A single-player first-person horde defense game made with Godot 4. Survive endless waves, construct defenses between attacks, adapt when fixed world gates open, and spend Technology Points earned on defeat to unlock equipment for future runs.
Planning offers optional Operator, Engineer and Guardian contracts: deal weapon damage, deal tower damage, or clear a wave without core integrity damage. Completing a contract grants bonus scrap and a supply cache; choose one item or salvage it for scrap. Contracts have no failure penalty, scale to bounded targets, and preserve progress and unclaimed rewards in checkpoints. Finish both the upgrade draft and cache selection before starting another wave.
Loot now includes Ammo Cells, Trauma Kits and Repair Kits with distinct field-guide models. Use them from Inventory during preparation or combat. Ammo Cells restore 3–7 active-firearm magazines to reserve, Trauma Kits restore 40–100 health, and Repair Kits restore 80–200 core health across Common–Legendary rarity. Full health, full reserves and disabled repairs retain the item instead of consuming it. Campaign equipment remains fixed.
Requires Godot 4.4 or newer, Windows 10/11, and an OpenGL 3.3 compatible GPU. Development verification uses Godot 4.7.2. No .NET, Python, Node, asset downloads, or external services are needed to play.
- Windows: double-click
run.cmd, or run./run.ps1in PowerShell. The launcher usesGODOT_BIN, Godot on PATH, or the locally cached development engine. - The local Windows export is
build/GridFall.exe; keepGridFall.pckbeside it. This export runs without installing Godot. - Any supported desktop platform: open
project.godotin Godot and press F6 onscenes/main.tscn, or rungodot --path .. - Editor:
./run.ps1 -Editor. - Forward+ rendering:
./run.ps1 -ForwardPlus. On a Vulkan-capable GPU, Settings additionally exposes screen-space reflections, ambient occlusion, and volumetric lighting. Compatibility rendering remains the default.
| Action | Default |
|---|---|
| Move / look | WASD / mouse |
| Fire / aim | Left / right mouse |
| Dash / jump or wall jump / crouch, slide or slam | Shift / Space / Ctrl |
| Reload / switch weapons | R / 1, 2 or mouse wheel |
| Collect nearby loot / use supply station / open door | E |
| Pulse ability / medkit / grenade | Q / C / G |
| Restoration matrix (Ascension unlock) | X |
| Inventory / planning | Tab / B |
| Begin next wave | Enter |
| Pause / close interface | Escape |
| Performance overlay | F3 |
| Sandbox controls | F4 |
| Grapple (hold) | F |
| Punch / timed parry | V |
| Unstuck / return near deployment | U |
Movement and interaction keys can be rebound in Settings. Mouse sensitivity, field of view, recoil, crosshair, HUD text size, audio, shadows, frame limit, VSync, fullscreen, and window size can also be changed.
Movement takes inspiration from fast arena shooters: run at 12 m/s, tap Shift for a ground or air dash, and hold Ctrl while moving to slide. Dashes use three recharging charges, shown below the crosshair, and briefly avoid direct damage; damage-over-time remains dangerous. Air control preserves momentum, including slide jumps and jumps out of a dash. Space near a wall permits up to three wall jumps before landing. Tap Ctrl in the air to slam down for one stamina/dash bar, damaging nearby enemies on impact. Slams require a full bar and charge once per slam. Jump just as you land for a higher rebound. A short jump buffer and ledge grace period make timing forgiving. Movement research and terrain still affect speed. Settings can disable movement speed lines, FOV expansion, and camera tilt; slide stamina is optional. Movement state persists in paused combat checkpoints.
Deploy into one of eight worlds. Start with a primary weapon, service pistol, 300 scrap, two medkits, and two grenades. Use preparation to build on permitted tiles, buy supplies, repair, upgrade, or sell structures. Construction and all structure management stop during combat. The core, roads, fixed walls, portals, and scheduled gate tiles are protected.
Switchback Quarry adds bridged water cuts, a covered 2.8 m firing terrace with stair access, and two supply caches. Reactor Spine combines a short western approach with a winding eastern machinery gallery and covered maintenance alcoves. Both have offset cores, additional fronts opening after waves 4, 8 and 12, and support Frontier expansion.
Built barricades protect against enemy movement and hostile shots while letting friendly towers target and fire through them. This includes rockets, mortar shells, chained attacks and Interceptor targeting. Fixed world walls still obstruct tower fire; player shots remain blocked by barricades.
Expand Wave intel in the planning sidebar to inspect scheduled enemy counts, active portals and newly available threats; hover an enemy row for tactics. The forecast uses the same composition logic as the actual wave, including bosses, events and adaptive factions, without consuming gameplay randomness. Purchases can change the forecast, which refreshes when you return to planning. Summoned reinforcements are additional. Forecasting remains bounded by the director's 600-entry queue plus compressed per-type reserves, regardless of the scheduled population.
Verification matched forecasts to actual queues/reserves for normal, boss, swarm and adaptive elite waves, repeated previews, save/reload and a supply purchase. A forecast for 791,521,206 scheduled enemies took 469 microseconds with 601 queue entries and 12 reserve types; this is composition timing, not simultaneous-enemy simulation performance. Native 1280 × 800 captures checked ordinary and 56-billion-enemy count layouts. A 60-simulation-second headless combat check cleared wave 10, defeated its boss and continued wave 11 with 14 kills and no core damage; its fixed-step frame figures are synthetic. Campaign and spectator modes exclude wave forecasts.
Barricades, razor spikes and proximity mines are the road-placement exception. Ordinary structures cannot be built on roads, water, ramps, stairs or platforms, even if an authored map includes them in its construction permissions. All nine turret types can mount above unlocked fixed wood, brick, concrete, steel, reinforced, glass, barricade and permanent walls. Their aiming and projectile origins use the mount height; destroying a breakable support lowers the surviving tower to the cleared ground. Hovering a planning tile shows its terrain, objects, collision, construction and placement restrictions. Animated minimap arrows aggregate occupied spatial cells to show the horde's direction.
Bombardiers join road-bound waves from wave 10. An amber warning precedes their first arcing shell by 0.9 seconds; shells burst across 3 m and can be parried into straight reflected shots. Barrier Field research also unlocks the Interceptor Array: a 150-scrap, 500-HP turret that destroys one visible hostile projectile within 11 m every 0.65 seconds. Tower rate, support, range and specialization upgrades apply; interference slows its cooldown. It never damages enemies or intercepts friendly/parried shots. Cover blocks its firing line and concentrated volleys can overwhelm it. Run results count interceptions. Enemy blast damage now respects the height of mounted defenses.
This addition passed engine-driven warmup, shell-impact, parry, interception, cooldown, range, friendly-shot, cover and checkpoint checks. At wave 10, one Bombardier dealt 51.54 core damage in six seconds; a covering array intercepted both shells and retained full core health. A mounted array at 3.5 m took and retained 50 blast damage across reload; its cooldown and an airborne shell's gravity/splash also survived. Native captures checked both models, and the Windows release exported successfully. Combined all-defense/1,000-enemy runs exposed frame spikes, including 136 ms p99; another active game was sharing CPU/GPU, so isolated performance remains unverified. The existing benchmark now reports up to twelve slow frames with simulation phases, physics/process timing and render timing to aid diagnosis. No population cap or new test files were introduced.
Choose Campaign · 20 chambers from the main menu for Research Trials. The rebuilt campaign starts with separate jump, dash, slide, wall-jump, grapple, slam and cover lessons. Its layouts include a straight stepping-stone course, a long runway, a U-shaped service duct, a compact climbing shaft, suspended islands, a high observation well, a continuous cover corridor, a spiral around a central column and a three-wing trial. Chamber dimensions, platform footprints, vertical routes and collision geometry are individually authored.
Chambers 8–14 combine pairs of taught mechanics; 15–19 are longer mixed exams. Difficulty spikes at 8, 15 and 20 introduce wider crossings, more required relays, overlapping firing lanes and shorter warnings. The final chamber requires all six movement mechanics through separate relays. Movement plates activate through the indicated action; E only operates release consoles. Slide tunnels physically reject the standing capsule, wall shelves exceed normal jump height, and Rebound's high return requires the slam-jump boost. Lessons 1–6 have no sentries. Cover training uses 1.8-second aiming warnings, 4.5-second recharge intervals and 12 damage; the final exam uses 1-second warnings, 2.7-second recharge intervals and 25 damage. Sentries lock their aim during the warning, allowing movement or cover to avoid the shot.
Checkpoint pads save automatically; falls, lethal hits and U return to the last pad with restored health and dash charges. Activated relays persist after recovery and restart. Completing a chamber unlocks the next, with unlocked chambers available for replay. Campaign progress remains separate from the defense checkpoint and uses fixed health and movement equipment without research bonuses, construction or TP rewards. Existing unlocks, completed chambers and retry counts survive the remake; an old-layout checkpoint and its relay state reset to the selected chamber's entrance.
Remake verification loaded all twenty chambers with grounded starts, validated their map data, and checked every route-point elevation. Engine input checks crossed the introductory jump and dash gaps, compared a blocked standing capsule with a successful slide through the duct, climbed a 2.4 m shelf with a wall jump, landed from the introductory grapple, triggered a slam plate, reached Rebound's 3.4 m shelf and crossed chamber 8's long diagonal with two air dashes. In cover training, standing exposed lost 24 HP while the sheltered position retained 100 HP. Separate checks verified that E cannot bypass movement relays, old-layout migration preserves unlocks, all six final relays and the last checkpoint restore in another process, and the final exit opens the completion screen. Native captures inspected the duct, spiral, final layout and first-person instructions. These are focused engine-driven checks, not a complete human playthrough or a claim that every player's preferred difficulty has been established. No test files were added.
Enemy specialists retain targets across small distance changes. Attacks check visibility even at melee range; enemies can break a supporting wall when an elevated turret blocks their route. Ordinary enemies retain road navigation, with local recovery constrained to roads. Menders position near wounded groups and heal up to 32 nearby allies per pulse; recovery and positioning use bounded local queries rather than new per-enemy navigation fields.
Defenses now have stronger chassis: attack towers have 400–525 base HP, utility machines 360–550, and barricades 780. Older checkpoints receive the new durability while retaining their damage fraction and upgrade level. Breakers arrive at wave 5 and Saboteurs at wave 6, cost more threat, and strike more slowly for less damage. They acquire visible targets within 12/10 m and retain them out to 14/12 m. Only two hunters deliberately commit to a structure; unassigned hunters follow the core road. Road-blocking barricades can still be attacked by passing enemies. An amber line warns of a hunter's first strike. Switching targets no longer transfers accumulated attacks to the new target.
Aim at a damaged defense within 4 m and press E for a field repair, including during combat. Each repair restores up to 20% of its maximum integrity, costs proportional scrap, and has a shared 2.5-second cooldown. The HUD shows integrity, repair amount and price while aiming. Full repairs in the planning screen cost less. No-repair Challenge rules disable both paths.
- Arc Carbine: Plasma Cell research unlocks an energy rifle whose hits arc to up to two visible enemies within 5 m for half damage. Secondary arcs cannot chain again.
- Frost Launcher: Cryogenic Culture research unlocks explosive rounds that slow blast victims by 35% for 2.5 seconds.
- Cycle Accelerator / Cryo Injector: alternative barrel and elemental-slot attachments trade spread or direct damage for firing speed or slowing. Slot replacement returns the previous attachment with its rarity intact. Fired projectiles retain their resolved effects after switching weapons or reloading a checkpoint.
- Relay Lancer / Regenerator: road-following enemies introduced at waves 8 and 9. Lancers warn before precision bolts; Regenerators recover 4% maximum HP per second after four seconds without taking damage. Damage interrupts regeneration, including burns and secondary arcs. Both use cached instanced models.
Campaign verification used native room/menu captures and engine-driven map, console, checkpoint, reload, fall-recovery and completion checks. The full unlock/save sequence was exercised by positioning the player at route points and sending viewport E events, rather than a continuous hands-on playthrough. Route gap/rise clearance was reviewed against the real movement constants. A stationary sentry target lost 18 HP over six seconds; moving during its warning avoided that hit. All eight attack towers built on a fixed wall, and a Breaker damaged a mounted tower's steel support. The Windows release's 15-second, 1920 × 1080 horde check with about 1,000 active enemies averaged 119.71 FPS, with 10 ms p99 and 13.29 ms worst frame, on the RTX 3050 6 GB laptop GPU with VSync disabled and a 120 FPS limit. Arbitrary populations and a complete human campaign playthrough remain unverified.
Normal enemies follow marked road networks and their terminal core connections. Runners and exploders can intercept the player; breakers and saboteurs can take direct routes toward defenses. Navigation caches separate road-following and direct route fields. Barricades on a route are attacked rather than causing ordinary enemies to abandon the road. Custom maps without a connected road network still retain a traversable route to the core.
Water ripples animate in the shader and its physical floor is 0.65 m below the visible surface. Entering it visibly lowers the player; jump to climb out. Ground penalties apply near the ground, so crossing walls or elevated platforms above slow terrain does not inherit the terrain's speed penalty. Unstuck returns to clear ground near deployment, preserves health/resources and has an eight-second cooldown. It works during combat.
Hold Ctrl while opening Deployment's world dropdown to reveal Debug tile gallery. The labeled grid contains one sample of every tile/layer type. It is spectator-only: WASD flies along the view, Space rises, Ctrl descends and Shift increases flight speed. It has no collision, combat, construction, checkpoint replacement or permanent rewards.
The west gate opens after clearing wave 4; the east gate opens after wave 8. These changes and additional build permissions happen once, before the next preparation phase. Enemies can destroy player defenses and designated destructible sandbox geometry. Permanent world geometry ignores damage. Defenses can be rotated, moved for a 10% base-cost fee, or replaced using their remaining resale value. All management actions require preparation.
Kills award scrap and XP, and can drop rolled weapons, consumables, armor, shield packs, or attachments. Pick up with E. The backpack starts with 20 slots; equip or unequip either weapon slot, compare rolls, use equipment, sell, scrap, reorder, or drop items. Weapons have independent optic, barrel, magazine, grip, and elemental-module slots; attachments can be detached and reused. The Shock Blade accepts only an elemental module. Higher weapon rarities roll up to three compatible modifiers. Ammunition is replenished partly between waves; supplies sell more. Each cleared wave offers up to three functional temporary upgrades or supplies. Stat upgrades stop appearing at their effective caps; salvage remains available when all other rewards are exhausted. Temporary progression disappears when the run ends.
The run ends when the player or core dies. Eligible runs award TP once, with persistent statistics, records, achievements, and 120 functional research nodes across six branches. Technology uses a division sidebar, compact connected nodes, and a persistent inspector showing rank effects, prerequisites, and the next purchase price. Search and availability filters retain prerequisite paths. Scroll to pan, Ctrl + wheel or the toolbar to zoom, and use Focus to center the selection; purchases preserve the viewport. Optional respec refunds 90% of invested TP. Research unlocks equipment or grants bounded upgrades that apply to subsequent runs. Challenge Defense defaults to reduced starting health, faster enemies, and disabled repairs. Configure these alongside limited ammunition, marked-pad construction, ballistic-only weapons, and permanent acid rain. Selected rules are fixed for the run and determine its TP multiplier. Sandbox unlocks all equipment and never awards TP or lifetime records.
Scavenger halves starting reserves, removes free wave ammunition, and raises supply prices. Siege Protocol unlocks at a wave-20 record and starts at wave 20 with 1,500 scrap and all approaches open; its TP and cleared-wave achievements use waves actually survived. Daily Defense uses the same UTC date seed for everyone. Research and selected settings still apply.
Ascension becomes available after reaching wave 25 and researching 20 technologies. Its cycle panel displays both requirements, the core reward, and what resets or persists; the adjacent augmentation grid shows permanent ranks and exact upgrade costs. It optionally resets ordinary research and TP for Ascension Cores while preserving records, achievements, settings, and Ascension purchases. Cores unlock a configurable secondary loadout, a restoration ability, Resonant tower evolution, and starting mutations. The mutation selector shows the bonus at its purchased rank. Ascension and respec require finishing or abandoning the saved run. Run perks are snapshotted in checkpoints.
The operations menu shows your record and starting arsenal. Loadout separates the weapon catalog from base specifications and slot assignment; locked weapons link to their required research. Settings are grouped into Gameplay, Display, Audio, and Controls. Deployment indicates locked Siege Protocol and confirms when a new eligible run may replace an existing checkpoint. Supply, pause, results, and lifetime records use separate readable columns; inventory adapts up to four columns beside its item inspector. Construction offers category filters, selected-structure specifications, affordable upgrade actions, exact resale values, and an expandable world schedule. The editor keeps file actions above the canvas, with selected tools and grouped authoring controls in its sidebar. Sandbox controls scroll independently of their header.
Combat hints prioritize the current phase, nearby pickups, and consumable counts. Numeric HUD readouts and contextual searches refresh at 10 Hz; meters, movement effects, and damage feedback continue drawing each frame. This reduces repeated UI work without imposing an enemy population cap.
Controls highlight the action awaiting a key and reject conflicting gameplay bindings; Escape cancels capture. Rebuilding bindings replaces existing mouse events. The editor preserves its tile palette on redraw, keeps copied tiles within their source layer, and clears selection, clipboard, undo and redo state when opening a different map. Research specifications distinguish flat units from percentages, including additional projectile counts.
New profiles start muted, with 70% master volume and 25% music volume ready when unmuted. Existing saved volume and mute choices are preserved. Changes in Settings → Audio save immediately, including when leaving with Escape. Master volume controls music and effects together; music volume adjusts its relative mix. The game always mutes itself in the background and restores your chosen mute state when focus returns.
Field guide is available from the operations menu and pause screen. Its searchable categories cover every weapon, enemy, defense, attachment, weapon modifier, equipment item, run upgrade, technology, world and tile, plus movement, combat and progression rules. Combat specifications come from the live catalog; enemies include a preview of their actual mesh. The separate Achievements page shows all 25 requirements, earned/locked filters and progress. Best completed eligible-run progress is saved; sandbox never awards achievements.
All five attachment slots and seven weapon modifiers have catalog-defined effects. Heavy barrels trade precision for damage; optics trade firing speed for accuracy; extended magazines slow reloads; grips improve spread and recoil; thermal conductors trade a small amount of direct damage for burning. Inventory shows resolved stats and modifier descriptions. Nineteen run upgrades include armor breach, chaining, burning-target explosions and damage against slowed enemies. Synergy drafts require a supporting weapon, research bonus or defense. Rarity weights affect offers, capped choices are omitted, and each draft includes a lasting upgrade while one is available.
Attachment rarity strengthens useful bonuses while retaining their drawbacks. For example, a Common heavy barrel grants +16% damage and +12% spread, while Legendary grants +32% damage with the same +12% spread. Rarity survives swapping, detaching, checkpoint restoration and projectile elemental-effect snapshots. Inventory shows resolved percentages; the guide compares all five rarities. Weapons, enemies, defenses, attachments, equipment and tiles have fitted 3D model showcases. Towers use distinct industrial models and scan while idle, yielding to combat aiming.
Research uses compact selectable nodes, prerequisite links and stage columns, with separate branch lanes in the All systems view. Drag or scroll to pan, Ctrl + wheel to zoom, and Focus to center the selected technology. Filters, the purchase inspector and saved view positions remain available.
Each of the first 30 earned levels adds 6 maximum/current health. Levels add 6% weapon damage per level after the first, capped at +150%, and up to 2 HP/s recovery after the existing damage delay. Visible close kills restore 3 HP plus 2% of maximum health. Older checkpoints receive their earned health growth once. Eligible TP awards now use cleared waves: floor((5 × cleared^1.35 + 0.12 × kills + 25 × bosses) × mode difficulty), with the existing award limits and duplicate protection. Sandbox still grants no TP.
Levels update as XP is earned during combat. Earned choices wait until wave clear, with at least one choice for every cleared wave. XP thresholds start at 30, grow by 24% plus 8, and cap at 250,000. Pending combat levels and partially resolved drafts survive checkpoints. Run upgrades and XP reset on a new deployment.
At wave clear, combat XP resolves every earned level. Each level grants an upgrade choice, with at least one choice guaranteed per cleared wave. Multiple choices resolve in sequence, respect current upgrade caps, and persist with the checkpoint. Older checkpoints retain their one pending choice. Explosive swarm and Elite invasion compositions now spend the same threat budget as ordinary enemies; Elite invasion also compensates its population budget for the 45% elite health increase.
From wave 12, optional adaptive factions react to tower, explosive, close-range, or elemental damage. Planning explains the next faction. Adaptation changes composition weights, with blast/status resistance limited to 20%; it can be disabled in Settings for subsequent runs. Elites and a third boss phase add later-wave pressure. Twenty-five achievements track combat, engineering, survival, modes, and Ascension.
Weather includes rain, thunderstorms, and acid-rain events; ceilings shelter the player from acid rain. Settings can disable weather visuals, enable restrained camera motion and slide stamina, or choose a pixelated Retro or High Contrast scene. UI text remains sharp. Concrete floors and walls use mipmapped surface textures; other terrain retains procedural normal maps. Standard quality enables 2× MSAA, while Performance disables it. Bullet marks use a bounded reusable pool.
Content includes eleven weapon classes, thirteen enemy types including a phased Siege Titan every tenth wave, seventeen defensive structures and machines, seven attachments, seven tower targeting modes, two upgrade specializations, elemental effects, chain lightning, penetrating rail fire, projectiles, grenades, shields, and procedural audio. Eight firearm classes have distinct imported weapon models; the new energy weapons have native mesh models, with firing and reload feedback. No AI-generated artwork is included.
The bundled firearm models and muted Variation A color atlas are from Kenney's Blaster Kit 2.1. The 1K concrete color, OpenGL normal, and roughness maps are Large Floor Tiles 02 by Rob Tuytel / Poly Haven. Both asset sets are distributed under CC0 1.0 Universal; see also Poly Haven's asset license. Sources are also linked in Settings → Credits & licenses. Assets are bundled for offline play; no runtime download is required.
The main-menu editor authors separate custom worlds. It supports seven layers, custom dimensions, paint/erase, rectangles, selection, flood fill, copy/paste, undo/redo, rotation, tile properties, visibility controls, zoom/pan, metadata, JSON import/export, validation, and immediate sandbox playtesting. From a playtest, pause and choose Return to map editor. The normal-run planning interface cannot access terrain authoring tools.
Maps use version-2 JSON with shared tile identifiers; version-1 maps migrate when loaded. Save maps independently of progression. Incomplete drafts can be saved and reopened; playtesting additionally requires valid markers and routes. Validation checks known tile types, coordinates, numeric properties, initial routes, and every evolution/spawn-unlock stage. Normal-run worlds are built into scripts/world.gd and scripts/levels.gd; custom map edits cannot make a run eligible for TP. Supply, shop, ammunition, upgrade, generator, and chest markers have runtime behavior. Tower-only and structure-only pads enforce their categories. Wood, brick, concrete, steel, reinforced, and glass structures have distinct durability and resistance.
World rules configure threat budget, spawn interval, enemy and objective health, boss frequency, hazard damage, sunlight, and fog. Evolution events can be added, edited, or removed with completed-wave thresholds, structure removal coordinates, newly permitted construction tiles, and player messages. Tile inspection exposes terrain movement costs, wall durability/resistance/destructibility, and light color, strength, and range. Navigation uses cached weighted paths, and movement respects the same terrain costs.
Generate seeded 8–96 tile worlds in Balanced, Foundry, or Overgrown styles, or begin from any of the eight authored world templates. Generate, template, and shared-map imports can be undone. The community sharing dialog copies a compressed GRIDFALL1: code containing the full playable world; another player can paste that code to import it. Imports validate the decoded map before changing the canvas and enforce an 8 MB decompression limit.
Eight selectable worlds include the industrial Crucible, the bridge network of Floodworks, tall grapple routes of Skyline Yard, Switchback Quarry and Reactor Spine. These contain functional chests, supply/shop/ammo stations, generators, hazards, destructible cover, sheltered compounds, and approaches unlocked after waves 4, 8 and 12. All eight are available as editor templates and for seeded/daily selection.
The authored layouts use physical route networks: Foundry switchbacks, Relay's ring concourse, Crossroads' weighted flood crossings, nested Crucible circuits, Floodworks canal loops, Skyline's rooftop yards, Quarry bridges and Reactor's service galleries. Fixed geometry enforces their bends and flanks. Frontier connects newly opened districts to those networks. Enemy silhouettes are cached combined meshes; terrain and decoration details use instancing. Crates, machinery and pipes have matching solid collision and block enemy routing and building placement. Ground holes, walls, stairs and elevated slabs retain physical collision. Editor palettes and the planning grid use shared semantic tile icons; cosmetic vegetation, debris and lights remain passable.
Expanding Frontier grows the chosen world by eight tiles east and south every four cleared waves. It adds building ground, caches, hazards and grapple anchors, extends roads and moves outer portals. Existing coordinates and construction stay stable. Growth stops at 96 × 96 tiles to bound memory and navigation costs; waves remain endless. Expanded territory and collected caches persist in checkpoints.
Hold F to grapple toward a surface or enemy within 48 m; release to preserve momentum. Jump, dash, ground slam, or opening a menu cancels the hook. V punches with a 0.55-second cooldown and opens a 0.16-second parry window. One frontal projectile or melee attack can be parried per punch: projectiles reflect at double damage, melee attacks retaliate, and a parry restores 5 health and 0.35 dash energy. Visible weapon or ability kills within 5 m restore 3 health. Punch slows briefly; shotgun damage benefits from slowed enemies. Bindings are configurable.
Falling below the world recovers the player at deployment during preparation or combat, preserving health, resources and the active fight. Ordinary combat deaths still end the run. Weapon balance gives the blade and shotgun the strongest close-range damage, the SMG burst fire, the LMG sustained fire, and sniper/rocket shots penetration or splash. Runners close faster, spitters threaten at longer range, and laser focus caps at twice base damage.
Godot stores data in its user:// directory, normally %APPDATA%/Godot/app_userdata/GRIDFALL/ on Windows. profile.json contains research, TP, statistics, settings, and loadout. maps/ is the default custom-world directory. Writes use a temporary file, replacement, and a backup; invalid primary profiles fall back to a valid backup.
Eligible runs autosave after a cleared wave, after the upgrade choice, before starting the next wave, and when closing during preparation or combat. Pause to save manually. Version-2 checkpoints also preserve remaining spawns, enemy positions, health and status effects, in-flight projectiles and their firing effects, cooldowns, player velocity/crouch/stamina, Challenge rules, and RNG state. Combat checkpoints restore paused so the player can resume deliberately. Legacy version-1 preparation checkpoints remain readable. A valid backup can be continued even if the primary file is missing; malformed combat data fails validation before it can replace a valid checkpoint. Sandbox defeat or abandonment preserves an unrelated eligible-run checkpoint. Abandoning a run awards no TP.
Install Godot's matching Windows export templates using its editor. They are installed on this development machine. Choose Project → Export → Windows Desktop → Export Project, or run:
./run.ps1 -ExportThe Windows preset exports the game and data catalog. build/ and Godot's cache are ignored by Git. Linux support uses the same source; add a Linux export preset in Godot to package it. Linux exports have not been verified. Godot and third-party engine license notices are available in Settings → Credits & licenses.
The game uses a 60 Hz fixed physics simulation. The grid owns map layers, construction, and navigation; meshes are representations. Cached reverse grid fields use BFS for uniform terrain and Dijkstra for weighted terrain over packed numeric tile arrays. Destroying already traversable cover preserves these fields; opening a genuinely blocked route still rebuilds them. Construction management also preserves terrain routes. Visible nearby players can be pursued directly on continuous ground. Small spatial buckets limit separation searches. Complete floors use one collision shape and grounded enemies avoid redundant floor sweeps; holes, ramp neighborhoods, airborne and elevated actors retain full character movement. The minimap caches static terrain separately from live actors and construction, and uses hashed construction permissions. Ground meshes are instanced and tracer effects use a fixed pool. Weapon/enemy/structure/research/upgrade balance, loot rarity weights, and modifier compatibility live in data/catalog.json.
Incoming hordes are represented by integer counts per enemy type, with at most 601 explicitly queued enemies and 64-entry reserve refills. Wave generation visits at most 600 individual enemies before compressing the remaining budget, keeping arrival scheduling practical for millions of pending enemies. Live enemies have no gameplay population cap. Every spawned enemy retains its own health, statuses, physics collision, navigation and checkpoint record; inbound reserves are scheduled arrivals rather than a substitute for live enemies.
Live models use spatially chunked MultiMesh batches by archetype. A persistent spatial hash handles local targeting, splash, chains, melee, hazards, support healing and separation. Separation samples a bounded number of neighbors in dense cells; exact damage queries still visit every affected enemy. Nearby enemies receive priority and the remaining AI work rotates through the population against a CPU time budget. Deferred updates account for elapsed simulation time, including damage-over-time and repeated attacks; GPU motion extrapolation smooths short update gaps. Large minimaps draw occupied-cell density rather than thousands of individual dots. Physics bodies and memory still grow with actual population, so very large crowds can reduce update frequency or frame rate. Earlier benchmark paragraphs below describe the previous 100/250-enemy implementation.
Limited proximity queries start with the containing spatial cell and then visit cached offsets ordered outward, instead of scanning from the northwest. Sampling within dense cells remains bounded and rotating; results prioritize cells rather than guaranteeing the nearest individual enemies. Complete damage queries retain their full scan.
No test files or extra documentation are maintained. The development benchmark runs actual AI, tower targeting, weapons, damage, waves, rewards, and world evolution in sandbox:
./run.ps1 -Benchmark
godot --headless --path . --fixed-fps 60 -- --benchmarkThe benchmark logs wave clears, damage, kills, navigation time, simulation phases, memory, queue counts, and frame-time percentiles after a two-second warmup. --benchmark-content exercises every weapon and structure. --benchmark-population=1000 or --benchmark-population=5000 seeds a minimum live population in bounded batches and replenishes losses; ordinary wave arrivals can exceed that population. --benchmark-stress defaults to 1,000. --benchmark-horde=1000000000 additionally queues a billion enemies in compressed form; --benchmark-duration=30 changes the duration. Measure rendered performance without --fixed-fps, which supplies synthetic frame deltas for accelerated logic checks. Rendered captures can be generated with godot --path . -- --capture=menu (also selection, planning, inventory, editor, research, settings, wave, boss, results, sandbox, credits); source captures go into ignored .godot/captures/. Cosmetic randomness is independent of seeded gameplay.
The automatic shooter considers a rotating batch of at most 64 enemies when its weapon is ready, then shoots the nearest visible candidate in that batch. This bounds benchmark-only targeting work without limiting the live population. Slow-frame diagnostics and periodic samples report the benchmark's own update time separately from simulation, along with physics collision pairs. Earlier autoplay measurements included a complete enemy scan every physics tick and are not directly comparable to this revised workload.
Benchmark phase data also records peak tower target-search and attack time with the responsible tower type. A 5,000-enemy investigation measured a 165.5 ms Railgun target search versus a 7.24 ms peak Railgun attack before paged acquisition was implemented.
Tower acquisition now walks spatial buckets incrementally within a shared 2 ms search window, rotating the first tower each physics tick. It does not build a full candidate list or cap the number of enemies a search can consider. Searches advance during firing cooldowns, and valid existing targets can still be attacked at the usual interval. Completed passes refresh about ten times per second when the workload permits; large crowds make targeting reflect recent observations rather than an instantaneous global ranking. Every firing target is rechecked for life, range and visibility. Runtime search cursors are rebuilt after loading and cleared when waves end or defenses are destroyed.
Verification performed without test files:
- Godot parser/import checks, rendered menu/editor/inventory/research/settings/combat captures, and Windows export launch.
- A fresh standard run started with 300 scrap, built defenses using actual earnings, cleared 12 waves, killed a boss, and ended naturally on wave 13 with 122 kills and 112 TP. Buying the 90-TP SMG research and restarting restored its unlock in a fresh run, with temporary progress reset.
- A 190-second content scenario exercised all nine weapons and sixteen structures through waves 10–17. A separate 300-second core scenario cleared 14 waves with 155 kills.
- A fresh 30-second Windows release stress scenario maintained 100 enemies at 1920x1080 with sampled rates of 102-142 FPS (140 FPS final) on an RTX 3050 6 GB laptop GPU. Simulation samples were approximately 2.4-5.1 ms. These measurements are not a universal performance or long-term balance guarantee.
- Live movement, collision, jump, crouch, shooting, melee, and plasma damage; attachment reload/detach ammunition conservation; projectile effect snapshots; loot/equipment save/restore; duplicate-award rejection; invalid-map rejection; and backup recovery in isolated data directories.
- An actual tower-combat wave checkpoint restored its pending upgrade, structures, resources, and player cooldowns. Backup-only Continue, paused-close autosaving, Sandbox checkpoint isolation, and pad restrictions on moving/replacing structures were also exercised.
Very large custom maps and prolonged runs need more profiling. Release builds may not expose Godot's static-memory counter; the overlay reports it as unavailable in that case.
The expanded-content run exercised all nine weapons and sixteen structures through wave 17 over 190 simulation seconds, with 152 kills and a boss defeat. Isolated runtime checks covered maximum-rank research effects, Ascension reset/purchases, restoration, Resonant upgrades, perk checkpoint snapshots, 90% respec refunds, adaptive composition, and the new modes. Procedural maps validated at 8×8, 33×33, and 96×96; a generated map round-tripped through its share code. Native Compatibility and Forward+ launches verified weather, Retro rendering, advanced lighting, and the new authoring/research interfaces.
After optimizing navigation, crowd searches, floor collision, and grounded movement, the 30-second 1080p Windows release benchmark with one billion queued enemies and up to 250 active actors averaged 84.36 FPS on an RTX 3050 6 GB laptop GPU. Frame times were 16 ms at p95, 22 ms at p99, and 39.39 ms worst; 0.21% exceeded 33 ms. Simulation samples were 3.66–4.76 ms, with a 30.15 ms peak. Enemies dealt 57,482 core damage and 57 were killed, verifying continued combat rather than stalled actors. The sampled spawn buffer contained 56–61 entries. A separate million-enemy source run used approximately 61.5 MB of Godot-tracked static memory. These are bounded local measurements, not an unlimited-run or hardware-independent guarantee; at that stage the default was 100 active enemies.
Combat checkpoint verification restored enemies with burn, poison, and slow effects, a live projectile, ability cooldowns, and the exact next RNG value; resuming another checkpoint completed its partially spawned wave through tower combat. Legacy checkpoint migration, malformed combat-data backup recovery, all-on/all-off Challenge rules, and capped-upgrade continuation passed isolated runtime checks.
The follow-up audit restored a billion-enemy reserve with exact RNG state, rejected fractional/unknown/overflow reserve data, and drained a small reserve through actual spawning and a single wave reward. A billion-scale wave composed in 461 microseconds with only 601 explicit queue entries. Another standard run naturally ended on wave 13 after 126 kills and awarded 112 TP; the content scenario exercised all nine weapons and sixteen structures, cleared wave 17, and defeated a boss. Engine-dispatched B opened planning and pause froze time and ammunition. Imported coordinate aliases are rejected, the boss HUD displays phase three, and enemies falling below custom-map terrain return to an active portal without healing, rewards, or consuming RNG.
The movement update passed live physics checks for 12 m/s running, dashing, momentum slides, jump/slam damage, wall-jump push and the three-jump limit, dash charge exhaustion, disabled FOV effects, and exact paused checkpoint restoration with legacy-save compatibility. Rendered captures verified the speed lines and dash meter. Movement decorations draw at most 20 speed lines and one fixed-segment impact ring; movement audio reuses the existing 16-voice pool. The updated Windows export's 30-second horde regression run averaged 92 FPS at an actual 1920×991 window size with 250 active enemies and a billion queued, with an 18 ms p99 and an 88.44 ms worst frame. This benchmark uses stationary autoplay; moving effects were visually checked separately.
The final audit added editable world rules and evolution events, version-1 map migration, intermediate route validation, weighted terrain navigation, and consistent platform collision/destruction. Runtime checks also verified construction and checkpoint restoration on newly cleared ground, lighting-event overrides, support healing cadence, and research unlock snapshots. Core and optional systems in this scope are implemented; the performance and platform limits above remain. World geometry combines engine-native meshes, procedural materials, and the credited CC0 assets.
The combat/world expansion pass verified projectile reflection and enemy damage, frontal melee parries, punch cooldowns, close-kill healing, enemy grappling and release on target death, combat fall recovery, all nine weapon classes, stair traversal, cache persistence, six levels and their unlock routes, and expanded combat checkpoints across a fresh process. A fresh standard run cleared 12 waves, defeated a boss, and ended on wave 13 with 127 kills and 112 TP. At 1080p with 250 active actors and a billion queued, the Windows export averaged 94.8 FPS; a moving-player source scenario on a fully expanded 96 × 96 Floodworks averaged 97.7 FPS after minimap caching and floor/navigation improvements, with a 34 ms p99 and a 149 ms worst frame during world updates. These stress measurements used the RTX 3050 laptop above; occasional navigation-update spikes remain.
The subsequent navigation pass eliminated redundant graph rebuilds after cover and construction destruction. An unchanged 96 × 96 weighted graph previously rebuilt in 101 ms; removing the same cover now takes about 0.03 ms and retains cached fields. In the same moving-player stress setup, p99 frame time fell to 18 ms and the worst frame to 58 ms, with 94.4 FPS average and 0.18% of frames above 33 ms. A separate 20-second Windows export run at 1080p with a billion queued and up to 250 active enemies held the 120 FPS cap, with a 9 ms p99 and 10.4 ms worst frame. Door reachability changes, overlapping ramp removal, construction move/replace/sell/destruction, live map markers, and combat checkpoint restoration passed runtime checks. These measurements reduce the earlier spikes without implying a hardware-independent or stutter-free guarantee.
The UI overhaul was checked in 29 native screen/dialog captures. Runtime checks covered all six template picker paths, editor paste permissions and unique player markers, supply purchases that preserve scrap when no refill is needed, and viewport-dispatched rebinding conflicts, cancellation and successful remapping. Seven earned level-up choices survived a partially resolved checkpoint; legacy pending upgrades also restored. A fresh automated standard run ended naturally on wave 9 with 69 kills and 61 TP, which persisted after exit. Audio checks verified default mute, saved preferences, and actual Master-bus muting through engine focus notifications. Native desktop window activation was unavailable, so these were engine-driven checks rather than a hands-on playtest.
The rebuilt Windows release, including the audio changes, completed a 30-second stress run at an actual 1920 × 1080 on the RTX 3050 laptop: 113.88 FPS average, 12 ms p95, 16 ms p99, and 76.15 ms worst frame; 0.13% of frames exceeded 33 ms. It maintained 250 active enemies with a billion queued, continued dealing damage, and recorded simulation samples of 4.80–5.33 ms with an 11.87 ms peak. The release static-memory counter was unavailable. Temporary verification profiles and captures were removed afterward.
The world/progression/guide expansion checked all 253 field-guide entries, searches, enemy previews, 25 achievement requirements, completed-run progress persistence and duplicate-award rejection. Native captures covered the guide, achievements, audio, editor palette, all six world layouts and all ten enemy silhouettes. Fresh profiles used 70% master volume; changed volume/music/mute preferences survived a fresh process, and focus notifications muted the actual audio bus. All attachment/modifier effects, slowed-target damage, elemental resistance, eight pending combat level choices and partially resolved draft checkpoints passed isolated runtime checks. All six authored maps validated through Frontier expansion to 96 × 96; an actual player controller stopped at solid machinery. A fresh engine-driven standard run reached wave nine after eight clears and 72 kills over 180 simulation seconds; it stopped at the scenario duration, rather than defeat. These were engine-driven checks, not a hands-on desktop playtest.
The uncapped engine restored 1,000 actual enemies from a combat checkpoint with matching spatial/render records and ray-hittable physics bodies. An exact area hit damaged all 1,000, and a lethal area hit removed all of them from combat and spatial indexes. Deferred burn/poison/slow clocks and their checkpoint lags are tracked separately from movement: newly applied effects do not inherit earlier AI delay. A delayed-enemy check retained the full new effect duration on application and dealt exactly six combined burn/poison damage over the next half-second, including when a second hit refreshed both effects midway.
Final uncapped Windows release measurements used native 1920 × 1080, Compatibility rendering and the RTX 3050 6 GB laptop GPU, with a 120 FPS limit and one billion scheduled arrivals. A 25-second scenario finished with 1,029 actual enemies: 120.00 FPS average, 9 ms p95/p99, 8.33 ms worst, and no measured frames above 33 ms after warmup. The oldest outstanding AI update was 333 ms; all 1,029 surviving enemies had already received an update. A 40-second scenario finished with 5,002 actual enemies: 50.22 FPS average, 34 ms p95, 51 ms p99, 133.33 ms worst, and 6.54% above 33 ms. The oldest outstanding AI update was 2.77 seconds; 4,978 survivors had received updates, with recent arrivals pending. Combat continued with 360 kills and 911,185 core damage. Rendering used 93 spatial batches at the larger population. These are local stress measurements: uncapped spawning does not guarantee responsive AI or constant frame rate at arbitrary populations, and millions of simultaneous physics bodies have not been verified. Release static-memory reporting was unavailable. Separate checkpoint verification preserved effect clocks independently from the longer movement update lag. Temporary QA profiles and captures were removed afterward.
The 2026-10-19 gameplay pass verified road routes reaching the core on all six maps, road barricades/traps versus rejected towers, mounted tower height and support destruction, and valid checkpoint restoration afterward. The actual player settled at approximately -0.649 m in water and crossed a wall above terrain with 4× movement cost at normal speed. Viewport-dispatched U recovered a stuck player without changing health; Ctrl exposed the seventh world option, entered the spectator gallery, and its flight moved 14 m vertically in one second. Gallery combat, construction, rewards and eligible-run saves remained disabled. Legacy road construction restored, and earned health migration applied exactly once across repeated saves.
All 93 model-bearing guide entries loaded: nine weapons, ten enemies, sixteen defenses, five attachments, four equipment/consumables and 49 tiles. Native captures checked fitted showcases, rarity comparisons, planning hover text, minimap arrows, water and the compact research tree. A purchase through its actual UI spent 90 TP and unlocked rank one; zoom and view storage were exercised. A Legendary barrel restored its +32% damage and rarity after detach/equip/save/load. A fired Legendary elemental module retained rarity four and 8 DPS after switching weapons; its combat snapshot validated. A ten-cleared-wave, 100-kill, one-boss eligible result awarded 148 TP; duplicate and sandbox awards returned zero.
A 180-simulation-second engine-driven autoplay combat scenario cleared five waves, killed 35 enemies and continued wave six with no core damage; it stopped at its configured duration. This was a sandbox combat regression, not a human endgame balance assessment or a rendered performance measurement. The rebuilt Windows release's 25-second 1080p horde check finished with 1,029 live enemies, averaging 119.75 FPS with 10 ms p99 and 14.10 ms worst frame; no measured frame exceeded 33 ms. Oldest outstanding AI work was 417 ms. After final tooltip and placement-policy changes, the exact exported binary passed another ten-second run with 1,031 live enemies, 119.90 FPS average, 9 ms p99 and 11.51 ms worst frame. Large-population and hardware limitations above still apply. GDScript parsing and the Windows export passed, and temporary QA data was removed. These checks were engine-driven rather than a hands-on desktop playtest.
The arsenal/defense pass verified exact two-target arc damage, cryogenic splash after switching weapons, attachment slot replacement with preserved rarities, and combat checkpoint restoration of projectile effects and the new enemies. A regenerator retained two seconds of overdue healing across save/reload. Six hunters reserved two slots on each of two towers; changing attack targets with a negative cooldown caused no burst damage, and death released the slot. A legacy half-damaged level-two sentry migrated from 234 to 585 maximum HP without healing its damage fraction or growing again on reload. Field repair restored 117 HP for 13 scrap, enforced and persisted its 2.5-second cooldown, and respected no-repair and range restrictions; a viewport E event also repaired a damaged defense. Native captures checked the new weapon/enemy showcases. The final standalone 20-second Windows release check at 1080p finished with 1,064 live enemies: 119.44 FPS average, 9 ms p99, 47.56 ms worst, and one measured frame above 33 ms (0.046%). Oldest pending AI work was 550 ms. GDScript loading and Windows export passed. These are engine-driven checks, not a long hands-on endgame balance assessment; arbitrary populations remain hardware-limited.
A follow-up found and removed stale benchmark roster limits: content rotation had stopped at weapon index 8, and stress spawning selected only the first nine enemy IDs. The routines now use the current weapon catalog and all unlocked non-boss enemy types. A 240-simulation-second sandbox content run cleared waves 10–15 with 87 kills, one boss defeat and no core damage; all eleven weapons were equipped, and both new guns fired (Arc Carbine nine shots, Frost Launcher eight). This preset includes sixteen defenses, epic weapons and a 6,000-HP core, so it is an integration check, not normal-run balance proof. Headless fixed-step frame figures are not rendered performance measurements. The rebuilt Windows release's corrected eleven-type stress mix finished a standalone 15-second 1080p run with 1,001 live enemies: 119.04 FPS average, 12 ms p99, 14.79 ms worst and no measured frames above 33 ms. Oldest pending AI work was 533 ms. No new test files were added.
The contracts/supplies pass verified Operator and Engineer damage goals, Guardian success/failure, cache-only and mid-wave checkpoint restoration, legacy saves, malformed-cache rejection, duplicate claims, full-backpack salvage, rarity scaling, and no-repair/campaign restrictions through engine-driven checks. Native 1280 × 800 captures checked planning, cache rewards and an equipment showcase. A 15-second standalone 1080p horde run finished with 1,002 live enemies at 118.00 FPS average, 11 ms p99 and 15.56 ms worst, with no measured frames above 33 ms; oldest pending AI work was 533 ms. Contract tracking reads existing aggregate counters and adds no per-enemy processing. This remains a focused integration/performance check, not a complete human balance assessment.
The Quarry/Reactor pass preserved the original six map hashes and validated both new worlds through gate unlocks and Frontier expansion to 96 × 96. Checkpoints restored their offset cores and damaged defenses. Actual player input climbed the Quarry's two stair flights to its 2.8 m terrace and opened a cache; all four caches rejected duplicate loot. Construction checks rejected water, fixed stairs and road towers while allowing road traps. Native captures checked both environments. Separate 60-simulation-second combat scenarios recorded 15 Quarry kills and 11 Reactor kills with no core damage. These are engine-driven integration checks, not a full human playthrough or rendered performance measurement.
Barricade verification reproduced a blocked sentry before the fix, then checked all eight offensive tower types damaging a target across built cover, including rocket and mortar impacts, flame burn and cryogenic slow. An Interceptor removed a hostile bolt across the same cover. Actual character collision stopped both player and enemy on opposite sides; ordinary rays and a player projectile remained blocked. A hostile projectile removed 25 barricade HP without hurting the protected player. A fresh-process checkpoint restored the barricade collision layer and a functioning protected sentry. Fixed world walls remained opaque to rail targeting.
Turrets now wait 0.1 seconds after finding no visible target instead of repeating the complete search every physics tick. A blocked rail-turret scenario with 128 enemies took 94.5 ms across 120 updates before the change and 14.6–16.0 ms afterward; newly visible targets were acquired after the retry, and the normal 2.1-second firing interval remained intact. The rebuilt Windows release completed a 25-second mixed-content stress run at 1920 × 1080 with 1,021 live enemies and a billion scheduled arrivals: 110.4 FPS average, 20 ms p99, 138.9 ms worst frame, and 0.59% of measured frames above 33 ms. Early crowd-startup spikes remain; this is a targeted reduction in repeated searches, not a claim that large-crowd stutter is resolved. No test files were added.
Separating benchmark overhead identified 13–15 ms automatic-shooter updates during the early thousand-enemy spikes, versus roughly 4–5 ms of game simulation. After bounding the shooter, the rebuilt Windows release's 25-second mixed-content scenario at 1920 × 1080 finished with 1,017 live enemies, 183 kills, 120.0 FPS average, 9 ms p99, 8.98 ms worst, and no measured frames above 33 ms after warmup. A matching 5,000-enemy scenario finished with 5,001 live enemies and 333 kills: 107.7 FPS average, 29 ms p99, 138.84 ms worst, and 0.68% above 33 ms. Oldest outstanding AI work was 0.52 seconds and 2.77 seconds respectively. These measurements use the same RTX 3050 laptop; the revised automatic targeting workload does not improve normal gameplay simulation or remove large-population AI latency. No live-enemy cap was introduced.
A fresh standard profile with the revised shooter cleared nine waves in 180 simulation seconds, recording 84 kills, 152 rifle shots and six defenses bought with actual scrap. It continued against the wave-ten boss with 535 core HP when the scenario duration ended. This fixed-step check verifies continued targeting, reloads, rewards and wave progression; its frame figures are synthetic.
The proximity-order correction reproduced a 0.62 m enemy receiving no priority slot or AI update while a western crowd filled the selection. The same scenario now selects and updates that enemy on its first pass. Runtime checks covered adjacent cells, negative coordinates, typed queries, removals, and complete splash damage across 70 clustered enemies. The rebuilt Windows release held 120.0 FPS average with a 9 ms p99 at 1,004 live enemies over 20 seconds. Its 25-second 5,000-enemy run averaged 97.8 FPS with a 33 ms p99, 143.45 ms worst frame, and 2.77 seconds of oldest pending AI work; combat continued with 308 kills. These are local mixed-content measurements, not proof of reduced overall AI latency or a nearest-enemy guarantee.
The final tower-acquisition pass verified all seven targeting modes, late visible targets behind 256 occluded candidates, searches progressing during cooldown without extra shots, and rejection of dead, out-of-range or newly occluded targets. Built cover remained transparent to friendly towers. Destruction, mine detonation, menu exit and combat checkpoint restoration cleared runtime scan state correctly. All eight world layouts and twenty campaign maps validated. A fresh standard run naturally lost on wave ten after nine cleared waves and 84 kills, earned 107 TP, then spent 90 TP on SMG research; a subsequent process restored the remaining 17 TP and deployed with the unlocked SMG. A separate 240-simulation-second content run cleared waves 10–15 with all seventeen defenses and eleven weapons equipped; direct firing checks covered every weapon class. New profiles start muted, saved audio choices persist, and focus loss mutes playback.
The final exported Windows build completed separate 25-second mixed-content checks at 1920 × 1080 on the same RTX 3050 laptop. At 1,025 live enemies it averaged 119.71 FPS, with 10 ms p99, 13.96 ms worst frame and no measured frames above 33 ms. At 5,001 live enemies it averaged 90.74 FPS, with 31 ms p99, 139.85 ms worst frame and 0.81% above 33 ms. Peak individual tower search fell from the earlier 165.51 ms observation to 2.04 ms; the shared search window does not bound physics, attacks or the entire frame. Oldest pending AI work was 0.53 seconds and 3.72 seconds respectively. These engine-driven checks establish practical local results, not a full human campaign playthrough or support for millions of simultaneous live enemies. GDScript parsing and Windows export passed; no test files were added and isolated QA data was removed.