Block cones on VALORANT's minimap lines, drawn on our walls

[SunkenInTime]

Stacked on #240. Merge after #241 as planned; the perf numbers below were measured with it. This PR no longer needs #241's runtimeWalls loader: the walls themselves are now the merged outlines.

What changes. Walls no longer carry heights measured from the 3D map. Every wall is our drawn wall art, and it takes its heights from VALORANT's own minimap vision lines, the 2D line set the game uses for its minimap cones. Riot's lines decide how tall a wall is, and our art decides where it is. So cones stop exactly on the drawn walls, with no gap.

• The rule. A viewer whose capsule centre (floor + 0.98 m) is in [th_k, th_k+1) is blocked by layer k's lines. The thresholds are absolute heights in Riot's data. The 0.98 m comes from a scan of where the lines' obstacles top out, which peaks sharply there. The assets don't state it.
• Walls.
  • Each edge of the wall art takes the layers of every Riot line running alongside it (within 2 SVG units).
  • Riot lines with no art beside them (glass, railings, crates we don't draw) stay as thin walls, joined to the nearest art where they hand over.
  • One line was added where Riot has none: the Lotus defense platform wall from your round 2 replay.
  • One spot has Riot's lines taken off: the Haven C Garage window, kept open from the garage floor as you ruled (2026-09-19, again 2026-10-04).
  • Touching strips with the same heights are merged into one outline each, and those outlines are the walls.
• Floor. Cones are drawn on the floor minus the wall art.

Checked against the 3D map (240 standing poses per map side). Only the part of each ray that lands on floor the app draws counts, because rays leaving the map are never drawn:

visible floor, both sides	measured (before)	this PR
leaked length	1,747,007	2,313,132 (+32%)
false-shadow length	32,714,287	33,816,279 (+3.4%)

• Better: Pearl (−88% leaked), Summit (−90%), Breeze (−65%), Ascent (−63%), Corrode (−52%) and Haven (−43%).
• Worse: Abyss, Fracture, Icebox and Sunset leak about twice as much or more, and Split and Bind about a quarter more.

The check is stricter than the game. Riot's minimap ignores props the 3D map has, like crates, poles and low walls. On Split's six worst spots, five cones look normal for exactly that reason.

Perf (all-maps drag, profile build, with #241): frame build is at most 3.35 ms at p99 on every side over two runs, under the 6.06 ms budget for 165 fps, and no frame's build goes over budget (13 frames did before this fix, 6 on #241's own run). Raster still has 2–7 single slow frames (10–27 ms) per run, on a different side each time; they were there before this PR too and don't depend on the wall data.

The first perf pass on the final models found a regression the drag test had partly hidden. Each frame of a drag steps the agent out of any wall it stands in (standablePointNear). With thousands of thin strips as walls, and a floor with every wall cut out of it, that step took up to 1.6 s on Lotus in a profile build (old models: 1.7 ms). Two fixes:

• The walls are the merged outlines rather than strips (builder).
• _Footprint containment files each edge under the 1-unit rows its height reaches and reads only the point's row. The answer is unchanged; test/svg_footprint_contains_test.dart checks it against reading every edge, and fails if the filing is broken.

Over a grid of every 1.37 units on Lotus, Breeze and Summit, placement now takes at most 2.0–5.6 ms in a profile build, against 1.2–6.7 ms for the old models with the same code.

Short stretches only. An art edge with no Riot line alongside borrows the nearest line only when it's a short stretch, at most 4 units, between edges that have a line: a bevel, a jog, a wall's end. Before this, a drawn ring by Split's B wall took the wall's line and cut cones beside it into slivers. Where several lines are equally near (Riot often stacks a ground line and an upper line on one spot), the stretch takes all of them; picking one by list order flipped a Haven corner when the garage lines were taken off.

Tried and dropped: making everything that is neither floor nor wall a wall. It blocked real sightlines across drops on Abyss and Icebox (+54% and +34% false shadow on visible floor) and changed nothing visible on Pearl.

Review. Astra found four issues and all are fixed: holes in walls were treated as solid, there was a gap where a Riot line hands over to the art, band tops were closed while Riot's layers are open, and this PR depended on #241's loader (it no longer does).

Tests. test/svg_riot_vision_test.dart:

• every band edge is a Riot layer edge;
• a viewer exactly on a threshold gets the layer above;
• a wall's hole stays floor;
• the Lotus platform wall blocks;
• the Bind container blocks (your call: follow Riot);
• Breeze Mid's slanted roof blocks (your call: follow Riot);
• a drawn ring by a wall doesn't slice a cone;
• the Haven garage window stays open from the garage floor.

test/svg_footprint_contains_test.dart: containment matches reading every edge, on Lotus's walls and floor.

Checksums are updated, all 161 svg/vision/cone/map tests pass, and tool/check_bundled_wall_heights.dart passes.

The builder, Riot's decoded tables and the line review tool are in the vision pipeline archive (scripts/riot, tools/vision-lines). Rebuilding from commit 0497bec reproduces the shipped models byte for byte. The method is in docs/vision-model.md, under "Blocking from VALORANT's minimap lines".

Your calls (2026-10-04). Abyss, Fracture, Icebox and Sunset look right in the dev build. The Haven garage window stays open. The Bind container and Breeze Mid's slanted roof follow Riot's lines.

🤖 Generated with Claude Code

Confidence Score: 4/5

Do not merge until the outstanding Breeze Mid sightline issue is resolved. The lower-floor projection concern is non-blocking.

Findings

1.  Breeze opening becomes opaque ▶
2.  Destination floor projection disappears ▶

Summary

This PR replaces measured wall heights with VALORANT minimap-line layers and updates the bundled models and visibility checks. The Haven garage window and Lotus platform test are fixed. Two earlier visibility concerns remain: the Breeze Mid opening is blocked from the perch, and lower-floor views no longer use destination-height projection. The Breeze sightline issue must be resolved before merging.

_{Reviews (13) · Last reviewed commit: "Keep placing an agent fast on the Riot-l..."}

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

• Configuration used: Organization UI
• Review profile: CHILL
• Plan: Advanced
• Run ID: e5090d51-808e-4b94-ac5d-c739d0c0fb00

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

• Autopilot · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[greptile-apps]

Breeze opening becomes opaque

An agent on the Breeze Mid perch could previously see the crate across this opening, which the test notes is clear in the 3D scene. The rebuilt model excludes the crate from the visible area, and the new assertion treats that loss as correct. Players therefore see a false shadow through the opening; preserve the sightline rather than accepting the blocked result.

Artifacts

Breeze Mid cone probe source

• The authored Flutter test loads the bundled Breeze attack model and exercises the actual perch-to-crate cone on both revisions.

Base and HEAD comparison command

• The authored script runs the same probe against an isolated base worktree and current HEAD using `--no-pub`.

Base revision cone result

• The executed probe at bf8f0c6 reports `receiverContains=true` and `seesCrate=true`, confirming the target was visible before the change.

Current HEAD cone result

• The identical probe at HEAD reports `receiverContains=false` and `seesCrate=false`, confirming the target is now blocked.

Current Breeze Mid assertion result

• The focused PR test passes on HEAD with its new blocked expectation, showing the test accepts the changed sightline.

_{Ran code and verified through T-Rex}

[greptile-apps]

Destination floor projection disappears

Dropping sightlineFloorSupportIds removes the runtime’s target-height projection for cross-height views. From a raised Lotus defense floor, the previous projected cone excludes a point on the lower destination floor, but the new horizontal fallback paints it visible. This changes the displayed lower-floor area without checking visibility at that floor’s height. Retain the projection or provide an equivalent target-height-aware path.

Artifacts

Flutter test for a raised Lotus observer and lower destination floor

• The authored test loads each ref's bundled asset and calls the runtime cone used by the painter, enabling the two runs to be compared.

Base run with destination-floor projection

• The Flutter test loaded the base Lotus asset and observed a projected path that excluded the lower-floor target.

HEAD run without destination-floor projection

• The same Flutter test loaded the HEAD asset and observed a null projected path and a horizontal fallback that painted the target.

_{Ran code and verified through T-Rex}