Take back the two features that only worked on one machine

[davidmckayv]

What this changes

Reverts #15 (the approval registry) and #17 (the repetition detector), and puts the deployment rule on the pull request template.

Why

Both features keep state that outlives a request in a Map inside the server process. That is correct on one machine and wrong the moment a deployment runs a second one, which is the deployment this product is aimed at: several server processes behind a load balancer, serving a whole company, with people bringing their own agents.

Neither fails loudly, which is why they are being taken back rather than left in place and fixed later:

• A question raised on one process and answered on another is reported as no longer open. That is indistinguishable from an expiry, so nobody investigates.
• Repetition counts split across processes, so a rule written as repeat.count >= 10 never fires. That is indistinguishable from a Bot behaving itself.

A boundary that quietly stops enforcing is worse than one that was never advertised, because the deployment is relying on it.

What stays

Everything that was already right. The stall watchdog from #19 tracks open streams in the process that holds them, which is where they belong. The client-side work in #16 and #19 is unaffected. The Did not happen audit filter keeps agent.stream_stalled and loses only the repetition entry.

What goes

The ask policy list and its surface, the approval routes and registry, the repetition context on the policy engine, COMPUTER_REPEAT_WINDOW_MS, and the computer.action_repeated and approval audit kinds. Migration 0001_gigantic_sumo goes with it; it added one column, action_policy.ask. Any database that already applied it keeps a defaulted column nothing reads, and drizzle-kit generate reports no drift.

Still to do

The gateway resolves every action's target from a snapshot cache that is also a Map in the process. It predates all of this and it is load-bearing for the action boundary, so it is the next thing to move to Postgres.

Proof

format:check, lint and typecheck clean. 647 pass, 5 skip, 0 fail. db:generate reports no schema changes.