feat(grouper): deterministic QUALITY GATE filter + broader noise strip

[belaytzev]

Summary

Architect's call after PR #32 merge analysis: stop trusting the LLM to follow negative drop rules. Add a deterministic regex/heuristic filter pass between extraction and dedup. Move QUALITY GATE prompt rules into the extractor as belt-and-suspenders.

This is the next iteration on top of PR #32. After running `/digest` on real bot, five categories of noise survived all the prompt-level rules:

Surviving noise	Example	Root cause	Fix here
Admin chatter	`🆕 В чате появился новый участник Denis Nogtev`	gpt-5.4-mini ignored prompt	Deterministic regex
Meta-empty	`🛸 В подборке упомянуты темы, но без дополнительных деталей`	gpt-5.4-mini ignored prompt	Deterministic regex
Speculation	`📊 Парк выглядит как сильный фотоспот`	gpt-5.4-mini ignored prompt	Hedging+entity check
Section header leak	`📌 Key points: (Вастрик Селфхостеры🔧, kyrillic🌍, ...)`	Strip regex didn't cover	Broader noise strip
Numbered emoji prefix	`1️⃣ 🍏 Intel и Apple...`	Cosmetic clutter	Strip in normalizer

Architecture change

Pass 2a extract → [NEW] QUALITY GATE filter → dedup chokepoint → Pass 2b classify
                  ^^^^^^^^^^^^^^^^^^^^^^^^^^
                  Deterministic. No AI dependency.

Same pattern as the dedup chokepoint introduced in PR #32: pull noise removal out of prompts and into code. Prompt rules stay as belt-and-suspenders but the actual safety comes from `_quality_gate_filter`.

QUALITY GATE rules (deterministic)

• Admin chatter: `r"новый участник|joined the chat|появил.{0,30}участник"`
• Meta-empty: `r"без (?:дополнительных )?(?:деталей|подробностей)|no details?|just a poll"`
• Too short, no facts: <30 chars AND no digits AND no @-handle AND no URL AND no proper noun pair
• Speculation without entity: matches `r"probably|похоже|вероятно|выглядит как"` AND no concrete entity

Substantive bullets pass through untouched (validated by `test_keeps_substantive_bullets_unchanged`).

Broader noise strip in _strip_channel_summary_noise

Five new regex patterns added to the existing 🚀 + 📎 strip:

• `📌 Key points:` / `Ключевые моменты:` / `Puntos clave:` / etc — multilingual
• `1️⃣` ... `9️⃣` standalone numbering prefix
• `[emoji]`, `[brief fact]`, `[brief subject]`, `[link]` template token placeholders

Why move QUALITY GATE into extractor prompt too

Architect: "will help marginally because the task is simpler, but you're still asking a weak model to follow negative rules at low temperature on Russian text. Don't bet on it."

So this is belt-and-suspenders only. The deterministic filter is the load-bearing fix. Prompt now explicitly says "DROP rules OVERRIDE extract-verbatim" to resolve the conflict the model previously had between rules.

What's next (out of scope for this PR)

Per architect, the bigger remaining wins:

1. Embedding-based fuzzy dedup (text-embedding-3-small, ~$0.0001/day) — would catch paraphrased duplicates that exact-string dedup misses (e.g. "Apple договорились с Intel" vs "Intel будет производить часть чипов Apple"). Solves cross-group duplication root cause.
2. Structured summarizer JSON output instead of prose — eliminates the entire class of template-leak bugs at the source.

Test plan

• 410 passed, 7 skipped (was 400 in PR refactor(grouper): split into 2a extractor + 2b classifier with dedup chokepoint #32, +10 new tests)
• Coverage 79.18%
• `uv run mypy src/` — no issues
• `uv tool run ruff check src/ tests/` — clean
• `uv run black --check src/ tests/` — clean
• `uvx flake8 --max-complexity=10 src/ tests/` — clean
• Run `/digest` on real bot, verify admin chatter / meta-empty / speculation gone

Files

• `src/grouper.py` — `_quality_gate_filter`, broader noise strip patterns, extractor prompt update, wire into `group_summaries`
• `tests/test_grouper.py` — `TestStripBroaderNoise` (4 tests), `TestQualityGateFilter` (6 tests), updated `test_classifier_error_propagates` to use substantive bullet