[Verslabs]

Summary

Closes #35. groundedProxy only samples the block column under the hitbox centre. A player sneaking at a ledge (void edge, tower top) with the centre past the edge therefore reads as airborne while standing still. A level airborne tick falls in the Fly(FlyB) friction band every tick (dy=0, prevDeltaY=0 ⇒ dy − expectedY ≈ 0.078), so FlyB alerted within a few ticks.

Changes

• FlyEnvelopeCheck: the grounded branch now also accepts a level tick (|dy| < 0.01) with a solid block within 0.05 under any column of the 0.6-wide hitbox footprint (isEdgeSupported). It only runs when the proxy already said airborne and the player is level. groundedProxy itself is unchanged, so other checks aren't affected.
• New WorldQueries.isSolidBelowFootprint and the pure Footprint geometry helper (1, 2 or 4 distinct columns; centre first; a face that only touches a neighbouring column doesn't count; negative coordinates floor correctly).
• New legit scenario legit-sneak-edge (sneaking on a 3-tall tower top with the centre 0.2 past the edge, 40 motionless ticks, then creeping along it). It asserts expectQuiet plus Fly(FlyB) / Fly(Blink) flag counts of 0. Before the fix: 57 Fly(FlyB) flags, peakVL 19.00 (an alert). After the fix: quiet.
• FootprintTest unit tests for the column geometry.
• scripts/checks.json fpGuards entry and docs/automated-live-testing.md coverage rows (FP-direction regressions 9 → 10).

Affected areas

• Detection/check logic
• Packet observation or inference
• Config or presets
• Commands or UI
• Rendering or HUD
• Replay, clips, camera, or input
• Persistence or local data
• Documentation/tests/tooling

Verification

Commands run (branch head 622c92b):

python scripts/verify_contribution.py --static
./gradlew test
./gradlew build
python scripts/live_selftest.py --check fly
python scripts/live_selftest.py --check sneak-edge
python scripts/live_selftest.py            # full three-pass run, see baseline note

Results:

• Static verifier: PASS (only warning: no build/selftest-manifest.json before the first live run)
• Tests: PASS, BUILD SUCCESSFUL, 54 tests, 0 failures, 0 errors (includes the new FootprintTest)
• Build: PASS, BUILD SUCCESSFUL

Live client verification

• Client launch/mixin application: PASS. The real 1.21.11 client booted through runClientGameTest, and pipeline-smoke passed.
• Scenario(s): legit-sneak-edge (new), legit-fly-ramp, cheat-fly-hover-meteor, cheat-fly-ascend-liquidbounce, cheat-fly-strafehop-fusion, cheat-fly-antikick-itami, cheat-elytra-fly-liquidbounce
• Expected result: the sneak-edge stance is quiet, and every fly cheat (including hover and antikick, which the edge-support concession could in principle hide) still alerts.
• Observed result: --check fly 7/7 passed, --check sneak-edge 2/2 passed, legit false positives: none, cheat bypasses: none. In the full three-pass run, all 23 legit scenarios (including legit-sneak-edge) and all 10 replay scenarios passed (legit false positives: none). The 12 cheat-pass bypasses in that run are all among the 15 that main itself fails on this machine (see the baseline note below).
• Not tested: no manual runClient session at a real void edge on a server.

Full-suite baseline on the test machine: the full python scripts/live_selftest.py run on main (5495a7f) on this machine (macOS, 8 GB RAM) also fails: 95 scenarios, 80 passed, 0 legit false positives, 15 cheat-pass bypasses in movement checks (flyEnvelope, noFallDamage, speedEnvelope, spider, noSlow, wallSprint, sprintHack, waterWalk, phaseClip). The same cheat scenarios pass when run on their own (e.g. cheat-speed-liquidbounce, cheat-fly-ascend-liquidbounce), so this looks like a resource/timing problem with the full run on a low-memory machine, not something this PR caused. The live evidence above therefore comes from targeted runs.

Safety and privacy review

• No telemetry, uploads, analytics, or remote reporting were added.
• No unintended outgoing gameplay packets were added.
• Fail-open behavior is preserved where appropriate. (isEdgeSupported returns false, i.e. no exemption, on any error.)
• Local-world, camera, input, and persistence state restore after stop/error/world change where relevant.
• Logs, screenshots, clips, and test data contain no secrets or private information.

Documentation and migration

• User-facing documentation updated, or no update is needed.
• Config/clip migration reviewed, or no migration is needed.
• Generated documentation source data updated if applicable. (scripts/checks.json fpGuards.)

AI assistance

• AI tool(s) used: Claude Code (Claude Opus 5.5)
• What the tool contributed: wrote the change, scenario and unit tests, and ran the static verifier, Gradle test/build and the live runs above.
• What I reviewed manually: the diff, with an AI-assisted review pass on 2026-10-12. Maintainer review requested.

Known limitations and follow-up work

• Same concession as groundedProxy: a flyer holding Δy ≈ 0 directly above any block the hitbox overlaps reads as grounded. The footprint widens that area from the centre column to the full 0.6 box.
• Merge conflict with main: SelfTestEntrypoint.kt conflicts with checks: exempt creative/spectator flight from movement checks only #37 (creative/spectator scenarios). This branch needs a rebase before merge. The coverage counts in automated-live-testing.md should be rechecked after the rebase.