feat: bot-per-persona — each persona as its own Telegram bot (#29)

[mattmezza]

Closes #29.

Run each persona as its own Telegram bot — its own contact in your address book. The persona that handles a message is determined by which bot received it.

Design — one composite channel name

A persona-bot registers as agent.channels["telegram:<persona>"]; the default bot stays bare telegram. Inbound passes channel="telegram:<persona>" to process(). That one convention does three jobs at once:

1. History isolation — in a Telegram DM chat_id == user_id for every bot, so the composite name is what stops two bots' DMs with the same user from sharing a (channel, user, chat) key.
2. Outbound + approval routing — agent.channels["telegram:<persona>"].send(...), scheduler job.channel="telegram:<persona>", and approvals reply via the bot that fired them (the channel string is threaded from process() straight to _await_approval → self.channels.get(channel)).
3. Persona resolution — telegram:<name> → persona <name> with no lookup (new top rung of the _resolve_persona ladder from Per-chat persona binding + isolated contexts (channel-agnostic) #14).

What changed

• Persona model — optional bot_token (+ optional allowed_user_ids, else inherit the global list). Empty token ⇒ not a bot, reachable only via the default bot. Stored on the personae table (new columns + migration for existing DBs); round-trips through markdown frontmatter and the admin editor.
• main.py — same process, N bots: the default bot plus one TelegramChannel per persona that carries a token, each polled in-process. One bad/revoked/duplicate token can't take down the others (or WhatsApp/the scheduler); partial bring-up failure tears down already-started bots so none are left orphaned.
• _resolve_persona — top rung parses telegram:<name>; falls through to the existing ladder for unknown names.
• Scheduler/jobs — channel accepts telegram:<persona> (back-compat: bare telegram = default bot, no migration). A persona job is generated as that persona while keeping the system execution mode (auto-approved writes, no memory/reflection).
• Admin UI — an Own Telegram bot card on the persona editor (token + allowed users); the token is redacted on read like the global one.
• Docs — personae.mdx (new Bot-per-persona section, rung 0, roadmap repointed to Group multi-agent rooms — turn-taking + loop guard #30), channels.mdx, README.

Out of scope (per the issue)

Group multi-agent / turn-taking (#30); per-bot config beyond token + allowed-users.

Quality

Built understanding-first, then implemented, then ran two multi-agent adversarial review passes:

• Review — 4 reviewer lenses (correctness, issue-completeness, concurrency/lifecycle, security), each finding independently verified against the real code. 13 confirmed findings; the substantive ones are fixed in this branch (startup-abort/orphan-poller hardening, persona-correct scheduled jobs, token redaction, token dedup, no cross-bot progress spam, no dead topic auto-bind on persona bots).
• Verify — 6 adversarial probes over the fixes: 0 broken, 0 regressions (it actually removed two pre-existing shutdown/startup bugs). GO.

All 418 tests green, ruff clean. New tests cover composite-channel resolution, persona-bot field persistence + token redaction, and persona-correct scheduled delivery.

[mattmezza]

The real test: two Telegram bots (the only true end-to-end)

1. Two tokens from @Botfather. Keep one as your default (channels.telegram.bot_token); the second goes on a persona.
2. Run the agent locally: make dev-agent (uvicorn + admin UI at http://localhost:8000/admin).
3. Admin UI → edit a persona (e.g. coach) → Own Telegram bot → paste the second token. Leave Allowed user IDs blank to inherit the global allowlist — but make sure your Telegram ID is in the global allowlist, or the bot will silently ignore you (_is_allowed). Save.
4. Restart the agent — tokens are read once at startup (_start_agent's bot loop), so a persona-token change needs a restart: the admin "Restart now" button (it re-runs _start_agent → re-reads the personae DB → spins up the new bot), or Ctrl-C + rerun.
5. You now have two contacts in Telegram. Verify the four guarantees:

• Resolution — DM the coach bot → it answers as the coach (voice/identity/scoped tools).
• History isolation (the headline fix) — tell the default bot a fact, then ask the coach bot about it → it won't know. Both DMs have chat_id == user_id, so the telegram:coach channel is the only thing keeping them apart.
• Approval routing — ask the coach bot to do a write action (one that needs approval) → the Approve/Deny buttons arrive on the coach bot, not the default.
• Scheduler — create a job with delivery channel telegram:coach (admin Jobs tab, or tools/jobs.py with --channel telegram:coach) → delivered by the coach bot, written as the coach.

6. Resilience (optional — exercises the review-hardening): set a persona's token to garbage → restart → the default + other bots still come up (log: Failed to start Telegram bot telegram:coach — skipping). Give two personas the same token → the second is skipped with a warning.