feat: eval output, scroll --until, session cleanup, snapshot compact

[detrin]

Summary

Four agent-friendliness improvements based on real-world usage bottlenecks:

1. eval returns output (feat: make eval return output to stdout #23) — dict/list results are JSON-serialized to stdout. Agents no longer need to fall back to MCP playwright for data extraction.

2. scroll --until <selector> (feat: add scroll --until for infinite scroll patterns #24) — scrolls in increments until an element becomes visible. Fixes the infinite-scroll pattern that fails the benchmark. Usage: brow scroll -s 1 --until "text=Article 35"

3. session cleanup (fix: auto-cleanup stale sessions on daemon restart #25) — releases all sessions. After agent crashes, a simple brow session cleanup unblocks profile locks without needing to identify individual stale sessions.

4. Snapshot auto-compact (feat: snapshot --compact mode to reduce token usage on large pages #26) — pages >500 lines auto-truncate to interactive elements + structural landmarks, with a hint to use --search. Prevents accidental 300K+ token dumps. Explicit --compact flag also available.

Closes #23, #24, #25, #26

Test plan

• brow eval -s 1 "result = {'hello': 'world'}" prints JSON
• brow scroll -s 1 --until "text=Article 35" on infinite scroll page
• brow session cleanup releases all sessions
• Large page snapshot shows truncation header
• Unit tests pass (pytest brow/tests/test_cli.py test_snapshot.py test_client.py)
• Lint passes