fix: address review findings for leader election recovery

[ssteele110]

Summary

Addresses findings from multi-perspective code review of #92.

Critical fixes:

• current_generation made public — fixes BuildRecord respond_to? always returning false (requeue data was silently lost)
• Fenced push() via Lua script — prevents stale master from overwriting active generation after lock expiry
• Lock refresh during population — prevents TTL expiry on large test suites
• wait_for_master detects immediate nil status (workers no longer spin 120s when master died before they started)

Code quality:

• Removed duplicate generation_key/learn_generation/generation_stale? from Worker (inherits from Base)
• Base generation_key raises instead of silent fallback to non-generation keys
• Supervisor rescues MasterDied alongside LostMaster
• generation_stale? check moved to idle path (reduces Redis GETs per poll iteration)
• Default master_lock_ttl increased from 30s to 120s

Tests (8 new):

• Election retry on master death
• Fenced push rejects stale master
• Generation staleness exits poll loop
• learn_generation raises when key missing
• max_election_attempts exhausted raises LostMaster
• BuildRecord reads generation-scoped requeue key
• Supervisor handles MasterDied
• wait_for_master detects immediate nil status

Test plan

• All 8 new tests pass
• Full test suite passes
• Manual verification of fenced push under simulated lock expiry