Demo dataset v2: realistically overgrown inbox (84,000 messages, 1,419 senders)

[zbuc]

Closes #17

Implements the dataset v2 design from issue #15.

Summary

• demo/summary.json regenerated: 84,000 messages across 1,419 senders in 12 categories — social-notification floods at the top (chatterloop 6,842; fotofling 4,517), heavy retail promo and newsletter mass, transactional receipts/finance/platform mail, a 90-sender gray-mail band, 120 human correspondents, and a 400-sender one-off signup tail.
• demo/dataset_spec.json — the machine-readable spec (category budgets, Zipf exponents, clamps, pinned top senders, brand vocabularies, fixed seed).
• demo/generate_summary.py — deterministic generator: pinned heroes take exact counts, Zipf tails are jittered/clamped and reconciled to exact per-category budgets, one-off groups use small count mixtures. It asserts every guarantee before writing: totals and sender counts match the spec exactly, addresses are unique, every domain is RFC 2606 reserved.
• Distribution realism: single top sender ~8% of volume, median sender has 11 messages, 368 senders appear exactly once.

Testing

• Generator runs clean with all assertions passing; regenerating from the committed spec is byte-identical (fixed seed).
• Rendered the page in headless Chrome: all 1,419 rows populate, tiles show 84,000 / 1,419 / notifications@chatterloop.example.
• summary.json is ~103 KB — fine for Pages.

No UI changes; the existing demo banner and constraints (fictional brands, reserved domains, no real data) are unchanged.