island-simulator

island-simulator is a peaceful first-person tropical island walking and near-shore swimming experience built in C++23 with CNA.

Current GPU safety status

AMDGPU hardware qualification is suspended after a confirmed GPUVM fault and desktop-session loss. The cause is still unresolved; do not treat successful software captures as a hardware fix. See the incident report. New wildlife and visual work is currently checked with CPU tests and forced llvmpipe/pixman. After building, a repeatable software-only visual check is:

python3 tools/run_software_capture.py build jungle

This requires kernel-journal visibility. If group membership was granted without a new login, use sg systemd-journal -c 'python3 tools/run_software_capture.py build jungle'. The command preserves its image, logs and kernel delta under build/software-captures/, forces at most eight llvmpipe workers, hides host GPU device nodes, and never clears the separate hardware qualification failure latch.

The project remains under an active long-form development and polish campaign. The current rendering policy requests 4x geometric-edge antialiasing on Linux/Web and 2x on Android through CNA's existing backbuffer path, with explicit applied-sample diagnostics and device fallback. See antialiasing evidence and cost; software results do not qualify hardware performance or resolve the GPU incident. The current playable slice renders a deterministic approximately 500 m island with portable procedural sand, wet-edge, soil, and slope material variation, including distance-faded interior substrate microvariation, broad leaf-humus and cool-moss floor breakup, close analytic soil relief and smooth pores faded by both distance and pixel footprint, sparse curved leaf litter, close dry-sand granulation, soft-edged pebble/shell flecks, subtle shoreline-raked ripples, restrained fixed-sun jungle-floor dappling, and a narrow noise-mottled, broken-edge earthen footpath that follows the same authored corridor reserved by vegetation. Two authored clearings use distinct sunlit-dry and leaf-rich ground treatments with broad irregular patch breakup derived from terrain noise that was already evaluated. Close jungle soil now adds one project-owned floor image with a 512-square mip chain and edge-safe rotated source windows, preserving the authored paths and clearing palettes. It is packaged on all targets; missing data logs a diagnostic and retains procedural soil. See material review for provenance, software-only runtime evidence and remaining visual limitations. Live jungle leaves now use a small mipmapped image to vary vein and tissue colour, preserving species tint and existing lighting; see leaf material evidence. An animated tropical ocean adds distance-faded crossing-wave colour structure, restrained shallow-ridge light and sunward ripple crests plus view-aligned grazing reflection bands from its existing slope field, with a shared nine-wave spectrum, gently irregular short-wave normals and per-wave pixel-footprint filtering. Grazing sky reflections become less transparent than downward views into the lagoon. Fine waves now vary in strength across slowly moving wind patches, with a wider spread of short directions and analytically consistent normals; broad water displacement is unchanged. Supplementary grazing contrast is bounded, and broad breaker foam is limited by shallow depth and shore distance rather than appearing across deeper water. Fine wave repetition still needs polish. The sea floor uses continuous sand colours across the waterline, with filtered shallow sand ripples; water tint and opacity increase with local depth instead of baking turquoise into the substrate. An exact shore-following foam/wash ribbon uses advancing, fading crests and pixel-footprint-filtered porous foam instead of a continuous white stripe. The landscape contains 370 procedural coconut palms in coastal groves, 277 broadleaf canopy trees (22 emergents in closed stands), 1,822 ferns/large-leaf plants, 579 low bushes, 3,116 small forest-floor plants, and 20,499 mesh-grass clusters placed from a shared canopy-light and ground-cover field. It also composes 157 irregular land rocks, 14 emergent near-shore reef outcrops, 13 mossy fallen logs with irregular near broken-end contours, 31 pieces of sun-bleached beach driftwood, 63 curved fallen palm fronds with asymmetric leaflets, and 265 near-field shell/pebble/dry-wrack clusters. Three varied driftwood silhouettes form a non-solid landmark arc in one sheltered cove; five rocks and two fallen logs form natural orientation landmarks around the outer rings of the two authored clearings without blocking their central walking room. Every dense near-island visual layer has spatial culling and appropriate distance reduction, while vegetation uses a calm breeze. Five unreachable, haze-softened low islets and 11 simplified palm silhouettes break up the open-ocean horizon in one intentionally always-present portable static batch. One beach shallows now contains a large original wooden sailing-ship wreck rather than a small boat. Its exact 20 m hull is grounded and heeled in the water with a central break, exposed ribs, partial deck, cabin, snapped mast, bowsprit, and rigging. Twelve planks and substantial fragments form a separate trail toward the sand. Six conservative hull proxies keep the player out of the major structure while the smaller debris remains non-solid and traversal-friendly. The complete 3,948-triangle landmark uses a finer curved hull section, localized missing planks, a split broken keel and rudder, broken deck rails, dark hull openings, and a splintered foremast stump. The debris trail varies its angle strongly in both directions instead of repeating the ship axis. One portable GLSL 330 / GLSL ES 300 material adds distance-faded longitudinal grain, weathering, restrained cracks, wet/submerged darkening, salt wear, and a small waterline sheen. It remains frustum/distance culled and is submitted as one batch before the transparent ocean colours its submerged portions; it introduces no quest, pickup, or objective. Nearby mesh grass bends away from the player: a small stationary body radius expands smoothly with walking speed and releases after movement stops, entirely inside the existing instanced shader. The 42-blade near mesh uses five curvature levels per blade, geometric V-folds, and portable two-sided daylight/root/transmission shading. Shared surface normals follow rest-shape curvature and instance scale; wind/player motion still uses approximate normal response. Pigment now stays separate from fixed side lighting, with restrained filtered longitudinal grain; see grass surface material and evidence. Its close blades are deliberately narrower than the middle tier, use small deterministic left/right edge asymmetry, and occasionally bend and droop at the tip instead of repeating broad symmetric ribbons. Their upper width now tapers progressively, and seed stems attach to the actual curved terminal centre/ridge. A sparse juvenile underlayer, six restrained seed-bearing stems in the near cluster, and deterministic per-instance blade stature, colour, and wind variation keep neighbouring tufts from repeating the shared mesh literally; the cheaper far mesh uses three levels and still merges into the terrain before culling. The 42 near blades form seven unequal loose tufts, with outward-facing leaves, short peripheral blades and related wind phases. This reuses the existing geometry and draw budget. Nearby palm crowns use nine paired, fuller tapered-leaflet stations on each of 10--12 curved fronds instead of solid ribbon silhouettes. A cheaper widened ribbon preserves the crown envelope at distance, while per-palm 125--145 m thresholds stagger the transition and avoid switching whole cells at once. The near tier also adds texture-free leaf-scar rings to the trunks and restrained faceted coconut clusters to two of the four variants without another draw submission. Five irregular faceted root wedges vary in angle, length, width, and height around each near trunk instead of ending it in a flat cylinder or a regular stamped star. The palm shader also derives a restrained green-only frond tint from each existing instance transform, so neighbouring crowns no longer share one literal green while trunks, roots, coconuts, near/far identity, buffers, and draw count remain unchanged. Two selected near variants also carry two or three irregular hanging dry-frond remnants beneath the living crown. Their broken ochre-brown leaflets add sparse age variation without turning every palm into a uniform dead skirt; the far silhouettes and submission count remain unchanged. Nearby broadleaf crowns use small irregular inner foliage masses, visible tapered support branches, and layered shells of 584 curved canopy blades per tree. Golden-angle sprigs, unequal leaf sizes and height-dependent tints break up smooth silhouettes without alpha blending or another draw pass. Eight shared crown forms vary branch architecture as well as leaf posture. The near meshes have 4,328–4,378 triangles; the 1,064-triangle far tier retains 104 original blades and primary branches. The foliage is still procedurally stylized rather than photorealistic. Terrain, grass and jungle ground layers share approximate fixed-sun shade baked from the placed broadleaf crowns, replacing unrelated noise-based sun patches. The static 4 MiB field preserves ambient skylight; it is not leaf-exact shadow mapping. Boars sample the same light per body vertex, retaining a 1 MiB CPU field without another GPU texture or draw. Monkeys reuse it through a smoothed body sample and the existing skinned material, including reduced direct specular light in shade. Other wildlife and palms are not yet integrated. Understory, bushes and small plants project their elevated leaves along the same sun direction, retaining ambient light beneath crowns instead of remaining uniformly bright against shaded soil. The same near meshes give trunks a restrained secondary growth bend and slight ring/azimuth girth variation, while each support branch tapers through a deterministic off-axis middle joint instead of reading as one straight cylinder. Near/far crown anchors agree and collision proxies remain unchanged; the expanded form set adds shared meshes and variant batches. Two of the eight near variants also carry one sparse hanging liana with a curved tapered strand and two small leaves, adding vertical jungle detail without another instance or draw. Their near trunks now meet the forest floor through four or five short irregular buttress roots instead of a uniform cylinder, using the same instanced mesh and draw submissions. Close bark gains restrained procedural longitudinal grain and narrow fissures that fade before the far tier, without a texture or new pass. A deterministic per-tree 105--129 m threshold staggers their transition to the far mesh instead of switching complete scene cells at once. The shared portable jungle shader uses smooth normals on curved and pointed leaves, with subtle footprint-filtered veins and a restrained waxy sun reflection nearby. Fine material detail fades between 16 and 38 m; bark and other geometry retain their geometric two-sided lighting. Leaf coordinates and normals share the existing vertex buffer; no texture or draw was added. Its vertex stage also recombines each existing instance transform's scale/yaw coefficients into a stable restrained foliage tint. The green-only mask leaves bark unchanged, and the shared transform keeps the same tint identity across near/far LOD without another attribute, texture, or draw. Low bushes use supported shoots and individual curved leaves. Both detail tiers retain the complete leaf outlines; far removes inner rib vertices without extra draw submissions. Two medium understory forms now use pinnate fern fronds, with tapered paired leaflets aligned to a curved rachis and attached through shared wind vertices. Far meshes retain original leaflet identities. This adds geometry but no material, texture or draw; mobile cost and moving LOD quality remain unqualified. Two other forms use broad paddle/lance profiles with mildly asymmetric margins and cupped surfaces, reusing the same mesh budget. One near forest-floor variant forms six upright broad leaves on visible petioles, while another uses five low folded split leaves instead of repeating the strap-leaf rosette. Their far representations retain reduced original leaf subsets with unchanged batching. Understory, bush, and forest-floor LOD and maximum-distance selection use actual instance positions after a conservative cell reject, so first-person detail no longer changes early at a coarse cell boundary. Land and reef rocks use shared smooth normals, scale/yaw-correct fixed-sun lighting and refined near silhouettes. Texture-free mineral mottling, filtered fine grain and restrained lichen replace the former triangle-facet lighting and repetitive strata. Fine detail fades before the far LOD; broad colour and daylight remain consistent. A fixed mean-sea-level band darkens their wet bases. This uses one additional normal attribute in the existing instanced pass, not another draw or texture. A restrained warm/cool tint derived from each existing instance transform varies neutral stone tones while preserving one identity across the near/far LOD and leaving wood and foliage colours unchanged. Nearby fallen logs use the same portable material path for gently irregular cut-end growth rings, restrained radial drying cracks, and longitudinal bark fissures. Both rock and wood detail fade out before their LOD transition. Wood surface normals now follow each instance's rotation and nonuniform scale, with outward-lit cut ends and separate hard rims. Far wood keeps the same daylight response without close detail. Sun-bleached driftwood adds a still softer near-only longitudinal weathering treatment. Its third near variant replaces a simple spike with three curved tapered fork limbs, while cut ends and all far meshes stay calm and inexpensive. Rocks, fallen logs, and driftwood select range and LOD from each instance after a conservative cell-bounds rejection rather than switching an entire cell. Nearby palms, canopy trees, land rocks, fallen wood, driftwood, monkeys, and boars receive subtle terrain-conforming contact shadows. The twelve-segment soft ellipses use one portable BasicEffect batch, fade before their bounded draw distances, and cache static terrain sampling plus movement-thresholded uploads so this grounding cue does not become a per-frame procedural terrain cost. Land-rock and wood ellipses are split at rendered terrain triangles to avoid coarse-surface mismatches. Fallen-log and driftwood model heading now matches placement/support and collision heading; terrain layout and walking routes are unchanged. Reef outcrops, the two broken hull sections and twelve wreck fragments now also receive soft grounding patches. These are clipped to the actual seabed/terrain triangles, drawn beneath the water in the existing batch, and preserve the gap between the hull halves. They remain contact approximations, not full mast/rigging or directional shadow silhouettes. It supports first-person terrain walking, near-shore swimming, trunk collision, and a soft ocean-current boundary whose always-shoreward return includes mild coherent coast-parallel circulation, dual-zone touch input, a subtle depth-responsive near-surface water veil with upper-surface light scatter that cools and strengthens as the offshore current develops, and speed/depth-responsive player ripples that expand across the same animated ocean surface without another draw call. Moving entries and exits also trigger one short speed-scaled pair of broken surface rings, with depth hysteresis and a cooldown that prevent shoreline chatter. A fixed sunny gradient sky adds a sun disc and sparse, layered horizontal high-cloud wisps without a weather simulation. A first flock of 24 procedural birds follows varied shore, canopy, and high-altitude routes with animated wing strokes, layered trailing flight feathers on the near silhouette, intermittent glides, banked turns, slow radial route drift, terrain clearance, visibility culling, and distance-throttled simulation. Six non-hostile wild boars now forage, rest, wander, and retreat from nearby players in bounded inland habitats, with solid-prop avoidance, adult/juvenile procedural models, continuous body/head profiles, folded ears, near nostrils and bent tusks, smooth fixed-sun shading and terrain-following contact shadows. An original 128x128 opaque directional coat atlas adds fine body strands through stock BasicEffect; a padded white region leaves eyes, tusks and hooves unpatterned. The three coat tints and juvenile markings remain shared with the mesh colors. Legs use connected tapered segments and split near hooves, with a grounded stance and lifted return phase; short swept bristles follow the same body profile. This is still approximate procedural animation: individual soles fit the rendered terrain with fixed-length limbs, but world-space foot locking and stride-speed synchronization remain unfinished. Use --viewpoint boars for their dedicated inspection camera. This original asset still needs visual refinement; its current runtime validation covers native llvmpipe and actual browser WebGL 2 through SwiftShader, not post-incident AMD or Android device qualification. The juvenile broadleaf ground plant now uses a curved leaf surface and retains real source leaves in its reduced-detail model. Use --viewpoint ground-broadleaf for its inspection camera; see ground broadleaf validation for exact scope and cost. The split-leaf ground plant uses connected, curved lobes with shared near/far source identities. Its --viewpoint ground camera and validation notes track the remaining terrain-intersection and LOD limitations. Shrubs now use rooted woody shoots and smaller individual curved leaves, retaining all source leaves and outlines at far LOD. Use --viewpoint bush for the targeted camera; shrub validation records the increased geometry cost and remaining coverage/animation limitations. The floor layer includes 199 irregular canopy-associated groups with 866 smaller companion plants, while preserving paths and clearing accents. See placement validation for bounds, cost and visual limitations. Jungle root origins now follow the actual rendered terrain triangles; see root-contact validation and remaining slope limits. Near rosette litter also follows slopes through the shared terrain height map; contact measurements and remaining limits document the change. Thirty small beach crabs occupy six dry/wet-sand colonies and alternate asynchronous pause, forage, and lateral scuttle states. They stay inside terrain-validated home ranges, retreat briefly from a nearby player, and use articulated near/far silhouettes in one culled batch. Eighteen butterflies occupy three sunny clearing-edge habitats, alternate rest, flutter, and nectar-hover behaviour, and stay inside low terrain-relative altitude and home-range bounds. Five palettes, hinged patterned near wings, a reduced far silhouette, and one culled batch make their motion readable without filling every clearing. Eight near-shore schools add 48 animated fish constrained between the seabed and water surface, with depth-following pitch, banked turns, forked tails, CPU-lit body facets, restrained depth/caustic colour response, and readable near heads. Five non-hostile sea turtles follow separate bounded near-shore routes between the seabed and surface. They alternate slow cruising, seabed foraging, and occasional surfacing, remain outside land and the outer-current region, and respond mildly to a nearby swimmer. Faceted scute-patterned shells, articulated flippers, three natural palettes, near/far geometry, and one culled dynamic batch keep them readable through the water without making them a constant attraction. One non-hostile shark occasionally patrols deeper offshore water with matching dive pitch, turn banking, a forked tail, depth-integrated colour, and restrained near eye/gill detail. Seven original procedural monkeys roam the interior using a compact, obstacle-aware state machine, three silhouette meshes, four material palettes, portable CNA GPU skeletal animation, and short state-transition crossfades. Near animals add restrained crown/cheek fur tufts plus bone-following finger and toe silhouettes in the existing body draw. Calm idle/observe animals gradually turn their existing head bone toward a nearby player within a restrained angle, then relax or retreat without becoming hostile. The current audio pass adds original procedural shoreline surf, two varied occasional stronger waves, light canopy breeze, a subtle daytime-jungle texture, speed/depth-responsive movement through shallow or swimming water, and two restrained water-entry/exit textures. The same continuous water layer now develops a subtle deeper undertow bed as the 42--72 m offshore current strengthens, without adding another persistent source. Two bird phrases and two monkey contact phrases share one positional voice attached to actual nearby agents. All four persistent loops remove their measured mean before PCM quantization, avoiding long-lived DC bias without altering one-shot silent endpoints. Distance-scheduled player footsteps select paired dry-sand or interior-soil textures from actual resolved travel, vary pitch/pan/gain, and remain silent while swimming or blocked. All audio is generated without external sound files. The complete scene runs on Linux/OpenGL 3.3, in an actual Chromium/WebGL 2 context, and in a KVM-accelerated Android API 35 x86_64 emulator through OpenGL ES 3. The accelerated Android pass also exercises movement and look gestures, AAudio startup, and background/foreground recovery. Android uses a centralized mobile quality default that retains all content and close detail while reducing only distant grass/shadow/wildlife work and the ocean grid; Linux and Web retain the full portable profile. Physical-device touch ergonomics, thermals, and sustained mobile-GPU performance are not yet qualified.

There are deliberately no quests, combat, inventory, crafting, survival systems, score, progression, dynamic weather, or day/night cycle.

Dependency baseline

The project consumes read-only CNA next and Sharp Runtime next checkouts. Exact qualified commits are stored in dependencies.lock. The current workspace checkouts are both on next; the project still creates an ignored, unmodified CNA source export so its build input remains pinned:

tools/materialize_cna_next.sh

This does not switch or edit the CNA repository. A separate read-only CNA checkout on next can be used instead with -DCNA_ROOT_DIR=/path/to/cna.

Linux development build

The machine-wide ccache is required. Sharp Runtime's stricter build policy limits builds that compile its sources to two parallel jobs.

export CCACHE_DIR=/rv/cnaccache
export CCACHE_BASEDIR=/rv
tools/materialize_cna_next.sh
cmake -S . -B build -G Ninja \
  -DCMAKE_BUILD_TYPE=Debug \
  -DCMAKE_C_COMPILER_LAUNCHER=ccache \
  -DCMAKE_CXX_COMPILER_LAUNCHER=ccache \
  -DCNA_GRAPHICS_RENDERER=OPENGL33
cmake --build build --target island-simulator island_core_tests island_world_report \
  island_audio_report --parallel 2
ctest --preset linux-core

The graphical tests need an isolated display; use the command in docs/build-linux.md rather than running the full CTest set on the owner's live desktop.

For a stripped, relocatable Linux release bundle, use the dedicated two-worker preset and packager:

export CCACHE_DIR=/rv/cnaccache
export CCACHE_BASEDIR=/rv
cmake --preset linux-release
cmake --build --preset linux-release
tools/package_linux_release.sh build-consumer

This produces build-consumer/island-simulator-linux-x86_64.tar.gz. The archive contains the game, the exact built libcna.so, dependency licences/notices, attribution, dependency pins, and a SHA-256 manifest. The executable resolves CNA through $ORIGIN/lib, so the extracted directory can be moved and launched from an unrelated working directory. Compatible system SDL 3, SDL_mixer 3, glibc, libstdc++, and an OpenGL 3.3 driver remain explicit host dependencies rather than silently copied distribution-specific libraries.

Run build/island-simulator. Click inside the window or press a movement key to capture the mouse, use WASD or the arrow keys to move, and press Shift to toggle between walking and running on land. Each keyboard toggle briefly confirms WALK MODE 4.3 M/S or RUN MODE 8.0 M/S near the bottom of the view; no key must remain held. The selected walk/run mode, camera-motion toggle, audio mute state, and Android touch-look sensitivity are saved automatically through CNA's platform storage and restored on the next ordinary launch. Isolated smoke/capture/benchmark runs do not read or change player preferences; --no-user-settings provides the same isolation for an explicit diagnostic launch. The executable may be launched from any working directory; current runtime content is embedded or generated and does not depend on a repository-relative asset path. On Linux, press Escape once to release the mouse and again to exit. In a browser, Escape only releases pointer lock; close the tab through the browser UI when finished. Mouse look remains active while a movement key is held. The controller follows the terrain, rejects overly steep uphill motion, slides along trunk, solid-rock, and fallen-log proxies, and enters swimming water with hysteresis. Swimming retains collision with solid emergent reef rocks. A static 16 m collision grid limits each capsule test to nearby proxies; bounded one-time recovery prevents a placement or debug teleport from leaving the player embedded in a static proxy. Geometric step-up and continuous recovery from moving obstacles are not implemented yet. Walking and swimming apply restrained, speed-responsive positional camera motion; press B to smoothly disable or restore it. Press F1 to toggle the developer overlay with frame timing, player/water state, renderer identity and the actual logical backbuffer size, culling counts, contact-shadow class/draw statistics, vegetation draw statistics, ambient mix levels, grass-interaction strength, player-ripple and transition-burst intensity/phase, player-water gain, wave/splash counts, undertow cue strength, and positional bird/monkey call activity. While F1 is visible, F2 toggles an x-ray world diagnostic: the authoritative shoreline and three ocean-current thresholds, nearby collision proxies, the player footprint, shared 8 by 8 scene cells, all nine terrain-contact samples and normals, wildlife headings, simulation-tier colours, nearby fish-school centres, the shark route target, monkey home ranges and active movement targets, three bird-family route volumes, all 24 authored bird spawn headings, and nearby live radial targets, the last exact attempted/resolved player step with rejection classification, the exact active inward/circulating-current vector, all six ambient crossfade boundaries, the nearest bird/monkey audio candidates, and any active positional wildlife emitter. The panel also reports active audio sources against the six-source ceiling. In debug builds, F3 toggles a bounded inspection camera while F1 is visible. Mouse look remains unchanged; WASD flies horizontally, Q/E moves down/up, and Shift increases speed. The player stays fixed, and hiding F1 immediately restores the first-person camera. The overlay also shows the four authoritative procedural-layout seeds. Press F4 while it is visible to copy a stable seed report; the panel reports COPIED only after reading back the same text and otherwise reports UNAVAILABLE. Press M to mute or unmute the soundscape. Press B to toggle locomotion camera motion. While the overlay is visible, F5 through F12 teleport among eight fixed QA locations, including views toward the south fish shallows, offshore shark, monkey troop, and the active offshore-current field. The overlay starts hidden in a normal run and can be enabled at launch with --debug-overlay for automated QA. Use --world-debug to enable both the panel and world visualization at launch. The 2026-10-08 AMDGPU GPUVM incident currently requires the fail-closed serial procedure in docs/incidents/2026-10-08-amdgpu-gpuvm-reset.md; a private compositor still shares the physical GPU with the desktop. After that incident is closed, tools/run_native_benchmarks.sh provides warm, uncapped JSON frame-time reports across five fixed scenes plus a 676 m moving traversal. Reports include min/max/mean scene-population ranges, timing distributions, and exact 120/60/30 Hz budget-miss counts; methodology and the current Linux release baseline are in docs/performance.md. Use --startup-report <path> to write a machine-readable first-draw report. It separates deterministic world construction, framework initialization, GPU/content creation, the interval before the first draw, and the first submitted draw. Combine it with --smoke-test --overview for an unattended run; the measurement ends before CNA presents the frame and is therefore not a GPU-completion timer. The native CTest suite validates the report after an actual OpenGL 3.3 draw. Use python3 tools/validate_assets.py to enforce the asset provenance gate and build/island-world-report --output <path> for deterministic terrain, placement, collision, wildlife, seed, and fingerprint statistics. Native overview captures are followed by tools/validate_capture.py, which rejects truncated, black/white, flat, or spatially degenerate P6 output and writes machine-readable metrics. See docs/asset-pipeline.md and docs/debugging.md. Use build/island-audio-report --output <path> --wav-directory <directory> to audit every procedural source and export disposable PCM16 WAV previews for focused listening. See docs/audio-review.md.

The automated browser smoke test verifies renderer initialization, the portable terrain-material, animated-ocean, and fixed-sky shaders, hardware-instanced palm/jungle/prop LOD, portable instanced grass wind, a finite first-draw startup timing line, four frames, clean shutdown, active overview/beach/jungle/prop/driftwood/reef/shipwreck/crabs/butterflies/turtles/water/ current/grass/frond/distant/bird/fish/shark/monkey screenshots, the transient run-mode confirmation, and portable overlay text rendering. The close prop view additionally requires a nonzero one-draw contact-shadow batch. The driftwood view frames the authored three-silhouette cove group, including the forked third variant, and requires a nonzero per-instance near-LOD population. The Chromium harness also drives first-click pointer lock and initialization of all ambient and positional-wildlife audio resources, WASD, arrow-key movement, Shift run/walk toggling, simultaneous forward movement and relative mouse look, a deterministic beach-to-water crossing with successful transition-audio playback, successful land-footstep playback, F2 world lines, F4 seed-copy execution with explicit clipboard status, F1, real-Escape pointer release, continued live frames, delayed click reacquisition, canvas focus, and responsive 16:9 sizing. A narrower Firefox ESR 140.15 WebDriver BiDi gate independently checks the production shell's first frame, reported WebGL 2 Radeon backend, trusted-click pointer lock, 44.1 kHz audio initialization, persistent Shift enable/disable edges, forward traversal with an actual footstep, Escape release, delayed pointer-lock reacquisition, and responsive 16:9 sizing. It also loads dedicated 1280 by 720 jungle, near-surface-water, 20 m shipwreck, beach-crab, butterfly, and sea-turtle scenes, requires their live state, and captures each. All Firefox scene frames were visually inspected. A visible interactive listening pass, the remaining fixed-scene matrix, long tab suspension, audio-focus behaviour, and fullscreen remain open. The automated Firefox path now does one actual second-tab activation cycle and requires visible to hidden to visible, resumed animation frames, and fresh game input after foregrounding.

After the WebGL 2 release build, tools/package_web_release.sh build-probe creates build-probe/web-release/ with verified uncompressed artifacts, deterministic gzip variants, and a SHA-256 manifest. The current gzip payload is 2,587,860 bytes, 71.5% below the 9,085,111-byte uncompressed package. See docs/build-web.md for required server content types and content-encoding headers; the quick Python development server does not exercise compressed delivery.

The Android project produces arm64-v8a device and x86_64 emulator APKs with an explicit GLES 3 renderer. The left half of the screen supplies a floating movement stick and the right half uses relative drag-to-look, with subtle generated premultiplied-alpha control indicators. Near-full movement-stick travel selects the faster land pace with a warm-gold visual cue; easing inward returns to analog walking. A small three-dot control at top right cycles low, standard, and high look sensitivity (0.72/1.00/1.35 times); the highlighted dot shows the current choice. Android Emulator 36.5.11 runs the API 35 x86_64 AVD with KVM CPU acceleration and host-GPU gfxstream. The complete current scene, independent movement/look gestures, all sensitivity levels, a same-process hot resume, force-stop/new-process relaunch, and AAudio pause/resume are validated there. Dedicated current-source GLES 3 views also validate the inland humus/moss floor material, the layered near-canopy edge plus irregular trunks and bent support branches, clearing butterflies, a near-shore sea turtle, and sparse hanging dry-frond variation through the mobile quality profile. The persistence-qualified run also requires a touch sensitivity change to survive the cold relaunch. The default arm64-v8a APK is 19,082,004 bytes, contains a 16,271,720-byte stripped libmain.so, passes app lint, and verifies with APK signature scheme v2. This is meaningful emulator qualification, but it does not replace physical-device ergonomics, thermal, driver, and sustained-performance testing. See docs/build-android.md and PLAN.md for exact evidence and commands.

Project records

• PLAN.md is the living development ledger.
• BUGS.md records confirmed CNA or Sharp Runtime defects and game-side workarounds.
• ASSET_LICENSES.md records all asset provenance before assets enter a runtime build.
• docs/asset-pipeline.md defines the manifest, validation gate, deterministic world report, and external-asset intake procedure.
• docs/audio-review.md defines deterministic source analysis, WAV preview export, and the remaining subjective listening protocol.
• docs/performance.md defines the repeatable benchmark method, current evidence, and open budgets.