Back up CLI --resume transcripts so idle tabs stay resumable

[SergKam]

Problem

Tabs left idle for weeks fail to resume with "No conversation found with session ID", even though the full chat history is still on screen.

Root cause: claude --resume <id> reads the CLI's own transcript at ~/.claude/projects/<encoded-cwd>/<id>.jsonl — not our chat history. The CLI prunes those by age (cleanupPeriodDays, default 30 days), so a long-idle workspace loses its backing file and can no longer be resumed.

Fix

Keep a rolling backup of each session's CLI transcript under ~/.config/flycrys/sessions/transcripts/<id>.jsonl, so it survives the CLI's cleanup (and manual cleanup, disk wipes, machine migration).

• Off the main thread. The locate + copy run on a worker thread during autosave — the GTK main loop only spawns it. Sync copy only on shutdown so the final turn is captured before exit.
• Zero idle churn. Gated on chat-history growth (reuses the existing autosave signal), so it only runs when a conversation actually advances. A stat-only up-to-date check skips redundant copies.
• Safe writes. Atomic temp+rename copy; a reader never sees a half-written transcript.
• Encoding-robust. Computes the CLI project-dir name from cwd as a fast path, falls back to scanning project folders (session ids are UUIDs, so a match is unambiguous).

Note on recovery

This PR is backup only — no auto-restore on resume failure (by design). Resume keeps working because the live transcript is preserved; pair this with a raised cleanupPeriodDays so the CLI stops pruning the originals. If an original is ever lost, restoring is a manual file copy from the backup dir back to ~/.claude/projects/<encoded-cwd>/.

Tests

Unit tests cover the project-dir encoding, the up-to-date check, and the atomic copy (no temp files left behind). cargo clippy --lib --bins -- -D warnings clean; full suite green.