* build: the main line on .NET 11 - SDK 11.0.100, net11.0, the three runtime-serviced packages at 11.0.0; Windows 10 stated as untested and unsupported (D94, D95) (beta.37 PR-1) The owner moved the main line to .NET 11 the day after it shipped (D94), setting aside condition (a) of 20_OPEN_QUESTIONS section G's table - two alphas on the released runtime. This is the alpha branch's switch applied to main at released versions, and what follows from it. - global.json 11.0.100 (allowPrerelease false), TFM net11.0-windows10.0.22621.0, Microsoft.Extensions.Hosting, Microsoft.Data.Sqlite and System.Diagnostics.EventLog 11.0.0; legal/licenses/nuget regenerated (78 packages -> 69). - Two CA1307 sites the released SDK raised in tests written after the last alpha was cut take the (char, StringComparison) overload. - IHost.RunAsync rethrows a failed BackgroundService on .NET 11: the Agent's --serve leaves through its crash report where it exited 0. Accepted, not wrapped; HostBackgroundFailureTests holds the runtime's behaviour, which nothing had observed. - release.yml installs a .NET 10 runtime for vpk 1.2.161; Qodana's container installs the .NET 11 SDK first (interim, until the Ultimate linter); Dependabot's and the workflows' comments follow. - Windows 10 22H2 is not blocked, and is stated as untested and unsupported (D95): NFR-8, LIMITATIONS, README, the guide, the wiki. A processor needs SSE4.2 and POPCNT. - docs: 12_BUILD (the main line's .NET; the alpha line rests after alpha.4, D98), 13_CI_CD, 20_OPEN_QUESTIONS row closed, 10_LOGGING, spike-notes section 28, CLAUDE.md, the decisions index (D94-D96, D98). Gate on the dev box: all passed in 338 s on SDK 11.0.100 (native 32/32, 2,561 tests, coverage floors held). Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> * build: what this pull request's runs measured on .NET 11 - CodeQL reads net11.0 whole, the delta from beta.36, Qodana's one finding fixed - HostBackgroundFailureTests: the message is kept rather than the exception, so no `!` is needed - the one problem Qodana's interim run reported (run 42916491047, InspectCode 2026.2.3.1 on SDK 11.0.100). - 13_CI_CD: what analyze (csharp, none) read of net11.0 (CodeQL 2.27.4, "scanned 1220 out of 1220 C# files", 363 references resolved, none unresolved), and the rule for C# 15 constructs (D96); the interim's first run. - spike-notes section 28: the table of this pull request's runs - check, CodeQL, SonarCloud, Qodana, and the release rehearsal with the first delta across two runtime majors (72.6 MB, 372 of 443 files patched). Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> --------- Co-authored-by: Claude Fable 5.2 <noreply@anthropic.com>
28 KiB
14 — Testing
Managed: xUnit.net v3 4.x (the xunit.v3.mtp-off variant since 2026-09-21 — build.ps1 runs the suites under VSTest for the coverage collector and results.trx; Directory.Packages.props says why) + FluentAssertions 7.x + NSubstitute. Native: Catch2. Coverage goal ≥ 80% on Domain + Application; Domain metric calculators ≥ 95% or the PR fails.
Safety-guard tests (highest priority — these protect users, not code quality)
The anti-cheat guard is the one component where a bug can cost someone an account. It gets the most rigorous treatment in the codebase:
These are Catch2 tests, not xUnit. 20_OPEN_QUESTIONS §S13(a) settled the guard into the C++ FrameLedger.Injector, so the fail-closed matrix below lives in the native suite and the guard needs seams — injectable enumerator function pointers — before any of these failures can be forced at all. Landing Catch2 is a prerequisite of the guard, not a later nicety (§S15). Landed long since — ctest runs 31 32 native suites on every gate (fl_process_stats since beta.12 — corrected 2026-10-04) (fl_guard alone is ~100 s); the sentence was left in the future tense until 2026-09-21.
- Blocklist matching: exact, case-insensitive, prefix rules; every family in
19_SAFETY§Blocklist seed has a fixture.Two families (Activision Ricochet, Valve VAC) have no data yet, so their fixtures must assert that absence explicitly rather than quietly passing on an empty rule set.(Corrected 2026-10-04: both have data since 2026-09-25 — Ricochet a driver and services, VAC per-title lists of Steam ids and executables, whichrules_budget_test.cppmatches back to "Valve VAC".) Since 2026-09-25 the fixture is generated from the shipped seed (rules_budget_test.cpp: every entry's every value matches its own family in its own group, which also catches a row an earlier prefix shadows), and Ricochet's fixture asserts its data; VAC's still asserts that no family entry fakes it. - A finding turns the game's hooking off (2026-09-22), and the proofs are about WHICH findings.
AntiCheatVerdictTestspinsIsFindingAboutTheGameper reason (in the game: yes; machine-wide driver or service, heuristics, failed scans, the latch: no) and that a finding without a family is not one;CaptureSessionTeststhat a start-timeBlockedModulewrites the block, keeps the consent stamp and saysHookingTurnedOff, that aBlockedDriverrefuses the session and leaves the row alone, and that the 30 s re-scan's finding writes the same block;RecordedSessionEventsTestsandSafetyNoticesTeststhat the flag travels to the notice. The override of 2026-09-21 had its own proofs for one day; they went with it, andNoSecondMatcherTestscounts the port's four methods again. - Fail-closed proofs: malformed
detection-rules.json, missinganticheatblock, unreadable target process,EnumProcessModulesExfailure, partial module list → all must refuse injection, never allow. Each is a named test. - Measured error paths that must be named tests (
spike-notes.md§1):ERROR_PARTIAL_COPY (299)from a suspended target,ERROR_ACCESS_DENIED (5)from a protected one, a driver-list parse whose paths are not native paths, and a service query returningACCESS_DENIED— the last of which cannot be produced on real services by a standard user, so it exists only as a fake. Every one means REFUSE. - Mid-session detection: simulated late-loading anti-cheat module →
unhookRequestedset, hooks disabled within one frame, session finalizedunhooked_safety. Cover the driver scan too, not only modules: a machine-wide driver can start after injection. - Absence-of-override test: no code path may reach the injection primitive without a passing guard result. The mechanism previously described here — a
sealedtoken type "only the guard can produce" — does not work and was disproved by compiling (20_OPEN_QUESTIONS§S8: C# accessibility flows inward,CS0122). The replacement is structural: the guard owns the chokepoint and calls the primitive itself, the primitive has internal linkage in the guard's own translation unit so no other TU has a symbol to call, and a build-time check asserts that nothing else references it. A token that escapes can be ignored; a symbol that does not exist cannot. - Static pre-scan: a game directory containing
EasyAntiCheat/rendershook_enabledun-settable; API-level attempt to set it is rejected — including viaSetWatchlist, which no longer carries hooking state at all (07_IPC§The pipe is not a trust boundary).The toggle half needs persistence and is not built yetbuilt in P2/P3:SqliteGameConsentStore.RecordGuardBlockAsyncforceshook_enabled = 0,AgentCommandHandler.SetHookEnabledAsyncre-runs the pre-scan and answersRefused, andGameDetailViewModeldisables the toggle on a blocked row (corrected 2026-09-21; since 2026-09-22 a finding at a session's start or at the re-scan writes the same block —CaptureSessionTests); the scan itself is ctestfl_prescan, whose matrix is mostly fail-closed: a listing that FAILED, one that was TRUNCATED, a directory that could not be established, unusable rules, and a missing evidence source must each be undetermined, never clean. Two cases exist only to stop the rest passing vacuously — a genuinely clean directory must be allowed, and a name matching the wrong group (a directory named like afilesentry) must not fire. One case removes a family from the rules and asserts the hit disappears, which is what proves there is a single matcher rather than a second one that happens to agree.
D48 (beta.14): the module check's one re-read is held both ways in ctest
fl_guard[D48]— a member still loading reads clean on the re-read (one pause, two snapshots); one still unreadable refuses; one that exited drops out of the fresh set; a target that exited refuses;ACCESS_DENIEDis never retried; a finding refuses at once and a finding the re-read sees refuses — and against the real source ([live]: a process created suspended and a pid with no process answerkRetry). The managed half — a refused re-scan of a game that has exited ends as its exit, never for a finding — isCaptureSessionTestsandSessionRecorderTests.
Native unit tests (Catch2)
- Ring buffer: SPSC correctness under a hammering writer thread; seqlock torn-read detection; wrap and overwrite-oldest semantics; power-of-two capacity assumptions. Include the one-full-lap case: a reader stalled exactly
capacityrecords must not validate a different frame as unchanged (this is what the never-resetseqcounter defends,07_IPC§Protocol rules). - Drop accounting: the reader-side computation (
writeIndex - readIndex > capacity) reports the right count and resumes at the right index. The writer has no drop counter to test — by design. - Torn records produce gap markers, not silently shortened streams; a golden test asserts the gap does not appear as a stutter.
- Record and header layout:
static_assert(sizeof(...) == 64)andoffsetofassertions forFlFrameRecordand allthreefour header structs (FlShmHandshake,FlWriterState,FlControlBlock, andFlDisplayStatesince layout 4, beta.10) — plus D33'sFlTolerance, mirrored the same way (corrected 2026-10-04) — the Agent reads across the process boundary in every one of them. Both halves exist as of 2026-08-05: the native side islayout_test.cppplus thestatic_asserts (ctestfl_shm_layout), and the runtime offset dump fromtools/fl-layout-dumpis consumed byShmLayoutMirrorTests, which walks the field list in both directions and additionally asserts blittability —Marshal.SizeOfand every offset stay correct for a mirror that uses[MarshalAs(ByValTStr)] string, whileUnsafe.SizeOfdrops from 64 to 40 and the struct can no longer be read from a mapped view. - Fault policy: injected SEH faults in a fake hook body → counter increments, original still called, self-disable at 3.
- Hook index verification, by behaviour. A vtable slot carries no identity, so "check that slot 8 is
Present" is not a question the runtime can answer — the earlier wording here required something unimplementable. What is answerable: patch the slot, call the method, see whether the detour ran (17_HOOK_ENGINE§Getting vtable addresses; proved for slots 8, 13 and 22 inspike-notes.md§H4). Runs headless on CI via WARP andCreateSwapChainForComposition. A deliberate mismatch aborts installation instead of patching. - Every probe must be shown to fail. A probe that ends in an unconditional assertion is green by construction and its ctest is decorative —
fl_proxy_swapchainshipped in exactly that state. Before a probe counts as a regression net, break the thing it watches and watch it go red.
Managed unit tests
Golden metric tests (write first)
- Hand-computed fixtures: time-based avg (not mean-of-FPS), median, p99/p99.9 with linear interpolation (include cases where interpolation and nearest-rank differ), min/max, σ, stutter rule, sufficiency guards at 999/1,000 and 9,999/10,000 frames.
FG ladder: records with API-sourced FG →Rewritten 2026-09-09 (P2 PR-A), because two of its rungs no longer exist and its last clause was the affirmative negativefg_source = api; present-count delta only →presentdelta; ETWFrameType→etw; cadence ≥ 1.5 →cadence; ambiguous →none. Factor always = Displayed/Native regardless of rung.03_METRICSforbids.presentdeltawas retired as structurally zero andetwwith Tier 2 (2026-08-28); "ambiguous →none" is exactly what rung 0 exists to prevent. The goldens now are03_METRICS§Frame Generation's:FL_MEASURED_FG_COUNTSclear →N/A, nevernone; a counted ratio ≤ 1.05 over a uniform window →none(reachable by COUNTING only); 1.05–1.5 → refused; a saturated count (255) → refused; a non-uniform window → refused rather than averaged; DXGI-counted Displayed may stand in the trio and inpresents/batchand never in a frame-time distribution. Each refusal is aFgRefusalKind, andFgWindowTests/DxgiCountedWindowTestsdrive every one.- Upscaling: exact ratio from render/output pairs; segment splitting when resolution changes mid-stream;
settings_changed_midsessionflag; dominant-segment selection. - RT: AS-builds-only stream (inline RayQuery case) →
rt = Yes— the specific regression the old design got wrong; dispatch-only →Yes; RT-capable device with neither →No; D3D11 →N/A. PT confidence scoring never yieldsYes. - Tri-state precedence:
manual > measured > inherited > N/A. - Tier gating: there is no Tier-2 record stream. A Tier-2 session produces
N/Afor every measured field — never a fabricated value, and never a zero standing in for one.
Parsers & infrastructure
- Shm reader against a synthetic mapping: version mismatch → refuse; torn records skipped; dropped counter propagated.
PresentMon CSV (Tier 2): header-map building with shuffled/unknown/missing columns, malformed lines, invariant decimal. Fixtures recorded from real runs.Removed 2026-08-27 with the tool. Whatever Tier 2 ends up parsing needs its own row, written against the format it actually reads — and the lesson from the one parser that did get written is worth carrying: its CSV's first column was namedApplicationwhile the value it counted was alsoApplication, so resolving by NAME rather than position was load-bearing (spike-notes§11).- Steam ACF (KeyValues), GOG
.info, Epic.item, itch receipt — real-world fixtures. Since P4 PR-4 (2026-09-14): the import's readers are driven by fixtures the tests write themselves (ValveKeyValuesTests,StoreLibrarySourcesTests— Steam's two-file shape across two library folders, GOG under a test HKCU key, Epic.itemJSON, an itch receipt gzipped in the test), not by committed real-world files; a real-world corpus is still worth adding the day a launcher's format drifts. RuleEvaluator: fixture directory trees per engine, including the "Unity markers present but UE structure too" ordering case.- Blob codecs round-trip for every series in
frame_blobs(NaN forbidden — assert), including the two-pairrender_resencoding and the three-bitrt_flagsbyte. SQLite migrations apply cleanly from an empty file to the current schema, and re-applying is a no-op. (There is no v1→v2 upgrade to test —06_DATA_MODEL§Migrations.) - IPC pipe: framing, split reads, oversize rejection, unknown fields/types ignored, protocol mismatch.
What the App draws (beta.14)
A page or dialog that only loads has proved nothing about how it looks: PagesLoadTests counted the import list's
rows while every row was white text on white in the dark theme (08_UI §Contrast). The rendered checks run on the one
STA thread with the App's three dictionaries (PagesLoadTests.OnStaAsync) and look at the result: Visuals lays a
control out at a window's size, pumps the dispatcher (a NavigationViewContentPresenter's navigation), swaps the theme
dictionary for one check and puts Dark back, and composites every Border and Panel background between a text and the
window into the colour actually painted under it; Wcag is the arithmetic. ImportReviewRenderTests measures each text
cell against that colour in both themes and counts the accent pixels of each ticked box in a RenderTargetBitmap; the
paging checks navigate a real page in WPF UI's presenter and send it the wheel. Each was run red against the unfixed XAML
first — a rendered check that cannot fail is the counted-rows check again.
A theme switch needs a window (beta.17). A change to Application.Resources reaches only the trees of
Application.Windows, and a Window registers itself there when it is constructed, shown or not. So
PagesLoadTests.EveryTextFollowsAThemeSwitch hosts each surface in a FluentWindow it never shows, swaps the theme
dictionary as ApplicationThemeManager.Apply does, and compares every text's colour with the same surface made fresh
under the new theme. PagesLoadTests.Surfaces builds every page, window and dialog content from loaded view models, for
any check that wants all of them. Colours are compared and texts only reported: another test class may change
Strings.Culture between the two builds.
A page is measured in the Frame the App shows it in, and a theme brings its accent (beta.18). WPF UI's
NavigationViewContentPresenter is a Frame, which passes no inherited value into the page, the text colour included, so
a page laid out straight in a window measures the window's colour and not the App's: the theme-switch check above did, and
stayed green over every page title drawn black (08_UI §Contrast, fifth hole). The rendered checks host a page in the
presenter (PagesLoadTests.Hosted), and PagesLoadTests.EveryTextIsReadableInBothThemes measures every surface's shown
texts at ≥ 3:1 in both themes. Visuals.UnderTheme switches as the App does — the accent resources derived for the theme
from Windows' default accent (Visuals.DefaultAccent, the App's derivation rather than the machine's own palette), then
the dictionary — and puts the application's own resources back afterwards. The accent lives in Application.Resources,
not in a theme dictionary; until then a check under Light measured the dark theme's link blue on the light window, 1.9:1,
a colour the App never shows. The theme-switch check compares what can be seen: a text only one of the two instances holds
counts where it shows, and a collapsed part is not compared (a switch re-applies the template of a DataGrid's idle scroll
bar, and its glyphs exist in the switched instance only).
Loaded needs a shown window, and even then it is not raised everywhere (beta.17). A tree laid out by hand never raises
Loaded, and a shown one raises it only where a subtree listens for it (16_WPFUI_SYNTAX §Gotchas) — so a test that calls
RaiseEvent(Loaded) proves a handler, never that WPF reaches it. ScrollBarHidingTests.EverySurfaceShownWithTheOptionOnDrawsNoBar
shows each surface in a real window, off the screen and never activated, with the option on before anything is built as
at start-up; WpfRaisesLoadedOnlyWhereASubtreeListensForIt checks the harness against WPF itself (a scroller holding a
Border is never loaded, one holding a CheckBox is), on CI too.
A test that sends input must also say which keys are held. Keyboard.Modifiers is the machine's real keyboard, read
from any thread (a fresh background thread's GetKeyState reports the machine's NumLock — measured 2026-10-06), so a
wheel test that read it held whatever key a person pressed while the gate ran: ChartWheelTests' plain half went red once
on the developer's machine and never on CI. The handlers read ChartWheel.CurrentModifiers, which the test sets inside its
own work item on the STA thread and puts back — and with the keys a parameter, the Ctrl half is driven end to end too.
Integration tests (CI-runnable, no game, no anti-cheat surface)
hook-harness is what makes this architecture testable without touching a real title:
- Harness presents 120 s at a controlled cadence → Overlay injected → ring drained → session written → assert every aggregate against expected values.
- Harness simulates: resolution change mid-run (segments), stub upscaler exports (
upscalerdetection), RT PSO + dispatch (RT flags), AS builds without dispatch (inline RayQuery path), PSO compile spikes (stutter attribution), and a fault-triggering hook path (self-disable). .partialrecovery: kill the Agent mid-capture, restart, assertinterruptedsession recovered. Built 2026-09-10 (P2 PR-G):PartialRecoveryEndToEndTests— the shippedFrameLedger.Agent.execaptureshook-harnesswith--partial-flush-seconds 2, the test polls until a flush has actually landed (never a sleep),Process.Kills the Agent, and then--console recoverturns the valid prefix into aninterruptedrow whoseframe_countIS the prefix, deletes the file, and finds nothing on a second pass.PartialSessionFileTestsalready kills at every byte offset; what this adds is that the Agent wrote a recoverable prefix on its own before it died.- Test hygiene: the harness's overlay logs leave with the run. The Overlay logs to the real
%LOCALAPPDATA%\FrameLedger\logsby design, so every test binary that starts hook-harness removes the logs it caused when its run ends, andtools/test-artifacts-check.ps1fails the gate for one left behind (17_HOOK_ENGINE§Native logging, 2026-09-15). A new native test binary that starts the harness needs the listener; a managed test project that stages it gets the sweep fromFrameLedger.DrainFixtures.targets. - Test isolation: a suite whose host resolves its target BY NAME consents to a run-unique copy of the harness
(beta.16).
build.ps1runs the test projects as parallel processes, andTargetResolvercounts every process of the consented executable's name — an unreadable one too, which is what ahook-harness.exeanother project is still starting is (its module list a partial copy). #269's first CI run lostAgentEndToEndTeststhat way:TargetAmbiguous, exit 6.AgentEndToEndTests,PartialRecoveryEndToEndTestsandCaptureHostEndToEndTestsnow copy the staged harness tohook-harness-<their scratch directory>.exebeside their data, asPipeEndToEndTestsalready did, register that directory with the sweep, and pin the copy to the staged harness byte for byte;AnotherHookHarnessRunningBesideTheSuiteDoesNotMakeTheTargetAmbiguousruns a second harness beside the capture (red against the shared name:TargetAmbiguous). A suite that attaches by pid (ShmDrainIntegrationTests, the nativefl_guard_test) is not exposed and keeps the staged name. - Test readiness: an injection test waits for the harness to say it presents (beta.27). The suites that start
hook-harness and then inject into it or look for it —
ShmDrainIntegrationTests,CaptureHostEndToEndTests,AgentEndToEndTests,PartialRecoveryEndToEndTests— slept 800 ms after starting it, a guess at when it presents. They now wait, bounded at 30 s, for the line the harness prints and flushes right before its present loop (tests/Shared/HarnessReady.cs, compiled in byFrameLedger.DrainFixtures.targets), and the pause case — which read the writer's status as INIT once under a full gate on 2026-10-19 — waits for the writer's READY (WaitUntilInitFinishedAsync) before it pauses. - The guard reads the test host too (beta.27). The guard's scan set is the target and its ancestors up to a platform
launcher, so for an injection test it includes the test host that started the harness,
dotnet testand the shell above them. Under a full gate the test host is still starting its other test classes, and a module it releases between the guard's list and the module's name refuses asModuleScanFailed— measured 2026-10-20, once in the gate and once in twelve runs ofFrameLedger.Infrastructure.Testswith coverage, never in a run of the class alone;ProcessUnreadable(2026-10-19, twice) is the same race met at the list.ShmDrainIntegrationTests.InjectAsyncasks again, four times at most, on those two reasons only. The end-to-end suites inject through the Agent or the capture host, which records such a session as Tier 2; they have the same exposure and have not shown it. - Guard integration: harness loads a dummy DLL named like an anti-cheat module → injection refused pre-launch; loaded late → safety unhook. (A renamed harmless DLL, not real anti-cheat software.)
Hook overhead measurement (NFR-1, per release)
-
hook-harnessin a fixed workload, present-rate uncapped: 3 runs hooked vs 3 unhooked → mean present-call cost delta must be ≤ 1 µs (measured with QPC around the call site in an instrumented harness build). -
Real game, fixed 10-minute scene, capture ON vs OFF, 3 runs each: game's own Avg FPS delta ≤ 0.5%, Agent CPU ≤ 1% of a core, Agent RSS ≤ 150 MB, Overlay resident ≤ 8 MB.
NOT MEASURED. The runbook exists as of 2026-09-10 (P2 PR-G); the run is the owner's.
tools/fps-impact-runbook.ps1sequences the six launches, holds consent granted for the ON leg and revoked for the OFF leg — the same launch path both times, so the only difference is the injection — reads the Agent's CPU and RSS itself, prompts for the number only the operator can see (the game's own benchmark average), and prints the block forspike-notes§14. It does not measure FPS and says so first: FrameLedger's frame times exist only in the hooked leg, so they cannot be both sides of this comparison. Dev-only, never inbuild.ps1 check— it launches a real game. The harness-level per-present cost (item 1's neighbour) is P0's 8.4 ns inspike-notes§3, and is not this number. Until the run happens, the honest state of this line is unmeasured, not passing. -
Results recorded in the release notes ("measured overhead this release: …").
Tier cross-validation (accuracy regression net) — GONE, 2026-08-28, and this is the entry to read before planning P5
Run the same game session with Tier 1 and Tier 2 simultaneously where possible (ETW does not conflict with hooking) and compare frametime distributions: Avg/1% Low must agree within 1%. Divergence beyond that means one of the two paths is wrong and blocks the release. This is also the check that would have caught the original detection-accuracy problem.
This gate is unsatisfiable and is struck rather than quietly dropped, because of its own last sentence. It described itself as "the check that would have caught the original detection-accuracy problem" — the problem the whole hook rewrite exists to fix. It required a second instrument measuring the same frames. PresentMon was that instrument; §S31 retired it and the owner dropped it, and the two-rung ladder has no second instrument by construction.
So state the consequence plainly: FrameLedger's Tier-1 frame times are not falsifiable by anything inside this project. Not "less well covered" — there is no second measurement to disagree with them. The suite still proves the ring, the hooks, the guard and the drain against fixtures and against the harness; none of that is an independent measurement of a real game's frame times.
This is the same absence that killed the frame-generation oracle, one layer over.
§S31 spent four oracles looking for an instrument that did not divide by our own
numbers; fl-baseline-probe fell to its own falsifier, two readings of Steam's overlay
were not independent because that overlay also hooks presentation, and PresentMon
classified every frame of a ×4 capture as an application frame. One root cause, two
casualties, and until now only one of them was written down.
What would restore it — none of these is chosen, and the row in 20_OPEN_QUESTIONS
§G owns the question:
- a Tier-2 mechanism of any kind, which restores both this gate and the FG oracle;
- an instrument outside the present path entirely (hardware capture), which is out of scope for v1 and is the only option that is independent by construction rather than by argument;
- the game's own frame counter, which §S30 records as still unrun and which is not a distribution, only a mean.
Until one exists, do not describe Tier 1 as validated. 03_METRICS §Accuracy
budget states the Tier-1 figure as a property of the mechanism — direct QPC at the
present call — and that is the honest claim: it says what is measured and where, not
that it was checked against anything.
Manual test matrix (per release)
| Axis | Values |
|---|---|
| OS | Win 11 dev machine · a Win 10 22H2 VM where one exists — since beta.37 (D95) Windows 10 is not blocked and is neither tested nor supported, so a run there is information, not a gate |
| GPU vendor | NVIDIA (primary) · AMD or Intel if available, else document as untested |
| API | D3D11 · D3D12 · Vulkan (layer path) · OpenGL · a 32-bit title, asserting it is correctly refused and recorded as Tier 2 with the reason |
| Upscaler | DLSS SR · DLSS-G · DLSS-RR · FSR2/3 · XeSS · none |
| RT | DXR 1.0 dispatch title · inline RayQuery title · non-RT title |
| Mode | launch mode · attach mode · mid-session settings change |
| Safety | game with anti-cheat → toggle disabled · simulated late AC load → unhook · double-crash → auto-disable |
| Tier | forced Tier 2 (records duration, available sensors and the reason — and nothing else) · Tier 1 → Tier 2 degradation notice |
| Accessibility (added 2026-11-09, beta.35, D88) | keyboard only with the mouse disconnected — every page, menu and dialog, Ctrl+1…5, Esc · Narrator — names, values, headings, the notices as they appear · a contrast theme on and off (Aquatic, Desert) · Windows animation effects off · a large display scale. The owner's first pass: 2026-11-09, keyboard only and Narrator on 0.1.0-beta.34 and 0.1.0-alpha.3, no fault (ACCESSIBILITY.md §How it is checked; the automated half is 08_UI §Accessibility) |
| Update | Velopack delta; update deferred while a game is hooked (FR-12) — a delta needs a previous release, so this row is reachable from the second release, not the first *(corrected 2026-10-04: no release has carried a delta — release.yml never fetches the previous release before vpk pack, 11_UPDATER §Built — so the update leg is a full package; since beta.13 a release carries a delta against the previous one, and the first |
| update that can use it is beta.12 → beta.13 — HANDOFF owner-only item 10)* |
Release smoke
Clean-VM install → Legal Gate → Agent setup (unelevated) → add a game → consent dialog → hooked 2-min session → charts render → CSV opens in Excel → Vulkan layer registered/unregistered cleanly → uninstall removes layer registration, task, and asks about data.
(Added 2026-11-09, beta.35, D88:) and an accessibility smoke, the same build: the Legal Gate, the consent dialog and
the session summary reached and answered with the keyboard alone; Narrator reads the notice a refused game raises; the
window readable under one contrast theme. The release's owner-only item in HANDOFF lists what that release moved.