Leader election recovery with generation isolation

[ssteele110]

Summary

Redesigns master election to handle worker death mid-population.

• Master lock with CAS: Lock acquired via SET NX EX with TTL, renewed during population via Lua CAS script. Final push uses a second Lua CAS to verify ownership before committing queue data — prevents split-brain when lock expires while master is still alive.
• Generation isolation: Queue-operational keys namespaced by UUID (gen:{uuid}:*). Each population attempt gets its own isolated keyspace. Build-wide aggregates (error reports, flaky reports, stats) remain unscoped.
• Automatic re-election: Workers detect master death (lock TTL expiry during setup), re-elect a new master, retry up to max_election_attempts (default 3).
• Staleness detection: Polling workers periodically check current-generation (throttled to 1 check/5s) and exit gracefully on mismatch.
• Bug fix: try_to_reserve_lost_test passed key('completed') to reserve_lost.lua, but Lua uses KEYS[2] as processed_key. Since acknowledge writes to processed (not completed), the guard was always false — acknowledged tests were being re-stolen after deadline.

New files

• redis/renew_master_lock.lua — CAS lock TTL renewal
• redis/push_queue.lua — CAS queue commit with ownership verification
• ruby/test/ci/queue/redis_generation_test.rb — 8 test scenarios

Test plan

• 8 new generation test scenarios pass (election, CAS rejection, staleness, retry, supervisor)
• All existing unit tests pass (0 new failures)
• 22 pre-existing integration test failures unchanged