fix(routing): expire stale pending-text + name session from its first message

[Time4Mind]

Problem

A message typed while no active session exists is stashed in context.user_data["_pending_text"] and forwarded once a session is created. user_data is in-memory and only a bot restart clears it — so a stash can survive for hours and then be injected into an unrelated session created much later.

Field incident (2026-06-28)

• 01:05 — user sends «почему смартфон на 100%». Bot had just restarted and marked all windows lost, so _resolve_active_window found no active session → opened the directory browser and stashed the text.
• 01:05–01:09 — the auto-restore path rebuilt the previous session and delivered the text there, but never popped _pending_text.
• No further restart for ~9.5 h → the stale stash lived on in memory.
• 10:34 — user creates a new session; _session_create forwards the leftover 01:05 text into a brand-new window, which then catches the user's unrelated «найди сессию про глотание / альфа страховка» message and gets auto-named medical insurance.

Forensics confirmed in bot.log: single Starting Telegram bot at 01:02, No active session: showing directory browser at 01:05:32, then Forwarding pending text … len=93 at 10:34:26 with no intervening re-stash.

Fix

• stash_pending_text / take_pending_text in handlers/directory_browser.py (TTL PENDING_TEXT_TTL_S = 10 min). Consumption drops a stale stash and always clears the slot, so a leaked pending message can never re-fire. Legacy bare-string format tolerated for state in flight across deploy.
• _session_create and window_picker now consume via take_pending_text (one clearing read) instead of get-then-pop.
• _session_create seeds the auto-name from the forwarded pending text (the session's real first human message) rather than whatever inbound lands first. maybe_auto_name stays one-shot via its re-entrancy guard.

Tests

• tests/test_pending_text_ttl.py — fresh round-trip, stale-drop, TTL boundary, custom max-age, no-expiry, legacy string, absent/empty, None-safe.
• Full suite: 736 passed; ruff + pyright clean.

🤖 Generated with Claude Code