Self-review of 3a69448 (plus an adversarial pass still in flight) turned up two
real defects in the work I just shipped:
1. THE STASH LIFECYCLE LEAKED ACROSS SESSIONS AND EGGS. _seat_defaults was
keyed by ROW INDEX and initialised once in __init__ -- never cleared on Start
Session, on Open/New egg (which rebuilds the table), or on a seat-count
change. Concrete harm: stop a session while a seat is OCCUPIED (the stash is
only popped when a seat empties, so it survives), reconfigure or open a
different egg, start again -- and when that seat next empties the GUI
restores LAST SESSION'S snapshot over the new configuration. Fixed by (a)
keying the stash by TAG, the seat's actual identity, so a changed roster
makes stale keys inert instead of landing on whatever row now sits at that
index, and (b) clearing it at both natural boundaries: session start (both
local and remote paths -- the clear sits before the remote branch) and the
table rebuild in _load_egg_into_ui.
2. POSITIONAL hostID->tag MAPPING IS WRONG AFTER A LAUNCH TRIM -- and 3a69448
widened its blast radius. The GUI resolves REGISTERED/dropped lines via
host_tag(), position into the UNTRIMMED egg roster; after a trim remaps seat
ids (e.g. the lone returning player reclaim-held at seat 3 becomes host 2),
those lines light the wrong row -- and, with the new seat_info pop on
dropped, CLEAR the wrong seat's identity, leaving the real one's name up:
the exact reported bug, resurfacing in the trimmed case. Fixed on both
sides: the relay's REGISTERED and dropped prints now carry tag='...' from
its LIVE (possibly trimmed) roster via a new _tag_of() helper, and the GUI
regexes prefer the explicit tag with the positional fallback kept for old
logs. SEATED/READY/LEFT already carried tags; only these two were
positional.
VERIFIED
* test_operator_roster.py grown to 22 checks, all passing: post-trim lines
light and clear the RIGHT row while the wrong row stays untouched; the
legacy no-tag format still resolves positionally; a stopped-while-occupied
session cannot leak GHOST into the next session's empty seat; a stale
foreign-tag stash never touches a live row.
* Live 2-pod rig: both new formats confirmed on the wire
(REGISTERED (1/2) tag='...' / dropped: closed tag='...'), mission launched
and stopped cleanly, round RESET intact. rearm 19/19, net 17/17.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
