Follow-up to #2356 addressing four issues that would bite in production:

1. channel_send: poll_options silently dropped non-string entries.
   The previous filter_map(as_str) pattern turned ["a", 42, "c"] into
   ["a", "c"], slipped past the min-2 check, and sent a poll with the
   user's third option missing. Extract a parse_poll_options helper
   that fails fast with a per-index error when any element is not a
   string, and pin it with eight unit tests covering the type cases,
   the missing-array case, and the 1/11 bounds.

2. telegram: PollContext.last_accessed was never updated on access.
   poll_contexts.get() returned an immutable ref, so the background
   cleanup removed entries 30 minutes after send regardless of whether
   users were still answering — long-running surveys silently lost
   their poll_question / poll_options metadata. Switch to get_mut so
   each answer bumps last_accessed, making the field honest.

3. telegram: poll cleanup task did not respond to shutdown for up to
   5 minutes. The loop had a bare interval.tick().await, so a stop()
   call would block until the next tick. Wrap the tick in a
   tokio::select! against shutdown.changed() and exit immediately when
   the signal flips.

4. telegram: api_send_media_group accepted 1 or 11+ items and let the
   Telegram API reject them with an opaque 400. Validate locally that
   the slice has 2..=10 items before composing the request.

…2364)

Follow-up to #2356 addressing four issues that would bite in production:

1. channel_send: poll_options silently dropped non-string entries.
   The previous filter_map(as_str) pattern turned ["a", 42, "c"] into
   ["a", "c"], slipped past the min-2 check, and sent a poll with the
   user's third option missing. Extract a parse_poll_options helper
   that fails fast with a per-index error when any element is not a
   string, and pin it with eight unit tests covering the type cases,
   the missing-array case, and the 1/11 bounds.

2. telegram: PollContext.last_accessed was never updated on access.
   poll_contexts.get() returned an immutable ref, so the background
   cleanup removed entries 30 minutes after send regardless of whether
   users were still answering — long-running surveys silently lost
   their poll_question / poll_options metadata. Switch to get_mut so
   each answer bumps last_accessed, making the field honest.

3. telegram: poll cleanup task did not respond to shutdown for up to
   5 minutes. The loop had a bare interval.tick().await, so a stop()
   call would block until the next tick. Wrap the tick in a
   tokio::select! against shutdown.changed() and exit immediately when
   the signal flips.

4. telegram: api_send_media_group accepted 1 or 11+ items and let the
   Telegram API reject them with an opaque 400. Validate locally that
   the slice has 2..=10 items before composing the request.

…hardening

Brings in:
- feat(channels): Telegram rich media, polls, interactive commands, channel_send tool (librefang#2356)
- feat(runtime): config-driven session_mode for agent triggers (librefang#2341)
- fix(agents): refuse direct DELETE on hand-spawned agents (librefang#2361)
- fix(channels): harden poll_options parsing and poll context cleanup (librefang#2364)

Conflicts resolved: bridge.rs (new trait methods), prompt_builder.rs
(9-param build_channel_section merging bot_username/group_members with
granted_tools), tool_runner.rs (cron-tz tests + poll_options tests).