Paginate active thread work without losing context

[SawyerHood]

Summary

• replace disruptive top-level message cuts with lazy, expandable active-work boundaries in the existing “Worked for…” presentation
• retain the initiating user message plus a bounded mutable live tail while paging older work exactly once
• preserve source ordering across same-turn work, delegated child turns, summaries, realtime updates, and scroll restoration
• enforce timeline row and aggregate UTF-8 payload budgets in SQLite before event payloads cross into JavaScript
• scope and sign pagination cursors, including turn context and delegated interval identity
• keep active query pages stable across frontier updates and preserve row references only when render semantics are equivalent

Validation

• Turbo typechecks passed for app, server, DB, server-contract, SDK, and thread-view across the implementation slices
• DB suites passed, including full 349-test validation plus focused byte-budget and wide-delegation regressions
• focused server validation passed up to 100 tests after rebasing
• focused app validation passed up to 66 tests after rebasing, plus 28 controller tests for semantic row updates
• contract, SDK, and thread-view focused suites passed
• Prettier and git diff --check passed
• each implementation slice was reviewed independently with GPT-5.6-sol; all completed gates currently approve

Still in progress

This PR is intentionally draft and must not be merged yet. The remaining performance slice is the conversation-outline path: replace replay of every streaming delta with a compact incremental outline projection and one bounded cache entry per thread.