link-draft: Add config file and ML-based tag/author classifier

[metaist]

Summary

Refactor link-draft.py to use a config file (link-draft.toml) as single source of truth for mappings, and add a light ML classifier for predicting tags and other metadata.

Design

Config file: link-draft.toml

Committed to repo. Training updates it in-place; user reviews git diff and commits what's correct.

# Auto-updated by: link-draft.py --train
# Last trained: 2026-02-03T15:30:00Z

[author_domains]
"simonwillison.net" = "Simon Willison"  # seen: 12x
"astralcodexten.com" = "Scott Alexander"  # seen: 8x

[domain_tags]
"x.com" = ["tweet"]
"github.com" = ["programming"]
"amazon.com" = ["book"]

[keyword_tags]
"\\bAI\\b" = ["ai"]
"\\bLLM\\b" = ["ai", "llm"]

[known_people]
names = ["Jeremy Wertheimer", "Shalev NessAiver"]

[rejected.author_domains]
# Entries here won't be re-suggested by training

Training pipeline

• Scans all existing link posts
• Learns patterns weighted by recency (recent edits override older patterns)
• Updates link-draft.toml in-place
• Trains sklearn classifier for tag prediction, saves to _ignore_link-draft-model.pkl

Training flags

• Auto-train if link-draft.toml missing or older than newest post
• --skip-train to bypass auto-training
• --train to force retrain

Draft generation flow

1. Load static data from link-draft.toml (user overrides take precedence)
2. Load ML model for fuzzy predictions
3. Combine: static rules first, model fills gaps

Tasks

• Define link-draft.toml schema
• Extract hardcoded mappings from link-draft.py to config
• Add training pipeline to scan posts and update toml
• Add recency weighting (recent posts weighted higher)
• Train sklearn classifier (e.g., SGDClassifier or MultinomialNB) for tags
• Add --train, --skip-train flags
• Add auto-train logic (train if config older than newest post)
• Add [rejected] section support
• Update draft generation to use config + model

Notes

• Training should be fast (<5 seconds for ~324 posts)
• Model file is gitignored; static config is committed
• If user deletes a suggestion and it keeps coming back, add to [rejected] section

---

Issue created by Claude Opus 4.5