Catch ActiveRecord::ConnectionFailed in the failsafe

[ajaynomics]

What

Adds ActiveRecord::ConnectionFailed to Failsafe::TRANSIENT_ACTIVE_RECORD_ERRORS so that a terminated cache database connection degrades to a cache miss instead of propagating to the caller.

Fixes #307.

Why

ActiveRecord::ConnectionFailed (added in Rails 7.1) is raised when a connection is terminated mid-query. Critically, it is a subclass of ActiveRecord::StatementInvalid, not of ActiveRecord::ConnectionNotEstablished — so even though the failsafe already rescues connection-related errors, this one slipped straight past the rescue list and surfaced to application code. Adding it alongside the existing transient connection errors restores the intended "cache is best-effort, never fatal" behavior.

How it's tested

A new SolidCacheConnectionFailedFailsafeTest reuses the shared FailureSafetyBehavior and mirrors the existing timeout test (the pattern moved to in #285), emulating a terminated connection by stubbing the adapter to raise ActiveRecord::ConnectionFailed. It guards the regression: cache operations drop to a miss rather than raising.

The test helpers were also tidied up — emulating_timeouts and emulating_connection_failures now delegate to a shared emulating_errors(error_class) helper to avoid the near-verbatim duplication.

All test configurations pass (TARGET_DB=sqlite bundle exec rake test): 400 runs, 0 failures, 0 errors across all 8 config matrices. RuboCop is clean.

Notes

No CHANGELOG entry — the repo has never maintained a CHANGELOG file, and recent bug-fix PRs (e.g. #286, #287) add none.

[a-abdellatif98]

@ajaynomics this still applies cleanly and is green: I ran it locally on Ruby 3.4.8 against sqlite, 400 runs, 0 failures.

Flagging it because #317 was opened for the same issue on 15 Aug without knowledge of this one, and #307 now has both. Your emulating_errors(error_class) naming is the better of the two approaches, since emulating_timeouts(ActiveRecord::ConnectionFailed) describes a terminated connection as a timeout. Full comparison in #317 (comment).

The one thing #317 has that this does not is a test asserting non transient errors still propagate, which is the guard against the failsafe list quietly growing until real bugs turn into cache misses. Might be worth pulling that in here so this PR is the complete one.