The shared prerequisite for both live steps was: does a next/ kernel process hold gen_server state
(flow_store) across HTTP requests? Confirmed yes. plans/ra_kernel.erl is a minimal kernel
(flow_store + register the publish-digest flow, then a blocking http:listen that keeps the
er-scheduler + gen_server alive); plans/ra-kernel-spike.sh boots it as a background sx_server and
drives it with two SEPARATE curls: GET /start suspends instance 1, GET /resume resumes that SAME
live instance → done. So durable suspend→resume across requests works on a persistent kernel.
Design decision (per the discussion): chose the persistent-kernel path (B) over host-side replay-log
(A). B serves BOTH durability (RA) and federation (TA) on one fed-sx-native substrate and exposes the
full next/ kernel (projections, outbox, actor model); A only solves flow durability and mixes Erlang
into the host process. The er-scheduler-context bug (which kills an in-process kernel, option C) does
NOT bite a separate-process kernel — er-bif-http-listen spawns each handler in-scheduler, so
gen_server:call completes. Gotchas recorded: a blocking listener hangs any in-process
erlang-eval-ast (the kernel must be a dedicated TCP-driven process), and binary =:= is buggy (always
true) so routes must pattern-match paths as byte-list binaries.
RA-live + TA-live are now BUILD work (a real kernel service + the host as HTTP client + the actor
model), not research — the prerequisite is proven.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
