feat(scheduler): ADR-0020 slice 1b schedule-through-gate

[chinkan]

Summary

ADR-0020 slice 1b: scheduled turns go through TurnGate with the same session_key(bot_id, user_id) serialization as user turns.

• TurnGate::begin_scheduled: idle → mark Scheduled and return; busy → ticket in waiting_scheduled and await until finish hands the gate (user always finishes first; leftovers stay pending behind).
• Job runner (main.rs): begin_scheduled before process_message_outcome, finish after every outcome (success / max-iter / error). ADR-0013 re-fires already share this runner, so they use the same path.
• finish → Next::User: spawn pending user turns via spawn_telegram_turn_loop (extracted from 1a Submit::Start). Bare agent loop fallback when no owning Telegram bot / unparseable ids.
• During a scheduled turn, user messages still enqueue (👀 from 1a) and are not drained (drain already no-ops for TurnKind::Scheduled; no steer source when schedule_id.is_some()).
• /stop clear: still drops only the user queue; waiting scheduled tickets are kept (unit-tested).

Out of scope

• 1c — voice / media on the queue path
• Portal 202 / generation identity
• P1: Evaluate schedule cron in local timezone (config override) #166 timezone
• Exa

Refs: ADR-0020 slice 1b

Test plan / QA matrix

• Unit: schedule serialisation in turn_gate (begin_scheduled blocks behind user; never two active; no drain during scheduled; waiting scheduled before leftover pending; /stop clear keeps tickets)
• cargo fmt
• cargo clippy --locked --all-targets -- -D warnings
• env -u RUSTFOX_GOOGLE_OAUTH_CLIENT_ID cargo test --locked (947 passed / 0 failed)
• Manual Telegram:
  1. Start a long user turn (e.g. a multi-tool prompt). While it runs, fire a scheduled task for the same bot+user — the schedule should wait and only reply after the user turn ends.
  2. While a scheduled run is active, send a chat message — expect 👀, no steer into the schedule reply; after the schedule finishes, a user turn should start with that pending message.
  3. /stop during a user turn that has a waiting schedule behind it — queue cleared, scheduled run still proceeds.

[chinkan]

TL ACCEPT — tip c48c3c1 (ADR-0020 slice 1b)

Reviewed against ADR-0020 Q8 / Product GO:

• begin_scheduled: idle → Scheduled; busy → oneshot ticket, user finishes first
• No steer into schedule (drain empty while TurnKind::Scheduled)
• User messages during schedule enqueue (👀 from 1a), then finish → Next::User via spawn_telegram_turn_loop
• /stop clear drops user queue only; waiting scheduled tickets kept (unit-covered)
• Waiting scheduled beats leftover pending; job runner wraps begin_scheduled / finish around existing ADR-0013 path

Unit coverage for the above looks solid. Merge when QA GO (CI green + manual three) lands. No tag.

Non-block (later): job runner stays globally serial (pre-1b); a long begin_scheduled wait can delay other users' jobs. Fine for personal single-user; revisit only if multi-bot concurrency matters.