Skip to content

Walking maps: step joins where players walk across in the capitals (#2860) - #2916

Open
fitzforge wants to merge 9 commits into
feature/city-pathsfrom
city/joins-steps
Open

fitzforge wants to merge 9 commits into
feature/city-pathsfrom
city/joins-steps

Conversation

@fitzforge

@fitzforge fitzforge commented Nov 18, 2026 •

Copy link
Copy Markdown
Owner

Issues

Builds on #2907 (the step and drop kinds); merge that first.

What and why

Places in the capitals where recorded players walk across and the walking maps had no way, or only a long way round: a door's threshold, a counter's gap, a platform's steps. Found from the players' own recordings: where a recorded walk jumps across a gap in our grid (two breadcrumbs on one level the grid can't join), the spot is clustered over players, the closest pair of floors across it is the step, and the shipped engine is asked whether it walks across. Kept only with 5+ distinct players stepping through (training players only: the gate's held-out players never chose a fix) and both sides on floors 3+ players walk. A second source, a navigation mesh's connectivity, proposed more spots; each was held to the same recordings evidence.

  • 28 joins: Ironforge 16, Stormwind 11, Orgrimmar 1. 26 are steps (a straight walk, at most 12 yd), two are routed on our grid.
  • Left out: a Stormwind step at 40.83, 49.01 (5 players). It joined the ground the city was built over, which lies under the streets, to the harbour side, so a walk starting on that buried ground found a 620 yd way out instead of none (nav_maps' walk from ground joined only far away).
  • Strict steps. A step's way gets an arc wherever the walking maps lack a move of their own, even where they walk round to it within 40 yd: that way round was the start of the long detours (Ironforge 36.36, 67.82: 86 yd round for 25). joins.json pairs may also be marked "strict": true.
  • Ironforge Path Routing #2860: Daryl Riknussun's platform (the cooking trainer) is joined to the Great Forge's floor. Before, the walk at his height found no way, and the walk asked again without his height left by the gate and crossed the mountain (3,515 yd).
  • Walking maps repacked.

Gate (held-out recorded walks in the capitals)

Same judge and baseline as #2907 (the shipped maps repacked unchanged). This pull request includes #2907's drop.

Baseline: the shipped maps repacked unchanged, city gate at the same commit (GAME = baseline-5's numbers in every city). After = this pull request, which includes #2907's drop; the step joins' own share is Ironforge's.

City GAME before GAME after WALK before WALK after
Orgrimmar 24 8 26 10
Ironforge 10 7 12 9
Stormwind 12 12 11 11
Thunder Bluff 15 15 15 15
Darnassus 20 20 20 20
Undercity 19 19 21 21
all 100 81 105 86

No walk newly fails in any city. Fixed in Ironforge: three walks onto Daryl Riknussun's platform, in both arms. The Stormwind steps change no held-out walk (their spots are walked rarely by the held-out players); they are kept on the recordings' evidence.

Checks

  • cargo test -p forge-nav --lib pack:: (a strict way gets its arc where the maps only walk round)
  • New nav-reported tests: the walk to Daryl Riknussun at his height from the reporter's spot is found, ends on his platform and is under 450 yd (before: no walk), and back; two of the busiest steps walked both ways at under 1.5x the straight line plus 20 yd
  • forge nav check; nav_maps, nav_memory and nav-reported pass on the repacked maps; forge data subzones rerun, nothing changes
  • Without the buried-ground step: the city gate in Stormwind, Orgrimmar and Ironforge is the same in both arms (SW 12/11, Org 8/10, IF 7/9)
  • Broad sample (2,000 walks): 8 walks change, all 1-19% shorter and all through Stormwind or Ironforge; nothing else changes
  • An AI agent wrote or helped with this — Model and tool: Claude Opus (Anthropic), omp

… Wizard's Sanctum drops

- joins.json pairs may carry "kind": "step" (a straight walk both ways, at most 12 yd, no steeper than the walk slope, where neither map walks across) or "drop" (a one-way fall from a ledge's edge to where players land, kept even where it hurts and his maps make another way).
- Orgrimmar: the drop off the wind rider tower's south edge into the Valley of Strength (29 yd; 121 recorded players take it). Stormwind: the 15 yd drop below High Sorcerer Andromath's floor in the Wizard's Sanctum (40 of 41 recorded players).
- Test: the walk from Doras to Kaja goes off the tower's edge and is under 160 yd (the ramp's walk is 195).
- Walking maps repacked in the next commit.
- forge nav pack --leffe (grid 1.60.1.70532, date 2026-11-16): one part changes, Polestar_Nav_Kalimdor/Nav1_30.lua, the tower's cluster (one arc).
- The Wizard's Sanctum drop is taken out again: the walking maps already make that fall (the pack logged "his maps make it already"), so it added nothing.
…mwind and Orgrimmar

- 29 joins from the players' recordings: places where recorded players (5 to 65 distinct ones each) walk across a gap neither our grid nor the walking maps join (a threshold, a counter's gap, a platform's steps), found where their walks jump across our grid and checked against the shipped maps; both sides on floors 3+ players walk. 27 are a few yards long and drawn straight (kind step); two are routed on our grid (one too long for a step, one too steep).
- Ironforge: Daryl Riknussun's platform (the cooking trainer, #2860) is joined to the Great Forge's floor; it was walked to nowhere at his height.
- 12 of the 29 add nothing to the shipped maps (they already make the way); they stay so a later pack whose maps lack it gets it.
- Walking maps repacked (forge nav pack --leffe, grid 1.60.1.70532, date 2026-11-16): 16 Eastern Kingdoms parts change, Kalimdor none.
- Tests: the walk to Daryl Riknussun at his height from the reporter's spot (found, ends on his platform, under 450 yd) and back from the forge; two of the busiest steps walked both ways at under 1.5x the straight line plus 20 yd.
…rted goal and Afadra Dunwall's spot

- A strict join's way gets an arc (or an edge across a cluster border) wherever the walking maps lack a move of their own, even where they walk round to it within 40 yd: that way round can start a long detour. Step joins are always strict; joins.json pairs may say "strict": true.
- Walking maps repacked: three Eastern Kingdoms parts change.
- #2860's test walks to the report's goal point (60.25, 36.79) at Daryl Riknussun's height, which the unchanged maps can't reach (his spawn's point they reach by a 356 yd walk).
- The second test now walks from Afadra Dunwall's spot (Ironforge 33.92, 49.31), cut off in the unchanged maps, to Thurgrum Deepforge and back.
@github-actions

Copy link
Copy Markdown

Other pull requests for the same issue

Please compare with them before going further. If one already covers the issue, close this one or offer the part it is missing there. One of them is merged, crediting whoever helped (CONTRIBUTING.md, "Before you start").

@github-actions github-actions Bot added the overlap Another open pull request addresses the same issue label Nov 18, 2026
@fitzforge
fitzforge marked this pull request as ready for review November 18, 2026 03:30
…e city

The step at 40.83, 49.01 (z 51, 5 recorded players) joined the ground the city was
built over, which lies under its streets and is joined to the world only far away,
to the canal side: a walk from that buried ground found a 620 yd way out instead of
none (nav_maps' walk from ground joined only far away). The other Stormwind steps stay;
maps repacked (Polestar_Nav_EK parts 11 and 12).

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

overlap Another open pull request addresses the same issue

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant