Rate-limited restarts for plain Erlang/OTP supervisors, accepting the extended supervisor2 restart type without a dependency on supervisor2.

mirrored_supervisor is built on rabbit's supervisor2, which accepts an extended restart type {permanent, Delay} or {transient, Delay} (Delay in seconds). A child spec ported from mirrored_supervisor can carry that form; portunus runs elected children under a plain Erlang/OTP supervisor, which rejects it.

child_spec/1 rewrites such a spec into a standard one whose start is rate-limited to one attempt per Delay: an isolated crash restarts at once, and only a restart within Delay of the previous attempt waits out the remainder. This is not exactly supervisor2, which restarts immediately until MaxR/MaxT is exceeded and only then delays on a timer; the practical property migrated specs rely on (a one-off crash recovers immediately while a crash loop is dampened) is the same. The wait runs in the local supervisor's own process (a supervisor runs a child's start synchronously), so a crash-looping child holds up the others under that supervisor for up to Delay; the running pid is always the real child, so introspection and links are unchanged.

Last-attempt timestamps live in one node-global ETS set owned by portunus_sup, keyed by {LocalSup, ChildId} so entries are namespaced per local supervisor; forget/2 clears one when the child is stopped, so a later first start (for example after ownership moves back to this node) is immediate.