Add Elixir and OTP CI coverage and stabilize distributed tests

[jeregrine]

Problem

Group needs automated coverage across supported Elixir/OTP versions without making every PR run fault campaigns or performance benchmarks. The local suite was one module, preventing useful ExUnit concurrency, and started distribution even for local-only runs. Partition tests also raced shard discovery and cleanup, while failures lacked durable peer diagnostics.

Fix

• Run latest-stable formatting, compiler checks and local examples/reference-model histories first, then a six-job runtime matrix with bounded real-peer coverage. Enforce warnings as errors throughout. Only push events on main run CI, avoiding duplicate branch/PR matrices.
• Add nightly fault campaigns at 03:23 UTC and manual/release triggers: multiple seeds, 1/2/4 peer schedulers, larger histories, repeated partition/churn/restart scenarios, and three targeted behavioral mutations.
• Add a separate performance workflow comparing a baseline and candidate on the same runner with a pinned Benchee harness and distributed load/recovery benchmarks.
• Split local examples into independently scheduled subsystem modules, keeping global-tracing and environment-changing modules serial. Start distribution lazily and use configurable, smaller scheduler pools for fresh correctness peers.
• Replace fairness synchronization sleeps with mailbox conditions, wait for shard state around partitions, acknowledge monitor installation, and prevent cross-VM node-name collisions.
• Retain failure output, seeds, generated histories, topology, configuration, bounded shard state, queue/buffer counts and peer diagnostics before teardown.

Supporting information

• Actions use current SHA-pinned releases with read-only repository permissions. PR failure artifacts and campaign reports are retained for 14 days; performance reports for 30 days.
• Runtime ranges exclude prereleases. The supported matrix covers Elixir 1.19 / OTP 26–28, Elixir 1.20 / OTP 27–28, and the latest-stable pair.
• Release publication reruns the workflows but is not automatically gated. Run them manually on the candidate ref and review performance against the previous release before publishing.
• Generated histories are sequential reference-model checks, mutations are focused contract probes, and nightly repetitions are bounded fault/churn campaigns—not exhaustive mutation or long-duration soak coverage. Mixed Group-version wire compatibility is not established by a runtime-version matrix.
• Runtime library behavior is unchanged; lane budgets, replay commands and limitations are documented in test/README.md.