Stop navgen seating waypoints past a ledge lip

[saworbit]

navgen samples hull 1, the player clip hull. That is the world expanded by the player box, so its floor reaches half a player width past every real ledge edge. The sampler read that as standable ground.

On e1m6 it seated n93 at -207 305 24. The true lip at that y is x -193, so the seat sat 14 units out over a 304 unit drop with lava at -128.

x	hull 1 floor, what navgen samples	hull 0 floor, the real world
-215	-280	-306
-207	24	-306
-199	24	-306
-191	24	-2

The engine never agreed. SV_movestep calls SV_CheckBottom, which point traces the box corners against the world, so walkmove refuses the step and Argus_MoveHazard refuses it too. A bot routed onto such a seat can neither arrive nor give up. A co-op companion on e1m6 spent its whole session at that lip: 213 stalls and 215 jumps in 175 s, mode 0 on 310 of 327 samples, three goals, never a death.

The fix

Section 4c runs the engine's own test instead of a guess at one, with the same constants and the same order. Corners solid just under the feet is the fast accept. Otherwise the midpoint must find floor within 2 * STEP, and every corner must sit within STEP of it.

e1m6, regenerated

	nodes	seat by the lava lip	seats over a 96 unit void	worst spawn reach
before	198	-207 305 24	32, 11 of them over lava	85%
after	188	-175 305 24	0	95%

536 lip overhang samples dropped. The seat moved onto real ledge rather than disappearing. Stalls ran 220, 213 and 200 across three tapes on the old graph, and 6, 2 and 5 on the new one.

Other maps are not regenerated

Those graphs are accumulated and tuned, so each wants its own ladder. Dry runs confirm the rule does not fragment them.

map	seats dropped	worst spawn reach
dm4	212	98%
dm6	209	95%
e1m1	501	99%

dm4 in engine, 180 s: cells 75, stalls 18, routefails 9, no freezes, against 72 / 51 / 17 on the pre-change binary.

Honest residual

e1m6 is still not co-op ready, and this change does not make it so. The old graph's only route to the gold key crossed the lava pit through n93. Removing the fake seat removes the fake route, so routefails run 400 per tape where they ran 9. That crossing is a func_door which rises on a trigger_once chain and stays up, and in a botmatch nothing fires it, so the route genuinely does not exist at runtime. Modelling the mover in navgen would only move the lie to a different height. What e1m6 needs is level progression handling, which is much larger work. Tracked on #249.

The trade this PR takes is a wedged bot for an honest graph. Stalls fall from about 210 to about 5, coverage moves 13 cells to 9, and routefails rise. By the project's own gate order that is the right way round, but the routefail rate is worth watching on a multi bot server, since an_busy serialises the router.

Tried and reverted, each on its own ladder

• A swimmer reach ceiling measured from the water surface rather than the bot. Offline it looked decisive, dm3 drag-down 45% of wet samples to 0%. It readmitted 1585 1185 208, the exact seat the graveyard entry names. Both that seat and dm3's lip sit 36 above their surface, so no height rule separates them.
• A reach test on the Argus_NearestNode fast path. It is a real hole, the neighbour loop applies no reach test and returns early at 120 units. It improved dm3 but cost e1m2 co-op half its coverage, 55 cells to 24 on two tapes.
• A routefail shelf on the co-op key commitment. It did not touch the storm, because the retries come from the goal already on the stack rather than from re-commitment, and it cost e1m2 routefails 0 to 77 and stalls 40 to 73.

Refs #249

🤖 Generated with Claude Code

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, you can upgrade your account or add credits to your account and enable them for code reviews in your settings.