0.3.1 — Assistants hardening

Hardening the assistant, after an adversarial review of 0.3.0. Nine defects in the concurrency, recovery and kill-switch guarantees the release advertised — every one now a regression test that failed on 0.3.0. Nothing changes what an assistant may do; it changes what can happen to a case while she is doing it.

• A customer's next question can no longer be answered around (R1). Every speaking path locks the ticket, then the conversation row chats updates inside every message insert, before reconciling — so an in-flight message blocks the answer until it commits and makes the turn stale. Holds on PostgreSQL and MySQL; on SQLite it is documented and doctor warns.
• Seats are decided under the lock (R2); assign! / release! take a turn: when by: is an assistant — the one API change.
• A lost registration is repairable (R3): the fold commits on its own, and redispatch_assistant_turns! looks for unregistered messages before it looks at clocks.
• Equal-timestamp messages no longer deadlock the turn (R4): the watermark is (created_at, id) in chats' transcript order, one rule in Ruby and SQL.
• The silent sweep rechecks before it escalates (R5).
• Stranded seats come back (R6): SupportDesk.reclaim_assistant_seats! / rake support_desk:reclaim_assistant_seats read the seats that exist, with no timeout and no clock; doctor seat checks run whenever assistant rows exist.
• Running out of turns publishes :ticket_escalated (R7).
• Disclosure on old messages stops moving (R8): name and mode are snapshotted onto the assistant row.
• The kill switch is read after waiting for the lock (R9).

Hosts with no assistant are untouched. Gem SHA-256: 0b7bd2310212553a911aa0643ef872aff63aae901a13308b689f709e4dd5b886