Persist step/stage state and improve error visibility

[nicpottier]

Summary

Replaces fragile multi-source state tracking (in-memory StageRunJob, in-memory runningSteps, DB step_completions) with a unified step_runs table as the single source of truth. Errors now persist across page refreshes and server restarts, per-step error messages display in the UI, and queued stages correctly show when a book is added.

Key Changes

• New step_runs table (schema v8): Tracks full step lifecycle with status (running/done/error/skipped), timestamps, and error messages
• All step events persisted: Backend writes every step-start, step-progress, step-complete, step-skip, step-error event to DB
• Pure stage state derivation: Stage state is now running > queued > error > done > idle, computed from step states + in-memory run queue
• Per-step error messages: Displayed below step labels when errored (e.g., "LLM rate limit exceeded")
• Fixed "no icon on book add": getQueuedStages() now includes the active run's stage range so not-yet-started stages show as queued
• Improved SSE race handling: Only cancel in-flight fetches if baseline data already loaded; prevents UI hang on page load
• Skipped steps show as done: Steps that are skipped (e.g., translation for single-language book) now render with checkmarks
• Startup recovery: Server startup sweep marks any step_runs with status='running' as status='error' to handle mid-run crashes

Tests

All 623 tests passing, including 13 new tests for step/stage completion logic, queued stages during active runs, and error state preservation.

Files Changed

• Backend: storage layer (new table + migration), stage-runner (event writing), stage-service (queued stage calculation), routes (step-status endpoint)
• Frontend: hooks (context + SSE logic), components (error display), pure helpers (extracted for testability)