Visible bug: with 1 wall ghost queued, all 3 pawns picked the same site;
2 stood idle while 1 built. Same shape would affect chop/mine/haul/etc.
Design: Job carries an untyped target_node ref (the tree/rock/build-site/
crop/item/patient/workbench/etc the job is acting on). Job.is_target_taken_
by_other(target, excluding_pawn) does an O(pawns) scan of live job state to
ask 'is anyone else already working this?'. Each WorkProvider's find_best_
for() now skips claimed targets in its scan and sets j.target_node before
returning. No per-entity claim state, no .claim()/.release() bookkeeping,
no save-format change — target_node is intentionally not serialized
because pawns re-decide and re-bind naturally after load.
Providers updated: construction / chop / mine / plant (harvest path) /
hauling (item AND corpse) / cleaning (target is Vector2i tile not Node;
field is untyped, doc'd) / doctor / crafting (workbench).
Not touched: rest (everyone shares the rest tile, that's fine), eat /
sleep (food and beds have their own availability gates; flagged as a
followup if multi-pawn food contention surfaces).
Verified MCP runtime: fresh boot, 3 pawns picked 3 distinct wall sites
(44,28)/(45,28)/(44,27) with distinct target_node refs.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
megaproxy
e7a2407af2