Visibility model

The offline pipeline this document describes (scripts/, the review data under scripts/data/, the audit and bench harnesses under tool/, and the per-map acceptance reports) was moved out of the repository on 2026-09-19. It lives in the icarus-vision-pipeline archive, copied from commit f46a4a4 of fix/all-map-vision, where every file keeps its history. The repository keeps only what the app and CI need: the bundled assets/maps/*_svg_height_*.json.gz models, tool/check_bundled_wall_heights.dart and this document.

Before declaring a visibility change complete, apply the acceptance contract (docs/vision-acceptance-contract.md in the archive). It defines source accounting, independent expectations, and the evidence required for completion.

Authority

The SVG artwork defines wall positions, endpoints, and thickness in the map plane. Extract those from the actual paths, transforms, stroke widths, caps, and joins. Some strokes are already expanded into filled paths; use their painted footprint. Cones meet the observer-facing edge of that footprint, with no arbitrary gap or global thickness adjustment. Preserve the artwork, map size, and saved positions.

The extracted 3D map supplies support elevations and vertical information for those SVG walls: solid height intervals, low cover, and openings. Associate that information with SVG geometry offline. Ambiguous associations stay explicit and reviewable. Exported mesh edges are not additional runtime walls.

An opening must provide a usable gameplay sightline before it makes an SVG wall see-through. A gap in an extracted asset alone is insufficient. Ignore cosmetic, inaccessible, or otherwise nonfunctional gaps and retain the drawn wall as a tactical blocker. Record gameplay review separately from measured source heights; an explicit gameplay decision takes precedence over an asset-only inference.

Some openings connect different floor heights. A horizontal slice through the observer's eye cannot establish visibility through such an opening. Version 3 data may name sightlineFloorSupportIds for reviewed, horizontal destination floors. Inside those exact support footprints, cast visibility to a standing target on the destination floor. Project the existing SVG wall volumes onto that target-eye plane, retaining their solid bases, frames, and overhead caps. The destination floor does not create additional runtime walls.

Fracture's September 14 corridor follow-up restores the opening in p13-stroke-3, reflected as p13-stroke-4. The user identified the free cone beside the corridor. Complete source rays from its 8.91172 m floor to the 5.5 m A-side floor clear the lower wall and overhead structure. Subtract only the measured opening from each existing wall band, preserving the varying upper heights and conservative end frames. The nearby A tunnel perimeter remains solid. The decision is recorded in scripts/data/fracture-reported-walls-2026-09-14.json.

The September 15 follow-up separates Fracture's bare A-platform endpoint from its taller barrier. Use the complete platform's conservative 9.0677929 m cap only at the reviewed endpoint, including the defense artwork's duplicate ink. Keep the column base and adjoining platform barriers solid. The exact review is scripts/data/fracture-platform-end-review-2026-09-15.json.

The same follow-up corrects Bind's balcony frontage, Haven's low shrine and separate framed window, Summit's tower frontage, and Pearl's Mid-to-B step. Use the complete local assembly and both sides' actual artwork boundaries. The decisions are recorded in scripts/data/remaining-bind-source-ownership-review-2026-09-15.json, scripts/data/remaining-haven-source-ownership-review-2026-09-15.json, scripts/data/remaining-summit-source-ownership-review-2026-09-15.json, and scripts/data/pearl-mid-step-review-2026-09-15.json.

Lotus's small crate and stepped crate require separate local tier heights. Keep the shared higher-crate edge; cap only the complete lower tier and small crate outline. The stepped crate has no authored internal tier wall. Rendered standing-target expectations therefore use the measured lower courtyard and upper tier destination floors, preserving downward obstruction by the base. Do not treat a horizontal target buried inside the upper crate body as a standing player. The exact reviews are scripts/data/lotus-small-crate-outline-review-2026-09-15.json, scripts/data/lotus-stepped-crate-review-2026-09-15.json, and scripts/data/lotus-lower-courtyard-projection-review-2026-09-15.json.

Haven's September 13 gameplay review supersedes its Mid Window floor projection. The window is now a single tactical opening below its header, with its solid end jamb retained. The sill remains a standing surface but does not block the cone. No bundled Haven destination requests floor projection. This deliberately omits sill detail to keep the view continuous as the observer moves. The user explicitly requested this simplification after the projected floor produced a detached patch. See scripts/data/haven-tactical-sightlines-2026-09-13.json.

The SVG floor fill limits where visibility is displayed. Its perimeter is not automatically an opaque wall. Height and opening information decides which authored boundaries block a sightline. Ignore decorative foliage absent from the SVG; retain represented structural obstacles such as fences and grates.

Classify each painted element before assigning a height. Zipline symbols and ramp/floor markings remain visible artwork with no blocking interval. For example, Icebox's A Site dashed zipline and the cross-hall Tube ramp marks are annotations. Stroke color, opacity, or a nearby mesh alone cannot identify a wall.

Verify the source object's role and local extent before assigning a wall height. A nearby monitor, decal, or snow cap does not describe a complete ramp or building. For stacked structures, test a named gameplay opening from its upper floor and test the solid base separately from below. Compare against gameplay images or Riot's documented map changes; agreement between two implementations of the same height assignment cannot establish that assignment is correct.

Elevation

Keep source standing-plane precision through capsule clearance. The clearance contact tolerance is 1 mm; rounding plane coefficients to four decimals can move a slope inside its own collider even when the height error is below the runtime comparison tolerance. Pearl's September 18 Mid slope lost its whole standing surface this way. The measured restoration is recorded in scripts/data/pearl-mid-slope-standing-2026-09-18.json and compiled by scripts/apply_pearl_mid_slope_standing.py. It adds the recovered floor without changing wall footprints or bands.

Standing eye height is the selected support surface plus standing camera height. On a box, use the box top as support; its sides are below that viewpoint and do not block sight as though the agent were standing beside it. By default choose the highest locally applicable surface on which a player can stand. Abilities can provide access: ordinary walking reachability, navigation connectivity, a named callout, and a recorded jump route are not eligibility requirements. Include isolated platforms, props, ledges and roofs when their local physical surface supports the standing player and clears the player capsule. Apply actual unwalkable, player-blocking and kill-volume rules. Interiors retain their usable floors, and explicit lower-floor selections remain available at stacked passages.

The September 14 covered-interior review gives the represented corridor or room priority at the reviewed exterior roof and ceiling projections. This is a recorded default-placement decision, not a finding that the roof lacks physical collision. Keep those supports and their heights, set only automaticStandingAllowed: false, and retain explicit saved roof selections. The exact domains and assembly evidence are in scripts/data/covered-interior-gameplay-review-2026-09-14.json. Offline source records move from domains to manualDomains without changing their measured geometry. This exception does not exclude unrelated ability-accessible props, ledges or platforms.

The collision-body addendum applies the same default rule to the reviewed covered interiors in Abyss, Corrode, and Summit. A collision volume can describe an exterior ceiling just as a mesh can. Its physical upper level remains selectable; only automatic placement prefers the represented room below. The exact domains are recorded separately in scripts/data/collision-covered-interior-review-2026-09-14.json.

Split's September 15 source completion replaces its legacy standing data with measured version 3 floors. All 8,335 scene objects have a resolved collision disposition, including 24 analytic capsules. Keep the 849 physical standing domains, with 206 exact exterior roof, roof-fixture and overhead-boundary domains reserved for explicit selection under the represented-interior rule. The independent crane and construction-panel standing surfaces remain automatic. The exact default review is scripts/data/split-covered-interior-review-2026-09-15.json. The source manifest and release certificate now require an independent source for Split as they do for every other map. No legacy source exemption remains.

Floor projection subtracts each projected wall face separately. This has the same result as subtracting their union, while avoiding overlapping thin path contours that can make Skia fail beside a wall endpoint. Keep the sill, header, and hole checks when changing this calculation.

Interpolated ground can pass through a raised wall at a stacked passage. If its standing eye is inside an active SVG wall, it cannot take priority over a verified physical floor below it. Choose the highest eligible surface whose standing eye clears the local walls. Preserve explicit saved elevations, including ground.

Reference interpolation also cannot outrank a verified physical floor when both eyes clear the SVG walls. Version 3 height data records measured ground through ground.standingTriangles, an explicit list of triangle indices. Only the first covering ground triangle supplies the local ground height; a physical triangle hidden behind an earlier reference triangle cannot certify that reference. Unmarked triangles still supply reference heights and saved-ground matching. Version 2 data keeps its existing selection behavior until its ground has been measured and the version 3 data passes the acceptance checks.

For version 3 data, resolve a saved eye to the closest local level, including ground, within the 2 cm source-height tolerance. This accommodates corrected collision heights and rounded source planes without jumping from a saved lower passage to the roof. Exact nearby levels remain separate choices. Equally close different levels remain unavailable; the UI reports the automatic fallback. The saved numeric value is not rewritten. Earlier data keeps its existing matching behavior.

Geometric height confidence does not establish gameplay eligibility. A bundled support must have automaticStandingAllowed: true before the default can select it. Record the local physical collision surface and standing clearance in the offline source decisions; navigation and gameplay references can corroborate those measurements. Evaluate physical standing tops independently of navigation. Previously unnamed or disconnected surfaces must receive this same physical test, rather than remain excluded for lacking a gameplay role. A rendered horizontal mesh face alone does not establish a physical standing floor.

A collision body wholly enclosed by resolved player collision adds no standing surface. For curved shapes, prove containment of the complete boundary using the actual scaled shape. Distance to the enclosing body's triangles alone does not establish clearance inside its solid interior. Excluding distant collision requires its complete bounds, including the player radius, to miss the declared region.

Dara's gameplay review can establish eligibility when extracted collision metadata is incomplete or its clearance prediction disagrees with the observed position. Keep that review separate from the measured floor height and the original collision evidence. The September 8 review is recorded in scripts/data/gameplay-standing-review-2026-09-08.json, with exact sample identities and source fingerprints. It covers 454 positions across seven maps. On Bind's fountain, exclude only the narrow inner ring; retain the center and the outer basin. Bake the measured ring footprint, not a filled center exclusion. Use a matched player-collision floor where available; otherwise retain the reviewed local source face as the geometric height reference. This review does not authorize unrelated faces elsewhere on the same source object.

Ascent's later review excludes sixty exact standing domains on boundary-volume caps, bell-tower ledges and the Tree-room upper trim. The canonical decision is scripts/data/ascent-playable-space-review.json. Preserve the collision bodies and all other standing domains. The diagnostic 13 m threshold is not a gameplay height limit; apply only the recorded domain footprints and source faces.

Breeze's ceiling review excludes exactly volume-21-0, the upper face at 51 m shown in the review viewer. Dara confirmed it is outside playable space even allowing agent abilities. The canonical decision is scripts/data/breeze-playable-space-review.json. Retain the player-blocking body from 19 m to 51 m and every other measured or reviewed standing domain.

Fracture's ceiling review excludes exactly volume-507-0, the upper face at 74.5399 m shown in the review viewer. Dara confirmed it is outside playable space even allowing agent abilities. The canonical decision is scripts/data/fracture-playable-space-review.json. Retain its player-blocking body from about 20.43 m to 74.54 m and every other standing domain.

Pearl's review excludes four exact upper faces recorded in scripts/data/pearl-playable-space-review.json. Its map-wide collision body spans 28.4976–48.6261 m, matching Dara's approximate 29–49 m approval condition. Retain all four collision bodies and all other measured standing domains.

Abyss's review excludes 284 exact standing domains: 15 collision-volume tops and 269 scenery/cliff domains across ten scene meshes. Dara approved both groups. Apply scripts/data/abyss-playable-space-review.json without removing their collision bodies or other surfaces on those objects.

The September 12 screenshot review additionally excludes Abyss's exact upper faces volume-132-0, volume-141-0, and volume-142-0, and Haven's volume-34-0 and volume-35-0. These collision-volume tops were selecting eyes above the reported Mid, ramp, and tower interiors. The agent matched the user's gameplay reports to these source faces; this is separate from the earlier viewer approvals. The canonical record is scripts/data/reported-sightlines-2026-09-12.json. Retain the collision bodies and the usable floors beneath them. Corrected openings must also restore any measured ledge or sill standing area previously clipped away by an incorrect opaque SVG wall.

Corrode's review excludes only volume-1026-0, the upper face at 43.3816 m on BP_BlockingVolume692. Dara confirmed it is outside play. Apply scripts/data/corrode-playable-space-review.json and retain the collision body. These reviews define no global height cutoff.

Summit's review excludes exactly volume-0-0 at 65 m and volume-328-0 at 42.75 m. Dara confirmed both upper faces are outside play even with abilities. Apply scripts/data/summit-playable-space-review.json; retain both collision bodies and every other measured standing domain.

Sunset's review excludes 190 exact standing domains on fourteen defender-spawn scenery objects, including sinkhole pillars, cliffs and pipe sections. Dara confirmed all highlighted domains are outside play even with abilities. Apply scripts/data/sunset-playable-space-review.json; retain all collision bodies and every other measured domain. The review records the exact match between the completed scenery measurements shown to Dara and the final whole-map source.

At stacked passages, verify a continuous walk on each level against its named source floor. A ground interpolation that switches between an upper hallway and the floor beneath it needs separate surfaces. Confirm a usable standing domain after clipping each support against the SVG and higher walls; a tiny residual polygon or a horizontal face on a wall is insufficient evidence of a platform. When one source object contains several levels, record the selected floor faces and audit those faces; an overhead beam is not the height of the floor beneath it.

Only emitted, first-covering physical ground can replace an explicit source level. Planned polygons and triangles hidden behind earlier ground cannot supply that level at runtime. Keep source choices along the two overlay-grid steps affected by projection and partitioning rounding; this reduces coverage credit without expanding the measured standing domain.

A lower box's exposed mesh rim is not automatically a standing position. Check whether a box resting on it leaves clearance for the extracted player capsule. Record the covering object and capsule source when rejecting a covered level. Do not remove a platform solely because its area is small. Do not make the SVG perimeter transparent to recover sightlines that exist only outside the drawn map.

Treat connected ramps and slopes as continuous tactical ground. Changes in floor height do not create extra wall edges or hide the connected downhill floor by themselves. Retain real intervening obstacles and meaningful separate levels. Crouching and changing door states are outside the current pass.

Implementation and acceptance

Bake compact SVG wall geometry, vertical intervals, and support information offline. Runtime visibility uses those records, rather than the extracted or warped 3D triangle scene. Earlier mesh-normalization candidates are reference experiments, not the implementation to extend.

First deliver Split for Dara's feedback before extending this model to other maps. Personally inspect the actual SVG and cone renders at normal size and enlarged wall edges. Check solid wall contact, corners with different stroke widths, preserved openings, low cover, standing on a box, and a connected ramp. Report unknown source associations, data size, and measured query cost. A passing synthetic test is not evidence that an unreviewed game sightline is accurate.

Validate the segments between cone vertices, not only the sampled ray endpoints. Two overlapping painted wall strokes can create a visibility corner at their intersection. A wall can also cross the cone's range circle while both endpoints lie outside it. Preserve these events or the polygon can cut diagonally short of correctly placed SVG ink. Neither event adds a new wall or changes its height.

Use scripts/audit_svg_cone_boundaries.py with the production polygon export in tool/export_svg_boundary_sweep_test.dart to check these contacts independently. The audit separates range-arc chords and radial corner shadows from wall contacts. Also inspect production painter crops; matching geometry does not alone verify pixel coverage, the selected elevation, or the gameplay meaning of an opening.

The four saved Icebox poses in review 1788823493671-212d5f8 are regression fixtures in test/svg_automatic_standing_test.dart. Their source measurements come from scripts/review_icebox_user_sightlines.py. Test the windows from above and their bases from below, the connector doorway and its solid jamb, and the local lower pipe step separately from the higher pipe on the same assembly. Resolve missing-floor findings before calling a gameplay-height review complete. An infinite structural-outline fallback is an assumption, not a reviewed height.

Haven's same follow-up excludes the six connected tower ceiling-volume tops, volume-32-0 through volume-37-0. The earlier review excluded only two and missed a narrow 45 m strip above the 9 m upper room. Retain all six collision bodies and the real upper and lower floors. The source classification must separate an overhead collision box from a usable standing platform before the highest-floor selection runs. A clear top face alone can still be an invisible level boundary rather than a playable surface.

The September 13 Breeze review excludes the exact upper standing face volume-583-0 on SuperGrid_Box343. The collision box spans 8 to 16 m above the 4 m passage floor; its clear top is not that passage's gameplay standing surface. The nearby box at the second reported position retains its real 9 m top. scripts/data/breeze-gameplay-sightlines-2026-09-13.json also records three continuous facade regions. Their complete structural assemblies supply heights; isolated trim and cistern tops do not create openings in those SVG outlines. Source hashes are pinned so changed geometry requires renewed review.

Run the gameplay regressions before packaging. The desktop and Store release scripts run the reported sightline, Haven movement, and Breeze gameplay tests against the actual bundled assets. Their release flag ignores diagnostic model folder overrides. Source-height agreement alone cannot replace these behavioral checks. Preserve a before-fix failure and test deliberate bad standing and wall assignments on isolated data copies when adding another regression.

Breeze's covered-opening follow-up excludes both connected Courtyard ceiling tops, volumes 387 and 388, while retaining the 4 m room floor. Its Mid Nest room retains the 9 m floor. Measure complete window sill/header assemblies; nearest-source height cells must not turn a roof above an opening into a solid wall. A single SVG parent can contain both a window and exterior box outlines, so classify those portions separately. The canonical decisions are in scripts/data/breeze-covered-openings-2026-09-13.json.

Check both end frames and the entire usable opening from moving origins on both sides. Numerical fragments can block runtime rays despite negligible polygon area. Classification clips use the 1e-8 source-overlay grid and cannot leave old-height fragments at reflected boundaries. The roof diagnostic in scripts/audit_breeze_overhead_candidates.py flags suspended collision tops above lower floors for review; it does not make roofs or isolated ledges ineligible automatically.

Run the cross-map detectors after fixing a recurring source-association problem. scripts/audit_all_map_overhead_candidates.py inventories suspended tops; scripts/audit_all_map_opening_assignments.py checks measured gaps against both wall assets and nearby retained standing floors. Their candidates require role review. Missing navigation, collision channels, and source gaps alone do not establish gameplay behavior. Keep weak source overlaps separate from positions inside the receiver with a wall-clear standing eye.

scripts/audit_all_map_wall_residue.py checks every wall footprint and samples mirrored vertical profiles. The compiler removes whole wall records that collapse on the 1e-8 source-overlay grid. It preserves every remaining record. Before packaging modified map assets, run svg_wall_footprint_integrity.py from the archive to certify all 26 sides, then update the checksums in test/bundled_map_models_test.dart; the certificate lives with the archive.

The focused September 13 review excludes Abyss standing domains volume-145-0, volume-146-0, and volume-1603-0, and Haven volume-215-0. Complete neighboring assemblies identify these as overhead collision or invisible boundary tops. Retain their collision bodies and every other measured floor. Both physical and measured support aliases must receive the same exclusion. The canonical record is scripts/data/systematic-roof-review-2026-09-13.json. Breeze bridge parapets 334/335 and Icebox connector railing 496 remain eligible.

Pearl's lower B Hall tunnel retains its 2.5 m floor beneath a continuous 6.175 m ceiling. Use the complete tunnel shell, including its end frame, from scripts/data/pearl-lower-tunnel-review-2026-09-13.json. Classify each side's actual painted mouth separately: the authored reflection differs by 0.0003 SVG units and otherwise leaves a thin old-height blocker. Restore only source standing area released by the corrected wall intervals.

The twenty focused opening families have no supported opening changes. Their decisions are recorded in scripts/data/systematic-opening-review-2026-09-13.json. In particular, Icebox ramp parents p7-stroke-10 and p7-stroke-18 have front railings above their facade tops. A cap measured from the ramp skin alone removes real represented structure. Preserve the railings as tactical blockers rather than treating spaces between their bars as windows.

test/systematic_map_gameplay_test.dart checks these reviewed roof decisions, retained physical parapets and railings, and Pearl's moving lower tunnel sightline with overhead and end-frame controls on both sides. Both packaging scripts include these checks against the exact bundled assets.

Wall bands re-derived by ray probing (2026-09-17)

The bundled wall bands were re-derived from the extracted 3D scene by measuring the quantity a band encodes: the eye heights at which a horizontal sightline crossing the painted stroke is blocked. scripts/derive_wall_bands_by_rays.py probes each stroke at stations along its centreline with a 1 m horizontal segment across the ink at every 0.1 m up to 40 m, against opaque and masked render geometry with decor excluded. A height blocks when 60% of stations are stopped; runs form bands relative to the wall's floor. Runs set back more than 0.3 m behind the wall's own face are dropped as neighbouring structure. A stroke with fewer than six faces in its corridor keeps its previous bands.

Three passes then decide what may replace the reviewed data: scripts/protect_reviewed_walls.py keeps every wall named in a recorded review, fixture or partition (both sides, mirrored through the alignment); scripts/smooth_wall_band_neighbours.py rejects a lowered piece whose touching neighbours on the same stroke stayed high, so no notch is cut into a solid wall; scripts/cap_unsupported_raises.py rejects a raised top that the narrow-footprint reading from scripts/audit_svg_wall_heights_vs_world.py cannot support. Split joined this pipeline on 2026-09-19; see below.

Acceptance was the per-map gameplay suite (ICARUS_VERIFY_BUNDLED_GAMEPLAY), now in the archive with the pipeline. The repository pins the resulting models by checksum in test/bundled_map_models_test.dart.

A wall that blocks every probed eye height up to the 40 m ceiling is recorded with an open top (null) on its last band. That is a measurement, not the old reviewed-label fallback, so tool/check_bundled_wall_heights.dart accepts an open top there and nowhere else; unknown walls, non-finite floors and unbounded lower edges still block the release.

Split brought onto the piece model (2026-09-19)

Split was the prototype. Its wall layer stayed at 69 records, one per painted run with a single hand-assigned band, while the other twelve maps were compiled from reviewed decisions into pieces about a metre long with bands measured from the 3D scene. Every rule on this branch works piece by piece, so Split was skipped by all of them.

Rather than author the decisions review Split never had, scripts/partition_split_walls.py cuts each record into pieces about two SVG units long along its medial line (<parent>-local-<n>), keeping the parent's floor and bands so the cut alone changes nothing (the ink union is asserted unchanged). Records whose names carry a reviewed prop cut (-low-, -counter-, -planter-, vent…-opening, …) stay whole and are protected. The pieces then go through the same passes as the other maps: ray derivation, reviewed-wall protection, the perimeter seal (its Split exemption removed), the narrow-footprint audit and raise cap, the anomaly rules, box outlines and notch closing. The neighbour-smoothing pass is skipped for Split on purpose: its baseline bands are upper bounds rather than reviews, and smoothing would restore three quarters of the measured lowerings.

Stacked areas and tunnels (2026-09-19)

Dara's rule for stacked areas: the top layer is the default, and an agent placed on a lower layer sees that layer's view. The data already carries this: the painted ground is the lower floor, the upper floor is a standing surface that automatic standing prefers, and explicit elevation selection reaches the lower one. Rays never consult the ground; the layers separate by height bands alone.

scripts/apply_reviewed_openings.py with scripts/data/reviewed-openings-2026-09-19.json records spots Dara ruled see-through where a protected record kept the ray derivation out: Pearl's B Hall tunnel west mouth is a passage under a 7 m header. Walls at least half inside the region take the derived bands; the sealed map edge beside a mouth stays sealed.

Two further passes were built and withdrawn the same day, and stay in the archive's scripts/ unapplied. remeasure_above_openings.py replaced the tunnel review's blanket above the ceiling with what the horizontal probe found, which was nothing until 11 m; Dara confirmed the ramp into that tunnel is walled on both sides, so the probe was missing a set-back wall and the blanket was right. seal_void_walls.py only existed to stop the leak that change caused. A reading of "nothing above the ceiling" on a covered passage is not evidence without Dara.

Drag performance on Windows (2026-09-19)

Dara's bar: dragging an agent with a cone must feel instant on Windows. Measured before the change, with the native query at the app's usual range (about 140 SVG units): 1.44 ms in the native query, about 0.3 ms of Dart around it, and a profile-build frame build time of 1.9 ms at the 90th percentile. Windows frame timings report raster time as zero, so raster was reasoned about rather than measured.

Changes, all exact (the polygon is bitwise the same):

• native/height/icarus_svg_height.cpp owns a small persistent thread pool. The arc rays, event generation and the event rays run in chunks across it; the polygon is assembled serially afterwards in the original order. Workers spin briefly between the runs of one query and sleep between frames. ICARUS_HEIGHT_THREADS overrides the worker count for diagnosis.
• Event generation culls vertices outside the aperture before any trig.
• SvgHeightVisibility caches the wall activity mask per eye height.
• A native result keeps its packed doubles; the cone outline path is built from them directly, and the polygon list is only materialised on demand (reports, tests).
• The drag preview no longer sits in an Opacity widget. The cone paints at the preview alpha and only the small agent icon takes an opacity layer, so the engine does not composite the whole preview offscreen every frame.
• The receiver clip path is rotated and scaled once per drag and translated on the canvas, so the clip path object is stable across frames.

Instruments: tool/svg_height_drag_bench_test.dart and tool/svg_height_native_phase_bench_test.dart (set ICARUS_SVG_NATIVE_LIBRARY to the built icarus_height.dll), integration_test/view_cone_drag_performance_test.dart and view_cone_drag_timeline_test.dart under flutter drive --profile -d windows.