fix(librarian): re-analyze prose edited via tools and rework analysis scheduling

[Viscerous]

Auto re-analysis only fired from the HTTP route handler, so prose edited by the librarian's own tools (editProse, updateFragment, editFragment) never triggered re-analysis, and the per-passage freshness indicator went stale until the next poll.

• Shared reanalyzeAfterProseChange() so every edit path (route + tools) schedules re-analysis and marks the freshness index stale.
• Reworked the scheduler: removed the 2s debounce in favor of a leading-edge, concurrency-guarded coalescer. Analysis (one of the longest steps) starts immediately when idle, and triggers arriving while a run is in flight coalesce to the latest fragment. Guarantees one run per story at a time.
• Run-scoped hold: while a write-enabled agent run is active, analysis is deferred and fires once on the final state, so a multi-step editing agent triggers a single analysis instead of one per step.
• Resolved the llm/tools scheduler import cycle by deferring the agents import to the one runtime-only call site.

Tests: tests/librarian/scheduler.test.ts (rewritten for the new model), tests/llm/generation.test.ts.

Fixes: #38
Note: touches the editProse loop, so it conflicts with #43, whichever lands second needs a trivial rebase.