Add PromoteMaster command for manual master promotion

[alytaphoenix]

Summary

Adds an API endpoint to force a specific node to win the next master election, for operators who want to move mastership without waiting for organic re-election (e.g. moving master off a node they're about to take down for maintenance).

• POST /master/promote/{node_id} sends a PromoteMaster command, rejected with 409 if any instances are running (every node's worker/download-coordinator restarts on master change — see exo.main._elect_loop — so this is idle-only) and 404 if the node isn't in the topology.
• Election._force_promote boosts the target node's seniority strictly above the highest seniority it has ever observed from any peer (floored at the same FORCE_MASTER_SENIORITY that --force-master uses at startup), then triggers a new election round.

A fixed constant here would let a second promote tie the first on seniority and fall through to the commands_seen tiebreak, which favours whichever node has been master longest — silently no-opping the newer promotion. Tracking the highest seniority observed and always going one higher keeps repeated promotions, and promotions away from a --force-master node, working. This was caught with a regression test (test_second_promote_still_wins_over_previously_promoted_master) that reproduces the exact scenario; verified by reverting to the fixed-constant version and confirming the test fails as predicted, then passes with the fix.

Known limitations

• Sticky master: like --force-master, a promoted node's seniority only ever grows, so it stays master for the rest of its process lifetime and won't hand back automatically. A later PromoteMaster targeting a different node still works correctly (see above). If the promoted node crashes, peers re-elect normally, and a restarted node rejoins at seniority 0.
• Not yet exercised on a real multi-node cluster — covered by unit tests against Election and the API handler in isolation.
• No dashboard UI for this yet (API-only); a "promote to master" control in TopologyGraph is a natural follow-up.

Test plan

• uv run basedpyright — 0 new errors (baseline unchanged: 319 errors, 110 warnings)
• uv run ruff check — passes
• uv run ruff format — applied
• uv run pytest src/exo/shared/tests/test_election.py src/exo/api/tests/test_promote_master.py — 23 passed, including a regression test that fails against the pre-fix code
• Broader uv run pytest src (excluding pre-existing environment gaps unrelated to this change — mlx not installed in this dev environment) — 330 passed, 3 skipped
• Manual verification on a real multi-node cluster

🤖 Generated with Claude Code

https://claude.ai/code/session_011rjSfwDBTkmySmfU6NgHKF