37d38ee4

refactor: the trap reads a register, not four hand-kept lists

a73x   2026-08-26 18:34

refactor: the trap reads a register, not four hand-kept lists

cleanup() named every daemon, socket and capture the suite makes: a kill
list of 137 $DnnPID variables, a stop list, a 350-name rm list and the
backstop beside it. Adding a leg meant four edits somewhere else, and the
edits got missed — 102 captures were absent from those lists by
2026-08-19, enough that `make soak` refused to start.

Now the statement that creates a thing registers it. start_daemon spawns,
registers the pid and the socket, and waits for the bind in one call;
defer_kill, defer_sock and defer_rm take everything a leg makes by hand.
Captures are not registered at all: every one is spelled from $OUT, and
the trap removes that whole set by the pattern leak_sweep already reads it
by — a pattern cannot forget a leg.

cleanup() goes from 542 lines to 117, and the 82 declarations that existed
only so the trap could name them under `set -u` go with it. The trap's
plain `kill` becomes softkill on the way past: a signal sent to a kcov
tracer is dropped, which is why the hub leg had to learn that already.

The long-lived daemon becomes $D1PID. start_daemon owns $DPID, and the
scenario 5,000 lines below that asks whether the tear killed the daemon
needs a name no later spawn can overwrite.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>