Quake II north window: continuous hull correction, 2026-11-08

The old x≈112 northern aperture route is false for the original standing player hull. A continuous trace stops at the sill, and the proposed elevated crossing grazes world brush 1603. Moving left avoids that world edge but hits the original window entity *328 / model 20, omitted by the old leaf sampling. With that entity intact, the entire opening blocks both standing and crouched player boxes. This is a tested geometric correction, not an explanation of the retained native player's death.

Implementation: tools/quake2-bsp-route.py; tests: tools/quake2-bsp-route-test.py. Evidence: scratch/runs/20261108-quake2-swept-hull-route-2125/.

Original bytes, opened read-only

PAK path: test/binaries/candidates/quake-2-demo-installer/installed-extracted/Install/Data/baseq2/pak0.pak. Size 49,951,322, SHA-256 cae257182f34d3913f3d663e1d7cf865d668feda6af393d4ecf3e9e408b48d09. maps/demo1.bsp occupies offset 1,264,679, length 1,991,772; SHA-256 ee13d2e5448d990ced8ca7b0595722518c5221feab68d5a07e2a291391da6457. The loader verifies both SHA/size pins and the unique PAK directory entry before parsing IBSP38. It never extracts or copies media.

The checker parses original planes, nodes, leaves, leafbrushes, brushes, brushsides, models and entities. Each model's reachable BSP leaves determine its brush membership: 1,758 original brushes, 1,641 world brushes, 35 models. The authored solid-entity scene contains 1,721 brushes; trigger brushes remain nonblocking. Model *20 owns brushes 1678–1684, none belonging to the world model. Its seven irregular convex brushes retain all original bevel/side planes and WINDOW contents, not just SOLID leaves.

Player origin bounds are standing (-16,-16,-24)..(16,16,32) and crouched (-16,-16,-24)..(16,16,4). Thus feet are origin z−24, full width is 32, and heights are 56/28. The player mask includes SOLID, WINDOW, PLAYERCLIP and MONSTER; monsters themselves are not synthesized from entity records.

Concrete comparisons

The following coordinates are world player origins, not native input targets. Horizontal sweeps run from y=96 to y=192 unless stated otherwise.

Check Result
Standing (112,-320,24.125) to (112,96,24.125) Clear approach on the start floor
Old path, corrected ground heights, first blocked segment y=96→112 at x=112, z=24.125 Fraction 0.998046875, stops y=111.96875, sill brush 239
Elevated x=112, z=40.125, world only, native ±16 hull Fraction 0.166341145833, stops y=111.96875, brush 1603
Same world sweep using old ±15 footprint Clear: reproduces the omitted hull-width contact
Elevated x=96, z=40.125, world only Clear
Same x=96 sweep, original entities, standing Fraction 0.249674479167, stops y=119.96875, brush 1679, entity 328
Same x=96 sweep, original entities, crouched Same stop/fraction, brush 1681, entity 328
x=96 18-unit up/across/down step from z=24.125, y=96→160 World-only succeeds over the 16-unit sill; original entity scene blocks both hulls
x=96 jump probe from (96,96,24.125), velocity (0,200,270), gravity 800 Both original hulls hit entity 328 at 0.11984375 s, y=119.96875; world-only clearance probe passes 0.6 s

Brush 239 spans x=−192..128, y=128..152, z=0..16. Brush 1603 spans x=128..272, y=128..152, z=0..112. A centered ±16 hull at x=112 touches x=128. The old ±15 samples reach only x=127 and miss that contact. The 1/32 collision epsilon stops a north-moving hull before its front reaches the blocking plane; it does not shrink the hull to fit the aperture.

Entity 328 is authored func_explosive, model *20, health 10, dmg 0, target t78, no spawnflags or origin translation. Model bounds are x=−72..128, y=136..144, z=16..112. Its WINDOW collision contents are 402653186 / 0x18000002. These fields establish original brush geometry, not a current actor, an enemy, a damage cause, or authorization to shoot it.

Continuous obstruction certificate

At the opening's middle plane y=140, world geometry confines any hull crossing this particular window to origin x=−48..112. The sill top z=16 and lintel bottom z=112 constrain standing origins to z=40..80, and crouched origins to z=40..108. Boundary contact is already blocking; including the boundaries makes this a conservative superset of admissible positions.

cover-plane expands each original brush plane by the player-box support, then proves obstruction of rectangular origin cells. All four corners of a cell must lie inside one expanded convex brush. Convexity proves its whole interior blocked; uncovered/subdivision-limit cells remain unresolved. This is not a dense point grid. The intact window covers the entire standing rectangle with one cell, brush 1680. It covers the crouched rectangle with nine disjoint cells, brushes 1679, 1680 and 1681. Both certificates have zero unresolved area. Tests independently enumerate hull corners, verify each cell, check disjoint interiors, and sum the exact rectangle area.

Every continuous route through this opening must cross y=140. Consequently walking, diagonal motion, stepping or jumping cannot pass a native box through the intact authored opening. This conclusion concerns this aperture and entity state; it does not rule out other routes through Outer Base or a changed entity state. Omitting the window is retained only as a diagnostic counterfactual, never described as a qualified native route.

Original collision algorithm cross-check

No local cmodel.c, pmove.c, q_shared.h or qfiles.h was found under /home/user, /opt, /tmp. Instead, the existing original quake2.exe (beside baseq2/) was inspected statically with objdump, never executed. Size 323,584, SHA-256 2fa987c9390663e230c5b6479285e7c55511e35f80091bfed96f43062c1c810b. It has no symbol table. The inferred box/brush clip helper at 0x4121d0 matches per-plane box support expansion, entering/leaving fractions, outside early rejection and startsolid/getout logic. At 0x41237b it subtracts the double 0.03125, stored at file offset 0x40c00 / VA 0x442600; at 0x4123a2 it subtracts −0.03125 for the leaving fraction.

Inspection corrected an initial checker convention: the original moving allsolid branch at 0x4123e1..0x41240d sets the flags without overwriting fraction. The stationary test at 0x412595..0x4125b9 sets fraction zero. The final checker and tests preserve this distinction. Consumers must check startsolid/allsolid, not just fraction == 1. Tests hash-pin the original EXE and relevant inspected instructions. Retained disassembly is an independent static algorithm check, not an executed original-collision oracle. Python double arithmetic versus x87 float spills is not bit-exact numeric parity.

The independent analytical oracle enumerates all eight hull corners for each plane and intersects time half-lines, rather than calling the clipper. It agrees on 2,000 varied traces against actual original convex brushes with epsilon disabled; explicit analytical tests cover the fixed collision epsilon.

Reuse and limits

python3 -B tools/quake2-bsp-route-test.py
python3 -B tools/quake2-bsp-route.py info
python3 -B tools/quake2-bsp-route.py trace 96,96,40.125 96,192,40.125
python3 -B tools/quake2-bsp-route.py --hull crouch step 96,96,24.125 96,160,24.125
python3 -B tools/quake2-bsp-route.py jump 96,96,24.125 0,200,270 --seconds .6
python3 -B tools/quake2-bsp-route.py --hull crouch cover-plane 140 -- -48,40,112,108

path FILE checks a JSON list of origin waypoints with continuous segments. --policy world explicitly excludes entity models for comparison. Default initial retains solid brush entities at authored positions, excludes triggers and initially disabled walls/explosives, and inventories all choices. Moving doors, trains and rotating models have unknown runtime transforms; authored positions are not a reconstructed live scene. Explicit authored rotations fail closed. The relevant window has no transform or mover uncertainty.

Up/across/down is a geometric step-clearance check with the native 18-unit limit and a walkable landing normal ≥0.7; it does not implement full PM_StepSlideMove, friction, acceleration or sliding. Jump clearance sweeps continuous chords with upward inflation g*dt²/8, enclosing the intervening parabolic arc. It can conservatively reject a near-ceiling arc; it never infers free motion by sampling ballistic positions. Neither helper predicts health, damage, AI, trigger effects or survival.

Native evidence and validation

The actual retained world-first, window-slide-right, missing-section-forward, lip-angle-left and onward-right PNGs from 20261108-quake2-repaired-startup-opening-2052/remote/original/ were viewed. They show start health100, pane/gap/sill, then health54 with red tint and health0 with rolled death view. Their paths/hashes and observations are retained without copying images. Exact native origin, crossing time and damage source remain UNKNOWN. Static obstruction is compatible with the failed approach but cannot register these pixels to world coordinates or attribute death. The old claim is corrected on geometry independently of those unknowns.

18 focused tests pass, including thin-wall tunneling with free endpoints, grazing, startsolid escape/allsolid, stationary solid/free, hull width, crouch clearance, collision masks, step height/headroom negatives, jump ceiling/wall/apex negatives, original sill/window comparisons, exact aperture coverage and the independent oracle. Required flock /tmp/wine-build.lock bash tools/build.sh passed (canonical 2,184,500 bytes; compat 2,187,146 bytes). All unique logs were gzip from first stored bytes; inventory independently rereads regular-file SHA/size and decompressed original SHA/size with CRC, and records symlinks without following them. Git SHA plus small owned patches provide source recovery, not a tree archive.

Historical publication failure (preserved): publication was blocked by the owner's disk floor: /home fell to 1,499,435,008 bytes free during final evidence preparation and subsequently 1,499,049,984, below 1,500,000,000. The completed three owned source files remain uncommitted in shared main; no code commit or push is claimed. This run occupies about 104 KB, so reclaiming its files cannot restore the floor. No older run or peer file was deleted. Source patches and the final inventory are preserved; restore headroom before committing/publishing these explicit paths. The build and focused test passes above preceded this blocker.

No game, Chrome, remote, box, original retry, input, worker, public deployment, configuration/backend/binding/clock/seed change, private guest operation, old-run deletion or peer edit occurred. Native configuration and prior proof scope are preserved. Box NONE, state NOT_CREATED, expiry NULL, stop receipt N/A. Full27, parity, audio and FPS remain open.

Next work should use this checker to assess a different route or an explicitly established native change in the window's collision state. Repeating the old intact-aperture movement is not supported by these bytes. Any new native run requires separate authorization; this worker releases its files and stops.

Successor publication: the released correction is continued and published with the explicit window-state checker. The preceding disk-floor failure and original run inventory/patch remain intact; the successor tests conditional clearance after validated entity328/model20 removal, without changing this intact-window obstruction certificate or the retained death damageUNKNOWN. See the successor run for final build/publication.