CydandClaude Opus 4.8
4c15197b34
RIOJoy lifecycle is now owned by the console + launcher (the auto-start Run entry was removed on the RioJoy side): Apps.xml autoRestart false -> true, so the launcher starts RIOJoy and keeps it alive. Safe with uninstall — the watcher only relaunches a still-tracked process, and uninstall's kill untracks it first. Uninstall now physically removes the product (was unregister-only): - Agent CmdUninstallApp: after kill + unregister, reports the product's C:\Games\<dir> to clean up — but only if no remaining registered launch entry still uses that folder (orphan check keeps multi-entry products like Red Planet's GameClient/LC/MR intact until the last entry is removed). - Service (SYSTEM): on UninstallApp, runs <dir>\pre-uninstall.bat if present (RIOJoy's removes ViGEmBus + config) then deletes the directory, on a background thread so it can't trip the Console's RPC timeout. Path-guarded to only ever delete a proper subdirectory of C:\Games. Logged to podconf.log. 74 tests green; the cleanup path is pod-behavior, not unit-tested here. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
TeslaConsole.DiffTests — differential equivalence suite
Verifies that the reconstructed TeslaConsole.exe (built from the decompiled
source in this repo) behaves identically to the original reference binary in
original/TeslaConsole.exe.
How it works
The suite loads each assembly into its own child AppDomain
(DifferentialFixture) and drives it through a MarshalByRefObject proxy
(Invoker). This is why the project targets net48 — AppDomains are a .NET
Framework feature. The original is the 4.11.3.37076 baseline; the recovered
build is the modernized 4.11.4.x line (same TeslaConsole assembly name, an
intentionally newer version). Because the two versions differ, the public-member
comparison strips Version= stamps before diffing — it compares type/member names,
not assembly versions.
Each child domain is given a probe directory (the recovered build's output, which ships every dependency DLL) so the original — which is distributed without its proprietary dependencies — still resolves its references for metadata inspection.
What is compared
-
Public API surface (
PublicApiSurfaceTests) Every public type and public member (signature-for-signature) exposed by the original must also be exposed by the recovered build. Compiler-generated members and property/event accessor methods are excluded — the README at the repo root notes those legitimately differ between a decompilation and the lost sources. -
Behavioral output (
BehavioralEquivalenceTests) The same deterministic, dependency-free methods are invoked in both assemblies over a battery of inputs and the results must match byte-for-byte:RPStrings.GetTimeString(mm:ss formatting + 0.5 s rounding)HostTypeHelper.Parse(...).ToString()(incl. invalid-input exceptions)PlasmaBitmaps.ConvertBitmap(1-bpp packing of a known pixel pattern)PlasmaBitmaps.GenerateString(full GDI text → 1-bpp plasma pipeline)RPMap/RPVehicleXML parsingSiteManagementwell-known application GUID constantsTuple.Create<,>generic factory
A negative-control test (
Harness_Distinguishes_Different_Outputs) proves the harness can actually see a difference, so a green run is never vacuous.
The project also carries two byte-compatibility guards — not original-vs-recovered comparisons, but checks that the modernized protocol/crypto stays compatible with the original binaries:
PodRpcProtocolTests— round-trips the framed-JSON RPC (Contract/PodRpcProtocol.cs) in-process: every request/response shape encodes and decodes back to the same values.SecureConfigCompatTests— asserts the source-builtOFBCryptoStreamproduces byte-identical ciphertext to the originalTeslaSecureConfiguration.dll, so the pod provisioning handshake stays wire-compatible.
Running
dotnet test tests/TeslaConsole.DiffTests/TeslaConsole.DiffTests.csproj
A project reference builds the reconstruction first, and the suite always tests
the most recently built bin/{Debug,Release}/net48/TeslaConsole.exe.
Scope / limitations
This compares deterministic logic. It deliberately does not drive the WinForms UI, the pod networking, secure-configuration, or hardware-facing code — those require the live console, its pods, and the proprietary services, and are not reproducible in a unit test. The API-surface test still asserts those types exist with matching signatures even though their behavior isn't exercised.