SavePoint 0.9.22 #274

Merged
ethan-hann merged 79 commits from integration-0922 into main 2026-10-30 14:53:10 -05:00
Owner

SavePoint 0.9.22: the Code Cleanup batch (#248, #257 to #263, #268, #269), two 1.0 items (#251, #267), the Later follow-ups (#238, #240, #241, #244, #249, #253) and the GitHub release mirror workflow (#270, D71).

Product changes

  • Profile table rows are hoverable and right-clickable anywhere on the row (#267).
  • Steam launch messages: Steam that can't be started at all gets its own message (Flatpak/Snap named on Linux only), the waiting card says it waits 90 seconds, and a Steam game whose program quits cleanly within 30 seconds gets a hint that names Steam or a launcher and offers Edit game (#251, D67).
  • Help topics for online games and anti-cheat and the Steam app ID; leave-out patterns and mod file counts explained (#253).
  • New profile wizard: a folder's marker problem shows on that folder's box instead of failing the preview (#241); a folder left on an older snapshot after Go back gets its own reason (#249, D72).
  • Library picker and Ctrl+Shift+F explain an all-added list, a narrow games list and an empty library (#249). The picker ends loading only after its list is built (#268).
  • Free up space estimate: starts when a game is picked, measures folders and games side by side with at most 6 walks at once, counts a cut-off run's records as freed, stops when Settings is left and before a game is deleted, keys its cache on the pack list (#238, #244, D66.4).
  • FAT32 and exFAT game folders on Linux are pinned by path, not by file identity (#244, D57.6).
  • The shared snapshot space check judges several folders on their reduced size (#240, D16).
  • Fewer git starts per capture (first 8 to 5, later 5 to 4, watched 7 to 5), and ref-file shortcuts limited to the files ref backend so a reftable repository is never misread (#262). The switch and status candidates are in #273.
  • The Games screen's reload ends loading in a generation-guarded finally (#269).

Tests and tooling

  • Real UI dispatcher in the editor, shell, library, Games and History view tests; a UiThreadWrites helper flags off-thread writes; the selection-setter rule in the view model standard (#248, #260).
  • Render checks draw only with SAVEPOINT_RENDER=1; the Material color goldens are one loop per table (#257, #258).
  • The Free up space estimate is checked by what the player sees, through a contract step on both backends; fake-echo tests removed and every contract step has a proof test (#259, #261). The Known violations list in the tests standard is empty.
  • Gate: an unreadable memory reading picks one App process, the container limit less its usage caps the reading, the real readings run in a test, the two count checks are cross-checked, and the macOS gate fails when the build script parse test skips or is missing (#263, #249, D70).
  • GitHub release mirror: .forgejo/workflows/github-mirror.yml and build/mirror-release.ps1, stub-tested; the token may go only to api.github.com and uploads.github.com; a release event without GH_MIRROR_TOKEN passes quietly (#270, D71). Adding the secret and the first real run are the owner's.

Checks

  • Opus review: 0 critical, 0 high; all 2 medium and 6 low findings fixed.
  • Gap analysis against the issues: 62 of 76 items met at the time, the remaining gaps fixed or recorded (#273; reasons on the issues).
  • build/gate.ps1 passed on 0c6a606d; build/gate-linux.ps1 passed on 0c6a606d.

Closes #248

SavePoint 0.9.22: the Code Cleanup batch (#248, #257 to #263, #268, #269), two 1.0 items (#251, #267), the Later follow-ups (#238, #240, #241, #244, #249, #253) and the GitHub release mirror workflow (#270, D71). ## Product changes - Profile table rows are hoverable and right-clickable anywhere on the row (#267). - Steam launch messages: Steam that can't be started at all gets its own message (Flatpak/Snap named on Linux only), the waiting card says it waits 90 seconds, and a Steam game whose program quits cleanly within 30 seconds gets a hint that names Steam or a launcher and offers Edit game (#251, D67). - Help topics for online games and anti-cheat and the Steam app ID; leave-out patterns and mod file counts explained (#253). - New profile wizard: a folder's marker problem shows on that folder's box instead of failing the preview (#241); a folder left on an older snapshot after Go back gets its own reason (#249, D72). - Library picker and Ctrl+Shift+F explain an all-added list, a narrow games list and an empty library (#249). The picker ends loading only after its list is built (#268). - Free up space estimate: starts when a game is picked, measures folders and games side by side with at most 6 walks at once, counts a cut-off run's records as freed, stops when Settings is left and before a game is deleted, keys its cache on the pack list (#238, #244, D66.4). - FAT32 and exFAT game folders on Linux are pinned by path, not by file identity (#244, D57.6). - The shared snapshot space check judges several folders on their reduced size (#240, D16). - Fewer git starts per capture (first 8 to 5, later 5 to 4, watched 7 to 5), and ref-file shortcuts limited to the files ref backend so a reftable repository is never misread (#262). The switch and status candidates are in #273. - The Games screen's reload ends loading in a generation-guarded finally (#269). ## Tests and tooling - Real UI dispatcher in the editor, shell, library, Games and History view tests; a UiThreadWrites helper flags off-thread writes; the selection-setter rule in the view model standard (#248, #260). - Render checks draw only with SAVEPOINT_RENDER=1; the Material color goldens are one loop per table (#257, #258). - The Free up space estimate is checked by what the player sees, through a contract step on both backends; fake-echo tests removed and every contract step has a proof test (#259, #261). The Known violations list in the tests standard is empty. - Gate: an unreadable memory reading picks one App process, the container limit less its usage caps the reading, the real readings run in a test, the two count checks are cross-checked, and the macOS gate fails when the build script parse test skips or is missing (#263, #249, D70). - GitHub release mirror: .forgejo/workflows/github-mirror.yml and build/mirror-release.ps1, stub-tested; the token may go only to api.github.com and uploads.github.com; a release event without GH_MIRROR_TOKEN passes quietly (#270, D71). Adding the secret and the first real run are the owner's. ## Checks - Opus review: 0 critical, 0 high; all 2 medium and 6 low findings fixed. - Gap analysis against the issues: 62 of 76 items met at the time, the remaining gaps fixed or recorded (#273; reasons on the issues). - build/gate.ps1 passed on 0c6a606d; build/gate-linux.ps1 passed on 0c6a606d. Closes #248
ethan-hann added this to the Code Cleanup milestone 2026-10-30 14:16:44 -05:00
ComposedApp and EditorViewHarness take an optional dispatcher; the row-click shell view tests now use the Avalonia one so idle work runs in its real order. New UiThreadWrites records any change a view model or list announces off the UI thread, so a handler that posts can be told from one that writes on the pool thread. New off-thread tests cover the open editor (artwork arriving) and the library shelf (job ends, then a card is clicked); the Activity and Games off-thread tests assert no off-thread writes. FakeBackend.WhenEventsDeliveredAsync gets a 30 second deadline through an overload. The coding standards gain the rule never to change a control's items from its selection setter, and lose the fixed known violation.

Observed: the new tests passed on first run, so each got a mutation check. Making ArtworkProvider's handler and its inner post run inline failed the editor test on "GameArt.Source ..." off-thread writes; running ShellViewModel.OnJobChanged inline failed the library test on "JobStatusViewModel.IsVisible ...". Both mutations were reverted. No product bug found.
The UI-thread rule is now reachable from the standard's "When to Apply" question, the index entry for the standard names it, and both edited standards carry a new Last Updated time. No code changed.
The push mirror sends commits and tags but no releases. A new workflow
(release published or edited, or run by hand with a tag or "all") runs
build/mirror-release.ps1, which waits for the tag on GitHub, creates or
syncs the release (name, notes, prerelease flag), and sends every file
under its own name, skipping, replacing or removing files so a rerun is
a sync. Each file is checked against SHA256SUMS.txt before sending and
against the digest GitHub reports after. The GitHub token goes only to
api.github.com and uploads.github.com, through curl's standard input.

Tests run the script against a local stand-in for both servers.
Observed red: 19 failures, each "build/mirror-release.ps1 (or the
workflow) does not exist". After the script, the first green run had
two wrong test strings, fixed. Mutation check: disabling the host gate,
the checksum comparison and the digest comparison failed the host,
checksum and digest tests; the upload-host test was strengthened
because it did not fail under the mutation.
Pin how many git processes each engine operation starts (GitStartCountTests),
then drop the starts that only confirmed what a file already said:
- a branch with no loose ref file and no packed-refs file is unborn, so
  ResolveBranchAsync answers without rev-parse;
- the staged-before check reads HEAD from its file instead of rev-parse HEAD;
- no index file means no skip-worktree marks and no tracked files, so the
  mark clear and the tracked listing skip git.

Observed reds before the change (assertion failures): first capture started 8
(bound 5), watched capture 7 (bound 6), estimate before the first capture 1
(bound 0). Pins for a later capture (5), status (2) and switch (9) passed at
today's counts.

Starts per operation, 3000-file folder, before -> after: estimate 1 -> 0,
first capture 8 -> 5, watched capture 7 -> 6; later capture 5, status 2,
checkswitch 5, switch 9 unchanged. Wall time on that folder is dominated by
the add (about 8 s for the first capture) and was too noisy between runs to
claim a gain.
A release published before the secret exists printed a failure and turned the run red. A release event now prints a notice and exits 0 without contacting any server; a manual run with no token still fails. The workflow passes the event name as MIRROR_EVENT.

Observed red: the release-event test failed on the exit code and the manual-run test on requests sent before the failure; the workflow text test failed on the missing MIRROR_EVENT.
The row border had no background, so its padding and the gaps between cells were not hit-testable: the pointer was over nothing, the hover highlight did not apply and a right-click did not open the row menu until the pointer landed on text or a button. The table-row style now sets a Transparent background.

Observed reds before the fix: Hovering_the_empty_part_of_a_row_highlights_it (Assert.True failed, Actual False) and Right_clicking_the_empty_part_of_a_row_opens_its_menu (expected 'Switch to this profile', actual null).
A Steam client that cannot be started at all now fails as steamCannotStart, apart from one that started and never signed in. The waiting card says how long it waits, using the wait's own timeout. A program with no process name that exits with code 0 within the relaunch window ends as ProgramQuitAtStart and gets the Steam hint. The editor and help Steam text match the launch hint, the leave-out help topic names the full pattern syntax, the mod-file answer says when and where the counts show, and the anti-cheat and Steam app ID questions link to new help topics.

Observed reds (assertion failures): steamCannotStart expected, steamDidNotStart actual; waiting card body lacked the seconds; ProgramQuitAtStart expected, ChildExited actual; quit-at-start dialog title was the play time; editor note and help text differed from the new wording; the new topic, title list, picture and section-name tests failed on counts and missing topics.
On a FAT or exFAT game folder on Linux the device and inode can change between reads, so a destructive step could stop with 'folder changed' with no swap. IPathProbe gains KeepsStableIdentity (true by default); the host probe reads statfs f_type on Linux and says false for vfat/msdos (0x4d44) and exfat (0x2011BAB0). PathSafety.Pin then takes no identity and falls back to the real path and the consent marker, as D57.1 already does when no identity can be read.

Red observed: A_folder_on_a_file_system_that_changes_identities_... failed with Assert.Null(): Value of type Nullable<FileIdentity> has a value (FileIdentity { VolumeSerial = 1, FileId = 137 }). The KeepsInodes table test was written against a scaffold returning true.
Red observed: A_repack_that_moves_no_ref_is_noticed_on_the_next_ask failed with Assert.True() Failure (the cached answer was served after a repack). The detached-HEAD and mid-walk-cancel tests passed on their first run; mutation checks (dropping HEAD from the kept set, and passing no token to both rev-list walks) made each fail, then were reverted.
The library picker said "No games match your search." when the search box was empty and Hide added hid every game. It now says every game is already added and names the check box to turn off.

Ctrl+Shift+F did nothing, silently, in the icon rail or with no games added. It now announces why through the page's live region. Chosen over widening the rail because the rail width follows the window and a key press must not resize it; recorded in D64.

The macOS gate has no skip list, so a missing PowerShell skipped BuildScriptSyntaxTests without failing. SkipCheck gains --never-skip <test>; with no list it checks only those tests, and gate.sh uses it on systems without a list. The developer guide and Mac/Linux checklist now agree that PowerShell 7 is needed on every OS. The checklist gains keyboard and screen-reader checks (search counts, Clear search, chips, interface size, profile menu, Orca, wider VoiceOver) and folder identity checks for macOS and FAT/exFAT on Linux.

Observed reds: picker NoMatchesText was "No games match your search." where "Every game is already added..." was expected; three SkipCheck --never-skip tests returned exit 2 where 0 or 1 was expected. The shell key tests were written before the change but the implementation went in before their red run, so their red was not observed.
Observed red (GateSplitTests): an unreadable memory reading ran the lanes (expected process 'app', actual 'app-endtoend'); a 4 GiB cgroup v1/v2 limit under 20 GB of MemAvailable was ignored (expected lanes layout swap to one process, '4096 MB' not in output). The one-process thread cap tests passed on first run; mutation check: raising the cap from 4 to 5 in gate.sh and gate.ps1 made them fail (Expected 4, Actual 5), then reverted.

The gate now takes SAVEPOINT_GATE_MEMINFO_FILE and SAVEPOINT_GATE_CGROUP_DIR so the reading can be tested with fake files. The memory fallback note is copied into the gate.ps1 header.
Mutation checks on GamesSearchShortcutViewTests, one break at a time:
- narrow-list announcement dropped: the narrow-list test failed (expected the message, got null).
- empty-library announcement dropped: the empty-library test failed (same).
- focus moved anyway (to the rail list): survived, because focus was already on the starting control or the list did not take it.
- focus moved anyway (FocusSearch fallback): survived while the test started in the profile search; after starting from a profile row it fails, with focus on the search TextBox.
All breaks reverted; the class passes (18).
The new test passed on its first run; mutation check: making the bash check count methods case-insensitively (sort -fu) failed the 'case' scenario (expected 2 + 1 = 3), then reverted.
Snapshot space check (#240): a folder that fits on its raw size now offers
the combined check a way to find its reduced figure (content the store
already holds). CaptureRoomGate judges the raw sum first and, only when it
is refused, asks each folder for the reduced figure and judges the reduced
sum, so two folders whose raw total exceeds free room but whose reduced
total fits are both saved, and a reduced total that does not fit still
refuses. CaptureRequest.BeforeWriting takes the reducer as a second
argument. The capture progress byte total uses the reduced figure when a
check has already found it (no extra reads). Files git streams still count
in full; that needs stat data stored per snapshot and is not done.

Marker problem in the new profile preview (#241): a missing, foreign or
damaged consent marker no longer fails the whole preview. The folder is
listed with FilesCounted false and a StartEmptyBlock (NotConfirmed for a
missing marker, Other otherwise), the other folders are previewed, and the
wizard shows the reason on that folder's box and a "can't count" row.
New catalog keys: profiles.wizard.startEmpty.notConfirmed and
profiles.wizard.preview.unreadableSubtitle.

Observed reds: gate tests "Folders_whose_raw_total_does_not_fit_but_whose_
reduced_total_does_are_let_through" (DiskSpaceException, 501000 needed, 401000
free), "..._refused_with_the_reduced_figure" (expected 411000, actual 501000);
CaptureSpaceTests reduced figure (expected 20000, actual 320000) and byte
total (expected 20000, actual 320000); preview tests failed because the
whole preview failed on the marker; wizard test expected the confirm
reason but got "Take a snapshot first" and "0 files" in the row.
Observed red (BackendWorkspaceWorldTests): with the old OperationCanceledException-only catch, a test that joined a build whose starter was cancelled got the XunitException the cancelled capture raised instead of building the world itself. The no-retry test passed first run; mutation check: dropping the starter-cancelled condition made it fail (the joiner built a world after a real failure), then reverted.
The game picker dropped IsLoading before it built its visible rows, so a watcher
that saw loading end found an empty list. Observed red: expected (2, True),
actual (0, True). The picker now refreshes first and drops the flag last.

The wizard preview already set rows, note and error before IsLoading; two new
tests pin that order. They passed on first run, so a mutation check moved
IsLoading = false to the top of Show and Fail, and both tests failed.
RemovedSnapshots.EstimateAsync and the all-games estimate ran one repository after another; both now run side by side like ReadStateAsync. Records named by an unfinished run's note are added to the doomed set, so their data no longer counts as kept.

Red observed: Data_of_records_a_cut_off_run_still_owns_is_counted_as_freed failed with Range (600000 - 720000) Actual: 301600; Every_games_are_measured_side_by_side failed with Assert.True() Failure. The folder-level overlap test passed on first run; a mutation that serialized the folder estimates made both overlap tests fail, then was reverted.
The estimate reads the game's packs without the game's lock. Taking the lock for it would hold up snapshots and launches for as long as a walk lasts, so instead Delete game cancels the game's running estimates and waits for each git to exit before it removes anything; an estimate asked for during the delete is cancelled too. The cancelled call returns error.operation.cancelled, not error.unknown.

Red observed: Deleting_a_game_stops_an_estimate_in_the_middle_of_its_walk failed with Expected error.operation.cancelled, Actual error.gitOperation.failed (the walk lost its repository), and after the first cut error.unknown.
After Go back a folder sits on an older snapshot, and the wizard's box said
SavePoint can't check the files and to close programs, which is the wrong step.
Core now reports StartEmptyBlock.OlderSnapshot for that case, and the wizard
says to load one of the game's profiles first.

Observed reds: Core, expected OlderSnapshot but was Other; after the enum
reported it, the refusal fell through to startEmptyUnsaved instead of
startEmptyUnchecked. App, the hint read 'Take a snapshot first...' instead of
the older-snapshot reason.
Observed red (GameSessionWatcherTests): NamedProcessSeen read null where True/False was expected (property stubbed first so the red is an assertion, not a compile error). The no-name test and the log assertion passed first run; mutation check: reporting _nameSeen even without a process name failed both, then reverted.

Also records the mutation checks of the UI sweeps, all caught and reverted: setting PillRoomHeight to 0 failed all 28 JobPillPlacementTests cases (text under the pill); dropping the toast room in MainView failed all 28; TabIndex=-5 on the editor's Save button failed the merged editor tab walk (expected Save changes, got Name); TabIndex=5 on BrowseSavesButton failed its card-order check (positions 4, 8, 45, 13). No test needed strengthening.
Leaving Settings now cancels a running Free up space estimate (and ignores its late answer), and the screen disposes its FreeUpSizeEstimate with the shell. Returning to the screen asks again.

Red observed: Leaving_/Closing_the_screen_stops_a_size_... failed with Assert.All() Failure: 1 out of 1 items did not pass (the estimate's token was never cancelled), after adding empty OnHidden/Dispose members so the tests compile.
Render checks: the Activity, style gallery, Games/History/Shell search render cases only wrote a PNG and asserted it existed or was large. They now skip unless SAVEPOINT_RENDER=1 (listed in both expected-skip files). Layout tests that also drew a picture (Activity, PinnedFooter, HelpAccess) draw only on request. RenderSettleTests checks the animation settle directly instead of drawing a PNG.

Goldens: HctTests, TonalPaletteTests, ContrastTests, DynamicColorToneGoldenTests and the per-role rows of Md3PaletteGoldenTests were one case per row. Each is now one test that walks the golden table, collects every mismatch with its row inputs, and reports them together, with a guard on the row count.

Observed mutation results: Hct.From tone +0.3 fails Hct.From with 776 mismatches in 280 rows; RatioOfTones +0.001 fails all 64 rows; DynamicColor.GetTone +0.5 fails 344/450 single and 426/400 pair mismatches; Md3Palette.Build fails 515 of 540 roles. Skipping the animation settle loop fails RenderSettleTests (opacity 0.87, expected 1). All mutations reverted.
The confirmation waits only 1.5 s for the size, so a store whose estimate takes longer never showed one. The footer now asks for the preview and starts the estimate as soon as a game is selected; the backend caches its answer, so the confirm-time ask is cheap. Picking another game cancels the earlier one.

Red observed: both new tests failed with Assert.True() Failure (no estimate call was ever made without a button press). A_yes_stops_the_size... now expects at least one estimate token instead of exactly one.
# Conflicts:
#	docs/developer-guide.md
Background delivery test: assert the capture call completed on return and that
each event's thread differs from the thread that made the call raising it, so
delivery on the caller's thread or inside the wait fails. Mutation: draining
on the waiting thread instead of a drain thread turns it red (NotEqual).
Comments in the test and FakeBackend now give the real cause of the earlier
flake: the wait resumes inline on the drain thread.

GamesScreenViewModel.ReloadAsync: the read runs in try/finally that clears
IsLoading only for the current generation, so a throw after the last await no
longer leaves the loading flag up and a newer reload's flag is left alone.
The error is set before the rows are cleared, and ShowLoading requires that
the game has not loaded yet, so it never shows with the empty state or on a
failed reload of a game that had rows.

Reds observed (assertion failures): ShowLoading true during a failed reload of a
game with rows; ShowLoading and ShowEmptyState both true on a reload of an empty
game; IsLoading true after a throw following the last await. Mutation: a plain
finally fails the newer-reload test. The comment on the polling tests is
corrected and the folded bool is split in two.
# Conflicts:
#	docs/testing-on-mac-and-linux.md
Deleted tests of test tooling that cannot fail for an app regression: the call-order and
cancellation-token recorder tests (FakeCallLogTests, now FakeFailNextTests, keeping the three
FailNext cases other tests rely on), the FakeDialogs queue-order and default-answer cases, the
empty-scenario case, and echo cases in FakeLibraryApisTests (update, paths, executables, import,
templates, relocation and clear previews, default startup status), FakeJobsApiTests (progress
replaces, LastStarted), FakeLaunchApiTests (running defaults to clock, scripted preview) and
FakeRecentlyRemovedTests (none removed lists none).

Added four BackendContractTests that name the shared-story steps no test proved (library, profile
and settings lifecycle; refusals and empty answers; switch-job blocking; second profile history).
They passed on first run, so each was mutation-checked: renaming one step in ContractScenario made
all four fail on an assertion. Mutations reverted. Removed the known-violation entry.
# Conflicts:
#	.agents/coding-standards/tests-meaningful-and-fast.md
FakeGamesApi and FakeSettingsApi set EstimatesStoppedAtRunStart themselves and two
tests asserted it, so they checked the fake's own bookkeeping. Remove the flag.

The shared contract story now asks a game's size from a real preview token, runs
Free up space with that token, and asks again after the run, then does the same for
every game, against both the fake and the real backend. The jobs the story asks for
are compared like any other job and no longer counted as work the backend did on
its own.

The Settings test now asserts the row never shows a size that arrives after the run
starts; the Games test asserts every estimate was cancelled and the question named
no size.

Observed red: the new contract assertion failed with "Item not found in collection:
size a game's free up space would free: 0 bytes". The two App tests passed on first
run; a mutation check removing the _size.Cancel() before the run in each view model
failed them (Games: 1 of 2 tokens not cancelled; Settings: row read "Deleting them
frees up about ..."). The view models were restored.
Red: A_test_named_never_skip_that_is_not_in_the_reports_fails_the_check_and_is_named, Expected: 1 Actual: 0.
# Conflicts:
#	.agents/coding-standards/tests-meaningful-and-fast.md
Red (three): file missing from the list and no list both exited 0 (Expected: Not 0, Actual: 0); same-size file with no digest stayed 'SETUP BYTES' instead of 'setup bytes'.
A later save started ls-files twice, once for what changed and once for what the index records. One ls-files with -t tags now answers both (a "?", "C" or "R" tag is work to stage; any other tag is a recorded file), so a later save starts 4 git processes and a watched one 5, down from 5 and 6. Which files are staged does not change.

Reading a branch from its ref file, and calling it missing when there is none, only holds for the files ref backend. A repository with reftable storage has no ref files, so a branch delete reported success and kept the branch. The shortcut now applies only when there is no reftable folder and the config does not name reftable storage.

Observed reds: the later save bound of 4 failed with "started 5 git processes" and the watched bound of 5 with "started 6"; deleting a branch in a repository migrated to reftable left refs/heads/spare in place. New listing tests were green before the change; dropping the "?" tag in a mutation failed 4 of them. LeftOutReadOnceTests' mark-listing detector now ignores a listing that also asks for new files.

Candidates left: merging the diff and the new-file listing (the diff also sees staged work, and the in-folder ignore files would change what counts), skipping the index refresh before the verify diff (reads run without git's optional locks, so only update-index writes the index), the empty-folder trailer read (nothing else reads that commit's text during a switch), and the case-clash ls-tree (it must see the whole target tree, not only the changed paths).
Red (Core): A_game_folder_without_a_marker_is_left_uncounted_and_never_blocks_starting_empty, expected None, actual NotConfirmed.
Red (App): A_program_that_quit_cleanly_right_away_..., NotSeenHint differed from the new text (EditGameLabel was empty).
The quit-at-start hint and toast now name both causes: Steam not running, or a launcher handing off to the game.
Share one gate of 6 walks across all folders and games of one estimate call, matching the listing. D66.4 states the bound.

The delete test now releases the blocked walk only after a probe ask for the game ends as cancelled, so it proves the stop.

Observed red: the new bound test failed with Assert.InRange (2 - 6), actual 18, before the gate. Mutation: replacing the StopAsync line in GamesApi.Delete.cs with a no-op made the delete test fail on Assert.True(stopped); reverted.
Red (six): The_gate_takes_what_the_container_already_uses_off_its_limit on both scripts, v1 and v2; one-process rows saw the app lane instead of app-endtoend after reading 16384 MB, the lanes rows lacked '15360 MB'.
Two more checks read ref files directly. In a repository with reftable storage there are none, so the stored-content check said no snapshot existed and the index rebuild took the folder for one nothing was saved from, rebuilding an empty index over real history. Both now use the shared files-backend check; stored content answers yes (reads are the cheap side of a wrong answer) and the rebuild asks git whether HEAD names a commit.

Red by mutation (the new tests passed on first run only because the guard was written first): making the files-backend check always true failed all three reftable tests, the branch delete, the unborn check and the snapshot check.
Observed: both classes pass (49 and 17 cases) on the real dispatcher, so no ordering bug was hiding behind the instant one.
Observed reds: the workflow text test failed (pattern for a fixed concurrency group not found); the three make_latest stub tests failed (expected true/false, actual null).
First run passed. Mutation: renaming the CIM class in gate.ps1 made the PowerShell case fail (no reading reported); reverted. The Bash case skips on Windows (listed in expected-skips/windows.txt) and could not be mutated here.
Not reproduced as a red here: Windows pwsh already printed the message unsplit. The two refusal tests and the rest of GitHubMirrorTests pass (27).
Observed red: ErrorCopyTests Steam_that_cannot_start_names_Flatpak_and_Snap_only_on_Linux failed for windows and macos (Expected False, Actual True).
A program with no process name that exited with code 0 inside the relaunch
window ended as ProgramQuitAtStart for every game, so a non-Steam game's short
run got the Steam hint (PlayJourneyTests failed: "The program closed right
away"). The watcher and starter now take a steamGame flag, set by the
orchestrator from SteamGames.NeedsClient; other games end as ChildExited.
D67, the architecture note and the user guide say Steam games only.

Observed red: GameSessionWatcherTests.A_program_that_is_not_a_Steam_game_and_quits_cleanly_right_away_ends_as_child_exited
Expected: ChildExited, Actual: ProgramQuitAtStart. The Steam orchestrator test
passed first time; passing false from the orchestrator made it fail (Expected
ProgramQuitAtStart, Actual ChildExited).

The legacy-storage test now names the startup warnings when the history is not
in the new default, so a load-only failure explains itself.
The test freed its hold from a polling task that raced the product's re-read, which retries only briefly, so a starved task could leave the file held through every attempt. The hold is now freed on the product's own thread at the log line of the first failed re-read. Red observed: delaying the old release by 8 s made the test fail with the settings IOException; 1.5 s did not, since each read retries internally for about 1.3 s.
Merge flaky-legacy-0922 into integration-0922
All checks were successful
Docs / site (pull_request) Successful in 1m30s
Gate / linux (pull_request) Successful in 17m19s
Gate / windows (pull_request) Successful in 34m34s
Docs / site (push) Successful in 1m30s
Release / windows (push) Successful in 2m41s
Release / linux (push) Successful in 2m42s
Release / checksums (push) Successful in 31s
GitHub mirror / mirror (release) Successful in 28s
0c6a606def
Sign in to join this conversation.
No reviewers
No milestone
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
ethan-hann/SavePoint!274
No description provided.