≤ max_sentences sentences passes through unchanged; max_sentences < 1 raises ValueError; CJK sentence terminators (。！？) recognized; Korean/CJK input handled without crashing. Limitation documented in the module docstring: whitespace tokenization means unspaced CJK scripts get coarse frequency scoring.
• Stdlib only (re, dataclasses). Full docstrings.

Tests

tests/test_summarize.py — 19 tests, no network, no mocks of external services:

• top sentences selected by frequency (dominant-term sentences beat filler)
• original document order preserved in key_sentences and summary_text
• empty and whitespace-only input
• short input returned as-is; single sentence without terminal punctuation
• max_sentences respected; < 1 raises ValueError
• stopword-only sentences never win
• very long input (500 sentences) produces a bounded summary
• Korean text summarized without crashing; full-width 。 splitting behavior
• deterministic tie-breaking by earlier position
• segment joining, boundary reassembly, empty segment list, duck typing

Gates (python3.14):

• unittest discover -s tests: 132 tests, no new failures (only the 5 pre-existing macOS os.listxattr errors remain)
• interrogate -c pyproject.toml: 100% docstring coverage repo-wide
• .coveragerc source list untouched; no existing files modified

🤖 Generated with Claude Code