Repository navigation
GPU globe by default, physically based atmosphere - #5
Merged
Merged
Conversation
{key:<service>} in a layer address was filled in whatever host the address
named, so a layer (or the assistant, through data.add) could send the TomTom
key anywhere. Keyed services now list their hosts and fill() refuses others.
Keys are also taken out of the text of failed requests to filled addresses.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TXuzbw772NsDthvWr8VnjE
data.add ran without approval, so text the model read could make it fetch an arbitrary address. A new Reach::Network gates data.add when it names a url (library presets still run unasked); "allow for this conversation" remembers the host. Addresses on this computer or the local network are refused for the assistant. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TXuzbw772NsDthvWr8VnjE
settings.json, open-kml.json, lang.txt, the key files, KML/KMZ saves, planet documents and exported sources were truncated and then written, so a crash or full disk left a short file. They now go through paths::write_atomic (temporary file, sync, rename). A settings.json that does not parse is no longer silently replaced by the defaults on the next save: it is renamed settings.json.corrupt-<seconds> and the status bar says so. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TXuzbw772NsDthvWr8VnjE
The time-zone suffix was sliced by byte position after checking its byte length, so a value such as 2026-11-04T09:38é1234 in a user's CSV or GeoJSON panicked and failed the whole layer. The suffix is now matched as bytes. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TXuzbw772NsDthvWr8VnjE
- http::body and conditional_body cut a response at its limit and returned the first bytes as success, so a large CSV layer loaded with rows missing. They now fail with the new Error::TooLarge; aircraft drops its +1 workaround. - The Nominatim client had no timeout; a stalled connection left the search pending for good. It now has 10 s to connect and 20 s in all. - User agents said 0.1, 0.2, 0.3 or 1.0; all now carry the crate version. - Removed a stale doc comment above http's test module. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TXuzbw772NsDthvWr8VnjE
…te failures - The Wayback release list was the last file in geo/, named after its URL by geo_cache_path (which could collide). It is now wayback_releases.json and written atomically; geo_cache_path is gone. - remove_retired_caches deleted every file in geo/ except that list on every start, so anything later stored there would vanish. With geo/ unused it now removes the folder outright; nothing may store files there. - Failed tile disk-cache writes and a failed Wayback list write were dropped silently; the first of each is now written to the run log. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TXuzbw772NsDthvWr8VnjE
…ayer - N, U, +, = and - on the globe were read whatever modifiers were held, so Cmd/Ctrl with + or - zoomed the interface and the globe at once. They now act only without Cmd, Ctrl or Alt. N and U run camera.resetNorth and camera.topDown, so both now refresh the view the same way. - layer.show offered cinematic_sky, which nothing reads since the layer was retired; the assistant could switch it to no effect. Unlisted layers are now refused and left out of the schema. - Named the key zoom rate and the pinch deadband. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TXuzbw772NsDthvWr8VnjE
Every finished tile scanned the whole request map under the admission lock, and the map kept every tile a source lacks for the session, so completions got slower as a session went on. Expired failures now leave in passes once the map has doubled, and past 100,000 remembered entries the missing-tile marks are forgotten together. retry_delay treats a failure past its retry delay as never asked, so the longer-lived entries cannot cause busy retries. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TXuzbw772NsDthvWr8VnjE
- Explorer::on_earth and App::earth_shown replace 22 copies of the body comparison; two now-unused Body imports go. - App::layer checks a refinement's parent directly instead of building the whole visible-layer set for it (same result, called many times a frame), and its doc comment is back on it. - Named the live-service heartbeat and the aircraft query radius factors. - One warmup_releases helper for the two tile warmups; the full-texture rectangle is built once; the marine label rectangles are no longer cloned every frame. - The theme conversion moved out of ui() into egui_theme. - model_http's single-use trait became redact_url. - fmt_dist/fmt_alt show 999.6 m as 1.00 km, not 1000 m. - English-only comment in math.rs; rustfmt over the files touched by the review fixes. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TXuzbw772NsDthvWr8VnjE
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TXuzbw772NsDthvWr8VnjE
The disk cache accepted only tile columns and rows below 2^zoom, the Web
Mercator grid. NASA GIBS polar grids (EPSG:3413, EPSG:3031) have two tiles a
side at level 0, so half or more of the polar tiles were refused and
downloaded again every session. The error used to be dropped; the new
write-failure report showed it in the run log ('invalid tile cache path').
Address::parse now uses each grid's own size for the range check and the
TMS row flip.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TXuzbw772NsDthvWr8VnjE
Diagnostics went to standard error only, which release builds on Windows and apps started from Finder or a desktop menu never show. A small std-only logger in data (run_log, log_line!) writes them to globecraft.log in the data folder as well, rotating a log over 1 MiB at startup and capping a session at 4 MiB. A panic hook logs every panic with its backtrace, including those the worker threads catch. All non-test eprintln! calls use it; README says where the log is and AGENTS.md asks for log_line!. main runs the EarthCraft migration before opening the log: the log creates the data folder, and an existing folder ends the migration. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TXuzbw772NsDthvWr8VnjE
Workers catch their panics on purpose (cpu.rs, render backend, assistant), but a panic while a std Mutex is held poisons it, and the next lock().unwrap() - often on the UI thread - panicked too, taking the interface down after the worker had recovered. The 97 non-test lock, read, write and Condvar wait unwraps now take the guard with PoisonError::into_inner. Tests keep unwrap. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TXuzbw772NsDthvWr8VnjE
…assistant - A globecraft://view link (clicked in a browser, pasted, or opened from the command line) could add and keep data layers from any address, including this computer and the local network. Such addresses are now skipped. - view.openLink adds a link's data through data.add internally, so the assistant could put web addresses in a link and skip the approval that data.add asks for. view.openLink is now a Network command and the gate looks at every address a call would fetch; allowing a call remembers its hosts. - The host parsing and local-host test moved from the assistant into globecraft_core::web_address so the app and the assistant share them; PRESET_PREFIX moved to view_link with ViewLink::web_addresses. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TXuzbw772NsDthvWr8VnjE
Admission counted queued tiles as well as running ones against the 48-slot limit. After a 429 or 503 a host pauses for up to two minutes with its tiles kept queued, so a throttled imagery host could fill every slot and stop elevation, night-light and polar tiles from loading. Tiles waiting for a paused host no longer count against other hosts (the paused host itself is still capped); retry_delay and prioritize_view use the same test. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TXuzbw772NsDthvWr8VnjE
…hout a build - app-build-test gains a Rust 1.95 job (cargo check, every target): the declared minimum was only exercised when a release was built. - New weekly dependency-audit workflow runs cargo audit on Cargo.lock. - On a system with no published download (an Intel Mac, for one) the update check said 'up to date' forever. A newer release is now reported as unsupported here, pointing to Releases. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TXuzbw772NsDthvWr8VnjE
settings.json, lang.txt, open-kml.json and the two key files lived in the cache folder (~/.cache/globecraft_tiles on macOS and Linux), which cache cleaners empty. They now live in paths::config_dir: Application Support on macOS, APPDATA on Windows, XDG_CONFIG_HOME (or ~/.config) elsewhere, or GLOBECRAFT_CONFIG_DIR. migrate_config moves them once at startup; a file already in the new place wins. Caches, the run log and the Wayback list stay where they were. A run that names only XDG_CACHE_HOME (as the screenshot tests do to isolate themselves) keeps its configuration beside that cache, so it cannot write the user's real settings. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TXuzbw772NsDthvWr8VnjE
globe_view was 777 lines of input, render scheduling, warmups, painting and the context menu. Contiguous blocks moved verbatim into methods: navigation_input, pump_warmups, paint_map_layers, paint_scale_bar, paint_live_map_credit and globe_context_menu. globe_view is now 298 lines and reads as the frame's order of work. No behaviour change: navigation_input recomputes the aspect and Shift from the same frame's rectangle and input. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TXuzbw772NsDthvWr8VnjE
Deterministic generated inputs (a small test-only xorshift generator in core and data, no new dependency) check that each parser returns or refuses without panicking: ISO times, coordinate queries, view links, web addresses, colours, MVT tiles, compact border tiles, PMTiles headers and directories, KML, and GeoJSON, GPX and CSV feature files. Binary and file formats start from valid samples and are damaged byte by byte. 10,000 to 50,000 cases each. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TXuzbw772NsDthvWr8VnjE
- help.copyDiagnostics copies the version, system, window and scale, the renderer setting, language, world, basemap, visible layers, the number of added layers, tile cache use and the last 40 lines of the run log, for a bug report. The camera position, data addresses and keys are left out. Listed in the Help menu; labels in both catalogs. - A test reads App::run and the dispatchers it hands prefixes to and fails when a catalog command has no match arm there (until now an unknown command only failed at run time). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TXuzbw772NsDthvWr8VnjE
…ery 250 ms The whole settings tree, user layers and style included, was compared on every frame. Animation frames now check at most four times a second, and quitting checks once more before the writer finishes (review P2). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TXuzbw772NsDthvWr8VnjE
next_job rebuilt per-source and per-host counts from every running download for each job; the counts now live with the set (review P5). prioritize_view keeps the reserved-set lock instead of cloning the set (review P4); reserve and release no longer skip a poisoned lock. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TXuzbw772NsDthvWr8VnjE
While on, every HTTP client in globecraft-data and the assistant's model client refuse with Error::Offline before connecting, and the update check waits. Tiles and data in the disk caches still show. The command is Interface reach so the assistant cannot turn its own network off (review I4). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TXuzbw772NsDthvWr8VnjE
An integration test renders every command (id, English name and help, parameters from its JSON schema, whether the assistant may run it) and fails when the stored reference differs (review I8). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TXuzbw772NsDthvWr8VnjE
Network commands are listed as asking first only when they name a web address, as the assistant's approval gate does. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TXuzbw772NsDthvWr8VnjE
Export writes every setting (added layers and saved sources included) as readable JSON; keys live in their own files and are never included. Import reads a file as a saved one is read, and refuses JSON that shares no field with the settings so the wrong file cannot reset everything (review I6). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TXuzbw772NsDthvWr8VnjE
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TXuzbw772NsDthvWr8VnjE
The camera reached the services as (lat, lon, alt, tilt) tuples and as four query structs repeating those fields; ViewQuery replaces them, with From<Cam>. The live-service block moves out of App::logic into app/live_services.rs with its constants, and one repainter(ctx) helper replaces the wake closures each call site built by hand (review A6). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TXuzbw772NsDthvWr8VnjE
App's nine live data services become one LiveFeeds field (app/live_services.rs), and the picked flight, satellite, What's here? card and context-menu point one Selection field (app/selection.rs). Field accesses are renamed only; behaviour is unchanged (review A2, first two groups). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TXuzbw772NsDthvWr8VnjE
The eight fields that load, merge and paint borders and names become one Borders field (app/borders.rs): loader, updates, log, source, drawn, place names and the overlay cache (review A2, third group). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TXuzbw772NsDthvWr8VnjE
The workspace's zip moves from 2.4 to 8.6, the version self_update uses, so one zip is built instead of two (review Q23). Debug builds optimise the workspace at level 1 and dependencies, built once, at level 3: rebuilds after an edit are quicker and tests stay fast (P6). The app you run is a release build and is unaffected. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TXuzbw772NsDthvWr8VnjE
… exists Settings › Keys gets "Keep keys in the system keychain" on macOS and Windows (keyring crate). Turning it on moves the service and assistant keys from their 0600 files into the Keychain or Credential Manager and removes the files; turning it off moves them back. It is off by default because macOS asks again before a differently signed build (every local rebuild) reads the entry (review I2). self_update's signatures feature is on: once crates/globecraft-data/release-signing.pub holds a zipsign public key, an update archive without a matching signature is refused. The release workflow signs the updater archives when the ZIPSIGN_PRIVATE_KEY secret is set and builds the checksums after signing. docs/RELEASING.md says how to make the key; until then nothing changes (review I3). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TXuzbw772NsDthvWr8VnjE
app/screenshots.rs (2,800 lines) keeps the harness and shared fixtures; the tests move to app/screenshots/: interface, assistant, map layers, added data, live symbols, tools and performance. settle() now returns once the application has been idle (nothing drawing, loading or flying) for half a second, instead of always sleeping the full time; the frame-time benchmarks keep their real-time pacing (review Q22). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TXuzbw772NsDthvWr8VnjE
globecraft-assistant becomes an optional dependency of the app behind the `assistant` feature. Without it the assistant pane shows a line saying the build has none, its two commands refuse, and the Settings › Network section for it is left out. CI checks the app without the updater and without the assistant, and the check script runs clippy for each (review I9, second subsystem). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TXuzbw772NsDthvWr8VnjE
SettingsInput, a second declaration of every top-level setting kept only to deserialize it, is gone. apply_json reads the file's object through a Fields helper (text, number, flag, value), with the same rule as before: a field of the wrong type keeps the current value. Adding a setting now takes the field, its default and one line (review A5); the unused json_fields::number and boolean helpers go with it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TXuzbw772NsDthvWr8VnjE
Every successful load writes settings.json.bak. A settings file that cannot be read is still set aside as settings.json.corrupt-<time>, but the copy from the last start is used instead of the defaults when it reads, and the status says so (review B3, third step). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TXuzbw772NsDthvWr8VnjE
…g the style places.remove, data.remove and style.reset have no undo (Edit › Undo covers the ruler only), so they are Destructive: the assistant shows its approval card first, as for other commands that discard work (review B2, related). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TXuzbw772NsDthvWr8VnjE
release-signing.pub now holds the zipsign public key made on the maintainer's Mac (the private key stays there, outside the repository), so builds from here on refuse an update archive without a matching signature. The release workflow installs zipsign 0.2, the signature format the app's zipsign-api reads, and fails rather than publishing unsigned updates when the ZIPSIGN_PRIVATE_KEY secret is missing (review I3). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TXuzbw772NsDthvWr8VnjE
Command ids are still strings, so a mistyped id in a button, menu item, shortcut or run call compiled and failed only when used. A test scans the app's source for the ids passed to the command functions (about 230) and fails on any the catalog does not list (review A4). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TXuzbw772NsDthvWr8VnjE
…y fuzzing) Entry points for fuzz targets (`globecraft_data::fuzzing`, behind the new `fuzzing` feature) feed every decoder of untrusted data as the application does. Their first runs found five defects, fixed here with unit tests: - PMTiles: a tile id for zoom 32 or more overflowed its shift; directory tile ids and offsets, and the header's root, tile and leaf offsets, were added unchecked; a header with reversed zoom levels made `clamp` panic; a root directory of any claimed length was read into memory. Headers with impossible zoom levels are refused, sums are checked, a range read is at most 16 MiB. - MBTiles: `(1 << zoom) - 1 - row` underflowed for a row outside its zoom level and overflowed for zoom 32 or more. Only rows that can exist count; the lowest level with such a tile is the sample. - Stored border tiles: coordinates are differences, and a damaged row's sums overflowed. A coordinate outside the tile is damage. - Images: a header may claim any size. Tiles and layer images decode up to 4096 x 4096, world pictures up to 16 384 x 16 384 within 512 MiB. - gzip: tiles, PMTiles directories, the time zone, sea name and airport files and flight traces unpack to a named limit. Both limits live in the new `bounded` module. To reach the decoders without a file or a network, `kmz::decode_from`, `Header::root_in` / `tile_location` / `leaf_location`, `archive::pmtiles_info` / `mbtiles_info` / `mbtiles_tile` and `Airports::from_gzip` were split out of their callers. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TXuzbw772NsDthvWr8VnjE
`fuzz/` is a cargo-fuzz crate (its own workspace, never built by a normal `cargo build`) with 15 targets: KML and KMZ, GeoJSON/GPX/CSV features, WMS capabilities, PMTiles, MBTiles rows, vector tiles, stored border tiles, image tiles (Terrarium included), planet files, the JSON feeds, the gzip readers, settings, typed text and assistant replies. Several assert a property besides "does not crash": a stored tile reads back as written, settings survive an export and import, decoded positions are in range, a PMTiles tile id lies in its zoom level's block, an MBTiles sample tile can be read. `fuzz/run.sh [seconds] [target ...]` builds and runs them with a 10 s and 256 MiB limit per input; `fuzz/seeds/` holds small valid inputs (written by `make_seeds.py`) and the inputs of the five defects fixed in the commit before. The `fuzz` workflow runs every target weekly for five minutes, and for 30 seconds when the targets or their entry points change. CONTRIBUTING.md gains the two rules the findings taught (checked arithmetic on numbers from a file; decode through `bounded`), fuzz/README.md the rest. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TXuzbw772NsDthvWr8VnjE
main gained the full split of App into feature states (roadmap 10.2), one command handler per feature, PMTiles files on this computer, places.export and its own review fixes. Where both sides touched the same code: - App state and commands: main's structure wins. This branch's partial groupings (live_services, selection, imagery, its Borders, its handlers per prefix) are dropped; its features are carried into main's states and handlers: Copy Diagnostics, Work Offline, Export and Import Settings, the `assistant` and `updater` build features, the settings check every 250 ms (now `SavedSettings::checked`), poisoned-lock recovery, the run log. - The split of globe_view into named steps is kept, with main's field paths. - The live services keep this branch's ViewQuery: main's LiveState and MapDataState pass it on. - Borders keep the place-name snapshot itself, not its address. - PMTiles and MBTiles: both sides hardened the same arithmetic; this branch's version stays (checked sums, `bounded` limits, MAX_ZOOM 31), with main's local files and vector-tile refusal. - Keys in error messages: `keys::redact_error` stays; main's `without_address` had no callers left. - One `paths::write_atomic` (with the `private` flag). - The screenshot tests stay split by subject, with main's field paths. - docs/COMMANDS.md regenerated (places.export). Checked on Linux: fmt, clippy -D warnings in four feature sets, and the tests of core, data, render and assistant. The app is only type-checked there (against stub GTK descriptions); its tests and the macOS-only code run on the Mac. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TXuzbw772NsDthvWr8VnjE
`live` (on by default) covers aircraft, vessels, satellites, the aurora forecast, traffic, clouds and radar. A build without it: - never starts a download for them: every place in the data crate that would start one asks `live::ENABLED` first, and no aircraft worker runs. The feeds' types and decoders stay, since the renderer and the interface are written against them; - does not offer their layers (`MapLayer::is_feed`, `App::offers`): they are absent from the Layers pane, the ribbon's Live tab and the command search, `layer.show` refuses them, and a settings file or view link that names them shows nothing. Tests that need a worker are marked `live`; two tests that run only without the feature check that nothing is asked for. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TXuzbw772NsDthvWr8VnjE
Release downloads, the Flatpak, the Snap and the window smoke test build with `--no-default-features` to leave out the experimental GPU renderer. Since the updater and the assistant became default features on this branch, that also left them out, and would now leave out the live feeds too. `standard` names everything a release has (updater, assistant, live); the default is `gpu` plus `standard`, and every release build now says `--no-default-features --features standard`: the release and smoke workflows, the release-configuration check, both packaging manifests (with their test) and the build instructions in RELEASING.md, GPU_RENDERER.md and CONTRIBUTING.md. CI also checks the app with each optional part left out. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TXuzbw772NsDthvWr8VnjE
A setting was declared four times: field, default and two lines of hand-written reading per section. `settings::lenient::apply` now reads a struct's plain fields through its derived `Deserialize`: the struct as it stands is serialized, the file's values go over it, and the result is read back; if a value does not read, the fields are tried one at a time, so a bad value costs only its own field. A new setting is a field and its default. Sections and tables keep an `apply` their owner calls, and fields with rules a type cannot say are named in `own_rules` and read as before (the startup position, filtered lists, the ribbon tab's older names, cache sizes). Ranges are a clamp after the read. `coded` and `applied` read back fields saved under a short name or by a reader of their own. No file reads differently: the old and new readers gave the same result on 13,894 settings files (the fuzz corpus, and one odd value at each of 163 places of a full file). New tests hold that settings read back exactly what they save, which the reader relies on. See docs/plans/2026-11-15-settings-through-serde.md. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TXuzbw772NsDthvWr8VnjE
… lights With default settings no frame ever reached the GPU renderer: it refused the WGS 84 ellipsoid (the default Earth model) and any view of a polar cap under satellite imagery, and its parity test had failed since the polar caps landed. - The shader intersects the ellipsoid (a sphere is the ellipsoid whose axes are all one) and uses its normal for latitude, level of detail and light. - Polar cap pixels are left to the precise CPU repair, as the sky is. - Night lights are sampled on the GPU: their tiles share the atlas and the tile table (tagged), with their own block of tile needs. - The parity test covers both Earth models, the caps' edge and night lights; the frame benchmark uses the default model. Software Vulkan (lavapipe): 7.3 million channels, none more than one level off; 2000x1488 with real imagery: one channel two levels off. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01KenNVpGqarXCfD9mvMtRLQ
…ards - Self-check: before its first frame each device draws a small view and is refused when it differs from the CPU renderer's picture. - Start marker (`gpu-start` in the settings folder): a start that crashed or was killed is not repeated; turning the setting off and on, or a new version, tries again. - A frame the device does not finish within 5 s counts as a device failure. - A failed device is replaced after 2 s, at most three times per session. - Settings -> Graphics shows the toggle on every platform, with the adapter in use or the reason the CPU draws; the diagnostics count the frames each renderer drew. - The parity workflow also runs when the renderer changes, and includes the device-loss test. Release downloads still exclude the GPU renderer. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01KenNVpGqarXCfD9mvMtRLQ
`App::run` matched command ids as strings in fourteen handlers, each reading its parameters by name from JSON: a mistyped id or parameter compiled and failed when used. `globecraft_core::commands::Command::parse` now reads an id and its parameters once. The id becomes a variant, grouped by feature (`Command::Camera(CameraCommand::FlyTo(..))`), and the parameters a struct declared with `params!` (`commands/params.rs`): numbers also as text, text trimmed, a switch of the wrong type ignored, as the handlers' helpers did. Each feature's handler matches its group, and must name every command in it: a command without a handler no longer compiles. One table in `commands/command.rs` pairs ids with variants. The JSON Schema text stays hand-written, since it is what the assistant reads; core tests check that the table holds every catalogue command once, that each parameter struct has exactly its schema's parameters (requiring none the schema does not), and that every command parses what its schema describes. Found on the way: `ui.timeline` read a parameter `visible`, while its schema (and so the assistant) says `on`; it now reads `on`, and the timeline's close button sends `on`. Other small differences: a number given as text is accepted by every command, a wrong-typed optional number (`style.line`'s width, Wayback's split) is now an error instead of being ignored, and a missing parameter is reported as "<name> is required". `command_state` (enabled/checked for buttons) is still keyed by id, as are the widgets' calls; the interface-id test covers those. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TXuzbw772NsDthvWr8VnjE
Checking all 52 branch commits against the merged tree found a few review fixes that main's restructured states had reverted: - The merged flight list keeps the list it was made from (an Arc), not its address, so a freed list whose memory is reused can no longer be mistaken for the current one. - Named structs again instead of tuples: HereAdmin (pointer card, border revision, areas), Frame (width, height, pixels) and DownloadLog. - One map of typed API keys with get/edit (review Q8) instead of one field per provider. - LIVE_SERVICE_HEARTBEAT names the 30 s repaint for live services. - Explorer::on_earth used in the command handlers again. Everything else from the 52 commits is present, moved or superseded by a later commit (A4, A5, I9, fuzzing). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TXuzbw772NsDthvWr8VnjE
… (8.3, 8.4) Single scattering of sunlight by air (Rayleigh), haze (Mie) and ozone absorption, from a transmittance table and a four-dimensional scattering table computed once at first use (Bruneton's 2017 parameters). It replaces the blue rim and the 7 % halo, which ignored the sun. - The limb is a thin blue line by day, red at the terminator and dark at night; the ground fades into haze toward the limb, and relief with distance; inside the air the sky has its day colour, its sunset and its night, and the stars go out by day. - Sunlight that crossed much air turns the ground orange toward the terminator and fades to the night's own light (8.4). - From space the ground uses the table's top layer; the GPU shader does the same sums on the same numbers (lavapipe: no channel more than one level off, also with real imagery at 2000x1488). - Inside the air each frame narrows the tables to the camera's height and sun once, so a pixel reads a few entries instead of dozens. - Straight down only 30 % of the air is drawn (the imagery is corrected for it already), rising to all of it at the limb. - The raster fixtures with the atmosphere on were recorded again (GLOBECRAFT_RECORD_FIXTURES); `atmosphere_stills` writes pictures with real imagery for judging by eye. - The frame benchmark allows slow runner disks 120 s to load its tiles. Cost on the CPU: about 7 % at 2000x1488 from 6,500 km; more in close tilted views. Light scattered more than once is left out. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01KenNVpGqarXCfD9mvMtRLQ
The Windows runner's disk needs more than 20 s for the detail tiles too. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01KenNVpGqarXCfD9mvMtRLQ
On Windows the source lines end with CR LF, so every call whose id is on the next line was skipped and the scan found 184 ids instead of over 200. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01KenNVpGqarXCfD9mvMtRLQ
On Windows the checkout turned every line ending into CR LF: the command reference no longer matched docs/COMMANDS.md and source scans missed calls, so the Windows tests failed. Binary fixtures, keys and fuzzing seeds are marked as such. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01KenNVpGqarXCfD9mvMtRLQ
On Windows the source lines end with CR LF, so every call whose id is on the next line was skipped and the scan found 184 ids instead of over 200. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01KenNVpGqarXCfD9mvMtRLQ
On Windows the checkout turned every line ending into CR LF: the command reference no longer matched docs/COMMANDS.md and source scans missed calls, so the Windows tests failed. Binary fixtures, keys and fuzzing seeds are marked as such. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01KenNVpGqarXCfD9mvMtRLQ
.gitattributes decides what a Windows checkout compiles and compares, so a change to it must run the tests. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01KenNVpGqarXCfD9mvMtRLQ
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01KenNVpGqarXCfD9mvMtRLQ
On by default wherever the gpu feature is built in, with the self-check and the start marker; release downloads still leave it out. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01KenNVpGqarXCfD9mvMtRLQ
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Builds on #4 (
review-fixes); merge that first, then this one shows only its own commits.Why
With default settings no frame ever reached the GPU renderer. It refused the WGS 84 ellipsoid (the default Earth model) and every view of a polar cap under satellite imagery. Its parity test had failed since the polar caps landed, and nobody saw because the test is ignored and its workflow was manual.
What changes
GPU renderer (roadmap 8.2, 14.1)
gpufeature is built in. Settings → Graphics shows the adapter in use, or why the CPU draws. The diagnostics count the frames each renderer drew.standard).gpu-parity.ymlruns on every change tocrates/globecraft-render.Atmosphere (8.3, 8.4)
Windows (also pushed to #4): the command-id scan accepts CR LF, and
.gitattributeskeeps LF line endings.Checked
app-build-testis green on Linux, macOS and Windows.Not checked
dependency-auditfails on every branch with three RUSTSEC advisories. This PR does not touch them.Generated by Claude Code