fix(npc): roll while walking, and refuse an out-of-box destination #238

Merged

sickday merged 3 commits from fix/npc-wander-ratchet into main 2026-08-19 01:47:29 +00:00

Conversation 0 Commits 3 Files changed 2 +151 -31

sickday commented 2026-08-19 00:15:37 +00:00

Owner

Copy link

NPC wander composed two correct rules into a ratchet. hold/1 keeps the
destination when a creature is in the way, and roam/1 refused to roll
while it had one -- the client's own pathLength == 0 guard. Together,
two NPCs whose next steps cross can neither move nor choose anywhere
else; nothing clears it, so every jam is permanent and the population
accumulates into knots rather than settling into a wrong distribution.

Measured over the whole spawn table, 16,942 roamers against real
collision and real entity flags, the share of NPC ticks lost to a
creature in the way ran 0.6% -> 8.2% -> 37.2% -> 72.3% at ticks 1, 10,
100 and 1,000, and was still climbing. Mean other NPCs within three
tiles went 1.43 -> 2.44 over the same run, with the share of NPCs in a
knot of four or more going 20% -> 38.7%. That is the clustering an
observer sees, and it has nothing to do with pathfinding.

The roll now fires whether or not the NPC is already walking. The source
states the client guard and then contradicts it for the server, from the
game: NPCs move in one direction and then suddenly the opposite one, so
wandering may occur even with movement already queued. A roll that names
nowhere legal leaves the journey alone -- it named nowhere, which is not
the same as naming here.

The leash now refuses an out-of-box destination where it clamped one to
the box edge. The clamp had no source: the client's clamp is against the
map edge, and the wander box is a server concept the client does not
have. It loaded the boundary -- over 200,000 ticks a clamped walk spent
9.6% of its life on the four corner tiles, which are 3.3% of the box.
Refusing is the rule the module already applied to a zero offset.

Refusal needs one companion, because a stray can no longer be clamped
home: an NPC outside its box walks back to its spawn instead of rolling.
Only a fight can put it there, and World.NpcCombat already does this
for a creature that breaks its leash.

After: held flat at ~5% from tick 25 onward, crowding flat at 1.33 and
17.9% -- below what the spawn table itself starts at -- corner occupancy
0.68% -> 0.27% per tile, and the median NPC's distinct tiles visited
29 -> 63.

Three new tests, each checked to fail against the old behaviour.

NPC wander composed two correct rules into a ratchet. `hold/1` keeps the destination when a creature is in the way, and `roam/1` refused to roll while it had one -- the client's own `pathLength == 0` guard. Together, two NPCs whose next steps cross can neither move nor choose anywhere else; nothing clears it, so every jam is permanent and the population accumulates into knots rather than settling into a wrong distribution. Measured over the whole spawn table, 16,942 roamers against real collision and real entity flags, the share of NPC ticks lost to a creature in the way ran 0.6% -> 8.2% -> 37.2% -> 72.3% at ticks 1, 10, 100 and 1,000, and was still climbing. Mean other NPCs within three tiles went 1.43 -> 2.44 over the same run, with the share of NPCs in a knot of four or more going 20% -> 38.7%. That is the clustering an observer sees, and it has nothing to do with pathfinding. The roll now fires whether or not the NPC is already walking. The source states the client guard and then contradicts it for the server, from the game: NPCs move in one direction and then suddenly the opposite one, so wandering may occur even with movement already queued. A roll that names nowhere legal leaves the journey alone -- it named nowhere, which is not the same as naming here. The leash now refuses an out-of-box destination where it clamped one to the box edge. The clamp had no source: the client's clamp is against the map edge, and the wander box is a server concept the client does not have. It loaded the boundary -- over 200,000 ticks a clamped walk spent 9.6% of its life on the four corner tiles, which are 3.3% of the box. Refusing is the rule the module already applied to a zero offset. Refusal needs one companion, because a stray can no longer be clamped home: an NPC outside its box walks back to its spawn instead of rolling. Only a fight can put it there, and `World.NpcCombat` already does this for a creature that breaks its leash. After: held flat at ~5% from tick 25 onward, crowding flat at 1.33 and 17.9% -- below what the spawn table itself starts at -- corner occupancy 0.68% -> 0.27% per tile, and the median NPC's distinct tiles visited 29 -> 63. Three new tests, each checked to fail against the old behaviour.

sickday added 1 commit 2026-08-19 00:15:37 +00:00

fix(npc): roll while walking, and refuse an out-of-box destination ... 951b8ebb7f

NPC wander composed two correct rules into a ratchet. `hold/1` keeps the
destination when a creature is in the way, and `roam/1` refused to roll
while it had one -- the client's own `pathLength == 0` guard. Together,
two NPCs whose next steps cross can neither move nor choose anywhere
else; nothing clears it, so every jam is permanent and the population
accumulates into knots rather than settling into a wrong distribution.

Measured over the whole spawn table, 16,942 roamers against real
collision and real entity flags, the share of NPC ticks lost to a
creature in the way ran 0.6% -> 8.2% -> 37.2% -> 72.3% at ticks 1, 10,
100 and 1,000, and was still climbing. Mean other NPCs within three
tiles went 1.43 -> 2.44 over the same run, with the share of NPCs in a
knot of four or more going 20% -> 38.7%. That is the clustering an
observer sees, and it has nothing to do with pathfinding.

The roll now fires whether or not the NPC is already walking. The source
states the client guard and then contradicts it for the server, from the
game: NPCs move in one direction and then suddenly the opposite one, so
wandering may occur even with movement already queued. A roll that names
nowhere legal leaves the journey alone -- it named nowhere, which is not
the same as naming here.

The leash now refuses an out-of-box destination where it clamped one to
the box edge. The clamp had no source: the client's clamp is against the
map edge, and the wander box is a server concept the client does not
have. It loaded the boundary -- over 200,000 ticks a clamped walk spent
9.6% of its life on the four corner tiles, which are 3.3% of the box.
Refusing is the rule the module already applied to a zero offset.

Refusal needs one companion, because a stray can no longer be clamped
home: an NPC outside its box walks back to its spawn instead of rolling.
Only a fight can put it there, and `World.NpcCombat` already does this
for a creature that breaks its leash.

After: held flat at ~5% from tick 25 onward, crowding flat at 1.33 and
17.9% -- below what the spawn table itself starts at -- corner occupancy
0.68% -> 0.27% per tile, and the median NPC's distinct tiles visited
29 -> 63.

Three new tests, each checked to fail against the old behaviour.

sickday added 1 commit 2026-08-19 01:03:56 +00:00

Merge branch 'main' into fix/npc-wander-ratchet 192bbb7b82

sickday added 1 commit 2026-08-19 01:41:04 +00:00

Merge branch 'main' into fix/npc-wander-ratchet 0f921e1975

sickday merged commit 0f921e1975 into main 2026-08-19 01:47:29 +00:00

Sign in to join this conversation.

Reviewers

No reviewers

Labels

Clear labels

No items

No labels

Milestone

Clear milestone

No items

No milestone

Projects

Clear projects

No items

No project

Assignees

Clear assignees

No assignees

1 participant

Notifications

Subscribe

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference

Revenant/Server!238

No description provided.