[ptescher]

Step 3 of plans/actors.md: one commander path.

What changed

• recreation_towers_ground now holds the seat client (seat), the station (station: the team's Ground over the air ports, the map from telemetry, Tally for the server's report) and the computer commander (commander::run(addr, team, Brain::Play|Idle, tally, &stop)), plus the binary towers_commander --team blue|red [--host H] [--port P] [--idle]. The commander app uses the same seat/station (its files re-export them).
• Seat protocol additions (existing shapes untouched): Join::Compute(Team), Join::commands(), ToSeat::Orders(Vec<(Option<Order>, bool)>) sent once after the welcome of a seat that commands.
• towers_server: the in-process commanders and the truth-based Ground are gone; the server decides and sends no plan. For a computer (or nobody, which gives no orders) team it starts commander::run in a thread that joins over loopback. Kept in the server: light-quad sticks from stored orders, take-off, referee, air. New flag --on-leave computer|wait (default computer); documented in the server's header and domains/recreation/README.md. A crash now prints a detail line (time, order, low/high, jammed, silent, nearest flying drone).
• Handover: a person (Join::Command) relieves a computer (Join::Compute) holder (it is sent Refused("relieved...") and disconnected) and starts from the orders the drones have; when a person leaves, a computer takes the seat after 1 s unless the person is back. Two people (one per team) can join.

Tests

• cargo test --release -p recreation_towers -p recreation_towers_ground: pass (includes tests/commander.rs: a commander against a fake server gives orders from umpire + sites alone, leaves when relieved).

• cargo test --release -p recreation_towers_server --bins: 29 pass, including the Gate A tests in seats.rs (same message kinds for person and computer seats; a Red scout in the open is in the spectator's truth and not on the map built from a seat's messages; takeover; computer takes a freed seat; two people).

• cargo build --release -p teleoperation_command, cargo fmt, clippy on the changed crates: clean. expectation check expectations: 0 errors.

• Gate B (real ArduPlane SITL, acceptance test): fails on "No drone crashes in a game" for plain actors-plan too (0 to 7 crashes per game here), so neither path passes it reliably. Every other expectation passes (10 in AGENT, 0 refused, 0 fog leaks, towers taken). Crashes per game, 4 games per batch at --speedup 10, one at a time, new path against plain actors-plan interleaved:

  Batch	Path	Crashes per game	Total
  1	new, first version	1, 2, 3, 3	9
  1	actors-plan	0, 3, 0, 0	3
  2	new + AGENT only after the transition, clock cap in wall time	3, 0, 2, 4	9
  2	actors-plan	1, 1, 0, 5	7
  3	new + dead reckoning of reports	1, 1, 0, 0	2
  3	actors-plan	4, 7, 4, 2	17

  Defects of the new path found and fixed: AGENT was asked for in the middle of the take-off transition (two games lost a drone at its base at t = 10 s); the station clock was capped at one game second past the last map (plans slowed at 10x); plans came from the position of the last report (now carried on by its velocity). Apart from those take-off crashes (and one on actors-plan) no crash had another drone within 30 m, so the loss of true enemy traffic does not show; the rest are tree strikes while taking a tower, in both paths.

Not done

• No new .expect steps: handover is covered by Rust tests with seat clients; the arduplane driver only plays computer vs computer (text added to its feature).
• The Orders echo is the server's record of what the drones last heard, not read from telemetry.
• No GUI check of a takeover.

🤖 Generated with Claude Code