Skip to main content

vm lifecycle

The marlin-vm and marlin-vm-orca51 CI jobs install the ordinary Marlin preview image built from source, then a CI-only fixture layer (packaging/marlin/vm-lane/Containerfile). The layer gives the image the published marlin:tuna login path: stock GDM with tunaOS's tuna-flavor GDM steps (tunaos.toml [desktops.tuna] after_install: greeter theme overlay, FallbackSession=tuna with version-appropriate global account templates, every other session hidden), vendored byte-for-byte in packaging/marlin/vm-lane/tunaos/ from tunaOS d98ed1e09b61635a2a9b115ce559944e42f97d74, on the package's own Tuna Desktop session entry. The fixture account has no saved session. No autologin or initial session exists: the lane drives GDM's user list and types the fixture password exactly as the published-flavor nightly does (V-GREETER-LIST, V-GDM-FALLBACK, V-GDM-LOGIN; docs/marlin-tuna-nightly.md), from the same scripts/tuna-vm-lane code. Logout must return to that GDM greeter. The ordinary two lifecycle/Orca rows retain five pristine boots, the default GPU, fresh-account journeys and all aggregate gates. These ordinary lifecycle fixtures do not change the shipped preview image; the independent paired DEV profiles below are separate members of the same existing VM producer. GDM's session worker opens a journal stream for gdm-wayland-session, forwarded to the serial port. The fixture's session environment comes from environment.d, which gdm-wayland-session imports from the user manager; the per-login Orca test secret comes from a user environment generator. The ordinary CLI profile is --profile lifecycle (the default); its two matrix members keep their original distribution/upstream51 Orca contracts and marlin-vm/marlin-vm-orca51 artifact identities.

The speech lane invokes the pinned package factory through bash; the factory checkout's script execute bit is not an admission prerequisite. Its actual native package build and all speech gates remain required.

Installed strip captures​

The strip probe binds the original native processes and all three installed fixture helpers to the selected package, source and final image. Both PNG and PPM acquisitions must pass all 25 original body-color samples; the captures remain separate acquisitions, not presentation-latency proof.

A real notification can cover those pixels. Before acquiring a frame, the probe may acknowledge its visible banner through ordinary QMP input: Super+N, Tab, then Return only after the original shell's public Close button is observed focused and sensitive. The read-only AT-SPI observer binds the control's bounds to original process/bus identities, never notification content. Every pixel check and the existing 10-second capture and 120-second phase deadlines remain unchanged. No updater service, notification policy or failure log is disabled; acknowledging a banner does not qualify the update operation that produced it.

The column-height capture also submits a real critical notification through the session's public notification service. It cannot pass until the original shell shows the fixed public fixture title, its sensitive Close button receives the actual QMP Return action, and the banner disappears. Only a boolean title comparison is retained. This control exercises the recovery path even when the updater happens to show no banner; it shares the same capture deadline and still requires every original PNG and PPM color sample afterward.

Close readiness uses GTK4's actual AT-SPI state mapping: an interactive control advertises SENSITIVE, the inverse of its DISABLED accessibility state. GTK4 does not advertise ENABLED; requiring that additional bit would reject a genuinely focused, interactive button. Both observed bits remain in the receipt, but disabled controls remain ineligible. See the GTK4 state mapping.

The published-flavor lane uses these same strict assertions. --published only records acquisition: no old-package or absent-instrumentation branch can pass by crediting a current-build lane. Every boot manifest explicitly records acquisition, qualification_mode: "strict", waived_assertions: [] and skipped_assertions: []; the nightly/repeat receipts preserve and classify every actual boot (receipt contract).

Required independent native automatic timezone​

Whenever marlin-vm is selected, its matrix also requires marlin-vm-autotz. This uses the same original package producer and VM image recipe, with --profile autotz --require-kvm --boots 1: one pristine snapshot boot, stock GDM and the original non-root standard account, unchanged default GPU, and no disposable disk or full Core application installer/inventory. There is no dispatch, secret or repository-variable waiver. The leg never reuses a retained native product: that capsule's receipt binds only the lifecycle members, so this image is always built fresh from the same run's package and checkout.

The final immutable image receipt binds the exact executing source commit, original package hash and original installed ELF hashes, lifecycle helper, normal GeoClue policy helper and four fixed native-session observers: native_session.py, autotz_guest.py, autotz.py, geoclue_fixture.py. The latter two are installed from the clock-only source files, not the full Core guest. The ordinary preview policy invocation remains unchanged. This is same-run source/image admission, not signed release proof.

The root autotz DOMAIN OPERATION JSON route delegates to the installed standalone observer. QGA acquisition and nested commands share finite absolute deadlines; reconnecting cannot renew them. Failure diagnostics contain only bounded domain/operation/type/sites and optional numeric codes, never raw arguments or exception text. Unconfirmed guest-child retirement remains fatal. V-AUTOTZ requires the consumer's explicit complete qualified result; an exception, missing native session or missing result fails the row and therefore the existing aggregate CI gate. The consumer retains its actual original-shell AddAgent, source/provider, daemon location/timezone and panel-clock gates and restoration requirements. Routing implementation alone claims no native runtime qualification.

Pristine release-disk qualification is separate​

The signed release-disk consumer and its native observer live in scripts/lib/try_tuna_desktop.py and scripts/lib/try_tuna_qualification.py. They do not turn the fixture journeys above into pristine firstboot proof. The observer requires the normal-release signed manifest and its exact declared bytes, with the observer checkout matching the signed producer source. It boots a writable copy while retaining the shipping disk's hash; there are no injected accounts, passwords, SSH keys or fixture services in that shipping disk.

Each native Linux/macOS amd64/arm64 host must execute its matching signed Corral binary, real local QEMU, browser and noVNC transport. Executable headers reject foreign host binaries even where Rosetta/binfmt could run them. The guest remains the declared qualified amd64/UEFI disk; ARM hosts explicitly use TCG. Receipts retain actual QEMU target/accelerator, firmware and binary hashes rather than claiming native guest acceleration.

Actual GNOME Initial Setup creates a passworded operator account in the copy. Normal GDM must then enter Tuna, including visible application search, with wheel membership and password-required sudo; logout/login and reboot must preserve that account and user data. RFB connection alone means browser/transport readiness, never successful guest boot or Tuna login. Terminal cancellation (SIGINT, SIGTERM or SIGHUP), stop, failed browser launch and failed acquisition must retire owned processes/listeners/locks without touching an unrelated VM. Normal launch preserves writable user data; qualification explicitly deletes only its own disposable copies after retirement. The owned stop socket remains serviceable while browser readiness is pending; an opener returning zero is not a connection. Native qualification must prove early stop after that real no-RFB opener, and each cancellation signal must follow an actual current RFB connection in native Chromium. Its retained executable/profile identity and actual owned-child wait/exit receipt must agree; a pre-browser cancellation or a stale readiness receipt cannot pass.

Native qualification also retains current noVNC milestone PNGs and four bounded actual-guest video clips: pristine setup, the first Tuna session, passworded re-login, and the post-reboot session. It reuses the VM lane's QMP recorder and VP9 encoder, not a slideshow of stills. Each clip contains at most the first 30 seconds of its observed UI journey, with capture timing retained; it does not claim to show every account-setup or boot step. The PNGs are actual noVNC canvas pixels; the clips are actual owned-guest QMP pixels, not browser chrome. The qualification hosts require ffmpeg with libvpx-vp9 in addition to their existing browser observation tools; normal launcher users do not.

media.json binds those files to the signed manifest and pristine disk hashes, desktop/package source, immutable image index and platform digest, exact signed Corral release and binary, observer source, and native host. Failed or incomplete qualification records qualified: false; neither historical diagnosis frames nor before-fix firstboot screenshots may be promoted as current documentation proof. Documentation intake must retain these identities and all native gates.

The consumer requires Python 3 and the native package manager (on macOS, Homebrew and its normal command-line-tool prerequisites). It obtains the checksum-pinned Sigstore verifier itself and acquires QEMU plus amd64 UEFI through native package trust, including when QEMU exists but firmware is missing. Linux without caller access to /dev/kvm explicitly uses TCG. Interactive Linux hosts use their normal sudo authentication, with its password input connected directly to the real terminal; the wrapper never reads or stores that password. Without a terminal, sudo must permit noninteractive installation or acquisition fails closed. Cancellation waits for an already-started native package transaction (and its normal authentication prompt) to finish safely; it does not start a VM afterward. No Podman or guest disk build belongs in this consumer. An artifact directory changes only where the same signed bytes are read; it does not replace identities, trust roots or descriptors, and is not a promise of network-free dependency bootstrap.

Corral keeps its normal policy for listener authentication. The consumer preserves CORRAL_AUTH_REQUIRED, CORRAL_ADMINS, CORRAL_PEER_TOKEN. It also preserves Corral's ordinary settings for URLs and tokens. It never clears authentication to make a readiness report succeed. It excludes overrides for backend discovery and plugin acquisition, not Corral authentication. Each lifecycle command names QEMU and the exact owned VM.

The existing public wrapper is not cut over until a real signed disk and manifest pass all four native firstboot/browser/lifecycle qualifications. Private helper code or fixture receipts alone do not establish that result.

Record-free fresh accounts (tunaOS#4092)​

tuna-test above has a session-less AccountsService record, so it proves GDM's FallbackSession path only. scripts/tuna-vm-lane --fresh-accounts is a separate boot of the same disk (QEMU -snapshot) that passes the tuna.vm.fresh-accounts systemd credential (SMBIOS type 11). Only then does tuna-vm-fresh-accounts.service run useradd -m before GDM starts, as an installer would: tunastd (standard), tunaadmin (wheel), and the control tunaseeded, whose record is pre-seeded. tunastd and tunaadmin get no per-user AccountsService record, no per-user Session preselection and no autologin. This is distinct from the global AccountsService templates: on an installed GDM 51+ with FallbackSession support, their [User] group explicitly sets Session=. GDM can launch the configured tuna fallback without propagating that name to the worker's persistent settings; empty templates alone have not qualified record creation. Older GDM retains global Session=tuna: it can prefer launchable GNOME even when that session has NoDisplay=true. An absent or unparseable GDM version must not opt into the empty-template policy. Neither policy edits existing per-user records or replaces a valid saved session choice.

For the general credential-only fixture, GDM's own org.gnome.login-screen disable-user-list=true asks directly for the user name. The fixture inserts its tuna-vm-fresh-accounts dconf database ahead of the preserved vendor profile and requires an actual gsettings get readback of true. The vendor profile need not contain system-db:gdm; writing gdm.d alone did not apply the setting in run 41234017707. In run 41230684089 the old Tab count to "Not listed?" selected the listed tunaseeded instead. The private native candidate retains the normal user list and the separately proven bounded navigation reset: its package repair must qualify the ordinary greeter, not that fixture shortcut. The installed manifest controls only this setup choice; strict candidate and installed-pair admission is still required before any native persistence gate.

The process/logind greeter predicate is not an input-readiness predicate. In run 41248530233, artifact marlin-vm-orca51 (ID 13061267386, SHA-256 1de554a272722c7640b767cde1f75cfa7603910bd7c0c3cda4b4895f38d79af2) retains a black fresh-accounts/fresh-greeter.png; its tunastd-first-password.png still shows Username, not Password, when the old driver sends the password. The serial journal records pam_faillock(gdm-password:auth): User unknown at 19.270579 seconds and no user session. The standard counterpart reaches a visible Password prompt and authenticates both accounts; its later admin sudo failure is separate. The run's PR source is df05bde92cd4392871db9d36d3d3a21ce43b846b; workflow checkout/manifest metadata is the synthetic merge b4d42581e7299323fd6ead9f76dcd94862990f1c. Both use published Marlin base sha256:6c1e25b1fa9ddc07d7e71f767456839676333d3ebdc332b642a467f56c8ff71b; the Orca test-layer image ID is sha256:0e3b49a6df3ee644fb9ec4f1dc5b0971da0e6d67b4aedbbc0bedcb5ff116b2e8.

The general credential-only driver waits for the actual greeter's focused, visible, showing, editable, sensitive and enabled AT-SPI entry before typing each credential. The probe runs as the active non-root greeter UID, binds the accessibility bus PID/UID to the original gnome-shell process and start time, and distinguishes text from password-text roles without reading names or entry contents or invoking accessibility actions. Prompt receipts and pre-input frames are retained. Service/greeter readiness uses the existing greeter deadline; name, password and session readiness share one existing login deadline. There is one credential pair per general-mode login, no fixed-sleep increase and no authentication retry. The admitted native candidate retains its existing bounded stock-user-list attempts. This fixes wrong-phase input in the general fixture; old receipts alone do not identify which input was lost during startup.

The first native observer run found 1267 distinct GNOME greeter actors, exceeding the unchanged 512-node observation budget when hidden actors were enumerated. This was not a cycle (maximum depth 15, no duplicate objects). The actual focused username entry had all required flags, and every non-root ancestor was SHOWING/VISIBLE. Pruning non-SHOWING branches visits 131 nodes instead, including hidden branch roots, without changing the node/depth/time limits. The application metadata root is exempt from this prune because it is not a visible component. This follows the existing shell-focus observer's SHOWING traversal rule.

GateActual operation and evidence
V-FRESH-RECORD-FREE-<user>Before any login: account exists with a home, no /var/lib/AccountsService/users/<user>, no GDM automatic/timed login, wheel only for the admin
V-FRESH-CONTROLThe same probe classifies the pre-seeded tunaseeded as NOT record-free
V-FRESH-LOGIN-<user>-{first,second,reboot}General fixture: owner-bound focused editable GDM text and password-text prompts precede QMP credentials. Explicitly admitted native candidate: existing bounded stock-user-list navigation. Both require exactly one gdm-password VT session of that account running tuna-compositor and no gnome-shell/gnome-session
V-FRESH-SESSION-<user>-{first,second,reboot}Read AccountsService's effective Session over the system bus: it must be tuna; if a cache record exists, any saved Session must also be tuna
V-FRESH-RECORD-CREATED-<user>Verified private native candidate only: a physical AccountsService record must exist during the first logged-in session; any saved Session must be tuna
V-FRESH-SUDO-<user>sudo -S -k true as the account with its password: admin succeeds, standard is refused by sudo policy
V-FRESH-LOGOUT-<user>-<round>logind ends the session and the greeter is ready again
V-FRESH-REBOOT / V-FRESH-SESSION-RELOADED-<user>A different boot_id and ready greeter, then effective AccountsService Session=tuna reloads before login; both accounts authenticate into Tuna again
V-FRESH-RECORD-KEPT-<user>Verified private native candidate only: physical records for both accounts remain before either post-reboot login

Every path retains the effective-session checks from #989, including the GNOME reference's valid record-free general case. Explicitly admitted native candidates additionally retain every original physical record creation and retention criterion; an absent or invalid candidate cannot opt into native qualification. The private candidate's 24 original gates are not replaced by the additional effective-session coverage.

The controller reads a nonempty old boot ID through the read-only greeter action, then calls GuestAgent.request_reboot() once. It sends QGA guest-shutdown with mode=reboot. QGA returns no response on success and does not guarantee shutdown. The method retires its transport after it tries to send; send errors propagate. A return is not an acknowledgement or proof of reboot.

The controller reconnects at once, with no extra sleep. It retains the 180-second limit for a new boot and ready greeter. A boot ID that did not change cannot pass, nor can a greeter that is not ready. Both post-reboot logins and logouts remain mandatory, as does physical record retention in the admitted native campaign. The native 60-second login limit, two authentication tries and 90-second trace stay unchanged.

Run 41215443626 failed both native VM jobs with the AccountsService record still missing after a real record-free tunastd manual GDM login on GDM 51.1.1-1 / AccountsService 26.27.3-1. First login, standard-user sudo denial and logout passed; admin, second-login and reboot coverage aborted. The inherited global Session=tuna suggested a same-value setter might explain the missing write. Run 41237623922 disproved that as the remaining cause: even empty effective settings left the record absent. An empty global template is not itself proof that persistence is fixed: only the unchanged native gates above can qualify the fix. No setup script may create the fresh accounts' records or invoke a cache operation to do so.

Receipts are in fresh-accounts-receipts.json; frames per attempt and session. Both native VM jobs must complete the standard and administrator first login, real record creation, sudo policy, logout, manual second login, reboot, record retention and post-reboot login gates. Only Arch/GDM 51 is covered by this native lane; helper-execution tests also cover legacy GDM template policy, but do not constitute a pre-51 native proof.

Run 41226473836 (desktop c3713c34) left the record absent after a real orca51 tunastd first login, and the helper's chosen policy unobserved. Two read-only diagnostics now record what happened; neither is a gate. The image build prints tuna-vm-lifecycle gdm-policy: the gdm the helper resolves and its realpath, bounded --version output/status and parse, the daemon/FallbackSession literal-probe status, the derived empty/legacy policy, pacman -Q gdm libgdm accountsservice and the binary's owner. Each fresh-state receipt adds an accounts block, collected in order: physical record presence, the global standard/administrator templates (candidate presence in /run, /etc, /usr/local/share, /usr/share precedence, the winner, and every [User] Session, SessionType, XSession and SystemAccount occurrence: [] is missing, [""] empty), Properties.Get without D-Bus activation of exactly Session, SessionType, Saved and LocalAccount on /org/freedesktop/Accounts/User<uid>, then record presence again. An unexported object, absent bus or wrong type is an explicit error (remote D-Bus error name or exception type only), never an empty Session or false Saved. Collection points: preflight, first session, after each first-round logout (before a missing-record abort), after reboot, and once when a login fails. Nothing calls FindUserByName, CacheUser or a setter, retries, or dumps other template keys, error messages or stderr. A record seen only after logout, Saved=true or an effective Session=tuna never substitutes for V-FRESH-RECORD-CREATED-<user> at the first session.

Observed GDM 51 persistence failure and bounded trace​

Run 41237623922, head f7208bba, failed V-FRESH-RECORD-CREATED-tunastd in both native profiles. The receipts show UID 1001, a real gdm-password VT 2 Wayland login running tuna-compositor and tuna-shell-gtk, successful standard-user sudo denial and logout, but no physical record during the session or after logout. The winning /etc/accountsservice/user-templates/ standard and administrator templates each contain Session=; effective Session and SessionType are empty, Saved=false, LocalAccount=true. Packages are gdm 51.1.1-1 and accountsservice 26.27.3-1. Administrator, second-login and reboot acceptance remains unexercised, not passed. The retained logs of jobs 124646121324 and 124646121372 contain none of the four allowlisted worker markers below and no session-setter call headers. This is an evidence gap, not proof that saving was skipped.

For review before the next native campaign, add just --fresh-gdm-trace to its existing scripts/tuna-vm-lane --fresh-accounts ... command. The opt-in helper tuna-vm-gdm-trace runs only around the first ordinary tunastd login attempt, stops before any existing selection retry, and is stopped in finally on login/input/guest-agent failure. The observer has an independent 90-second lifetime, a 1 MiB header-input bound and 128-event bound; these add no login timeout, wait, retry, session or acceptance gate. Transient-service startup allows at most two seconds for observer readiness. Existing-marker queries additionally stop at 64 KiB or three seconds. The trace receipts are tunastd-first-trace-{start,stop} in the existing fresh-accounts-receipts.json; collection failure is explicit and non-gating.

Privacy contract:

  • dbus-monitor --system --profile, not normal monitor/binary mode, prints tab-separated message headers and no argument or error bodies. Match rules select only SetSession, SetSessionType, SetXSession calls on /org/freedesktop/Accounts/User1001. Return/error headers have no path, so the observer consumes their metadata and retains only replies correlated by call sender/serial and reply destination/reply-serial. Unrelated headers are discarded in memory; no raw bus log is written.
  • Existing gdm-password worker journal entries are reduced to four fixed labels: saving account details for tunastd, session name set to tuna, save failed, and SESSION_OPENED. PID, time, method and reply metadata are sufficient; usernames other than the fixture, journal text, setter argument values, error names/bodies, PAM conversations and environment values are not exported by this observer.
  • Never enable GDM debug or send SIGUSR1. The audited upstream worker source logs complete PAM query payloads at line 757, selected choice responses at line 576 and PAM environment entries at line 1537. Filtering this observer's output cannot make a broader journal/serial export safe. Missing pre-existing markers are reported as unavailable-or-debug-disabled; no conclusion about the save stage follows from their absence.
  • No probe invokes FindUserByName, CacheUser or an AccountsService setter, writes a per-user cache record, changes GDM configuration, or offers a fake second desktop session.

A matching setter error header proves an error reply, not its cause. A successful reply without a record points at persistence, not at missing session propagation. No setter observed with unavailable worker markers remains incomplete evidence. The physical-record gate is unchanged.

Root-fix candidate and actual Arch package boundary​

The upstream GDM 51.1.1 normal-session open path sends display mode and XDG_SESSION_TYPE, then Open; it does not send SetSessionName. Explicit selection does send SetSessionName(get_session_name(self)). With one offered Tuna session, GNOME Shell 51.0 hides the session menu. The settings save routine can succeed with null session/type/language and, for a local account, without a setter or CacheUser. Registration is not this save trigger. Inference: fallback selection may never reach the worker's settings.

Run 41245509890 had healthy observers in both native profiles. Neither saw a call to set the User1001 session. Worker markers: unavailable; debug off. Their absence proves nothing about the save stage. Neither run proves the cause.

The native image gets GDM from Arch through pacman, not this repository's packaging/arch/PKGBUILD (which builds only Tuna Desktop). The canonical tunaOS tunaos.toml desktop package list requests plain gdm, without a Tuna GDM version/patch override. This lane starts from ghcr.io/tuna-os/marlin:gnome; both the preview Containerfile and VM fixture layer run pacman -Syu, so a future replacement must survive the final package transaction rather than being silently overwritten.

The Arch recipe at revision 36e76b2b splits gdm and libgdm from one build. It uses the upstream tag and an Xsession patch. Its prepare() uses git apply -3. The build uses arch-meson, meson compile and meson test.

Package source PR963 adds a private candidate at 7b52bdf34be22bc60fef62b157be64b589b13f6a. It builds GDM and libgdm 51.1.1-1.1 together. It retains the PAM configuration, conffile declarations and split ownership. Both sources and patches have fixed hashes. prepare() applies the patches to the release archive.

The candidate queues gdm_dbus_worker_call_set_session_name() before Open. It uses get_session_name(self) on the same worker connection. It does not change the resolver or program sessions. Explicit, saved and fallback choices keep their precedence. It does not force tuna or change registration. No root probe writes AccountsService state.

The proposed insertion, inside the existing non-null-conversation block immediately before gdm_dbus_worker_call_open(), is:

if (!self->is_program_session)
gdm_dbus_worker_call_set_session_name (conversation->worker_proxy,
get_session_name (self),
conversation->worker_cancellable,
NULL, NULL);

The approved local factory build produced both unsigned packages. Meson: one pass, zero failures. The staged verifier installed both packages. It checked exact versions, owners and ELF closure. Both containers enforced 8 GiB memory, zero swap and CPU quota 200%. We removed both containers.

The local archive hashes are:

PackageSHA256
gdm 51.1.1-1.1 x86_645fa15579210678f6168ea5102b79050503b768fd93476910855281f4b3d60d92
libgdm 51.1.1-1.1 x86_64ca4587fe886490c1e1925c9419bf5714fe9327cb9367685fb8d306b9a2243797

These checks do not prove login or physical record creation. The C smoke used actual functions with stubs. It checked explicit, saved, fallback and program choices. It does not prove GDM, PAM, D-Bus or AccountsService runtime behavior. We deleted its temporary source and binary.

Private native GDM fixture​

The CI job builds the pinned package source once. It uses the existing factory build and staged verifier. Both native profiles use the same run's artifact. The artifact has package hashes and the source identity. Those hashes need not equal the local hashes above. The manifest also binds the package members to their hashes.

The final gdm51 overlay runs after every pacman -Syu and Orca step. It checks both archive hashes and package metadata. It installs both packages in one transaction. It checks their versions, owners and installed bytes. It checks the existing shipped-payload manifest without changes. No package upgrade follows.

CI retains this evidence in vm-lane-artifacts/gdm51/. The final image ID and kernel provenance describe the image after the overlay. The fresh journey receives TUNA_VM_GDM_NATIVE_CANDIDATE and TUNA_VM_GDM_NATIVE_SOURCE_SHA. The first names the bounded manifest; the second pins the expected source. The journey checks the manifest and guest package identity before input. The existing read-only gdm-policy action runs once for this check.

Neither the manifest nor the package versions can satisfy the physical-record guard. All standard/admin, first/second login, sudo, logout and reboot gates stay intact. Candidate admission alone does not exercise either native journey. No source, function or package check counts as runtime acceptance. The source PR stays unmerged. This fixture does not sign packages, publish a repository or change production pins.

Actual candidate campaign 41255271531​

Run 41255271531 ran for desktop PR head 31bf6bb30ee9949918b161a3a6035e906c2efb79. CI uses a merge checkout, not the PR head directly. It built package source 7b52bdf34be22bc60fef62b157be64b589b13f6a once. The factory build and staged verifier passed. Both containers reported 8 GiB memory, zero swap and CPU quota 200%. Cleanup completed.

These are the CI hashes of the package archives, not the local hashes above:

PackageSHA256
gdm 51.1.1-1.1 x86_648c67e6be0e707f157cd901d319dd2af522f9666c352af42958d4a8bbaca62f66
libgdm 51.1.1-1.1 x86_64a559fae8e6d3166294c6d11ff76b39258ed0ca2c5e62f4e9ccdd5bc5d24f9af4

Both profiles installed that exact pair after all upgrades. Their metadata, owners and installed bytes matched its manifest. Both checks left the shipped-payload manifest unchanged. Both profiles passed the ordinary hardware-session lane.

ProfileFinal immutable image
distributionsha256:92de3d18015d68e81e4fc769154ac5f130bfa719f3ed70043a0fb8ca6cd90491
upstream Orca 51sha256:6a00f62d298ad8e3f79f5b9bfde0247f65976cc474a8d0c931f60c8f2a6ec3f4

Both profiles began with no physical record for UID1001. Both passed its first real gdm-password login. Both saw a physical record with Session=tuna during that session. That record also survived logout. The Orca profile passed the same first-record guard for UID1002. These results cover those logins, not the complete journey.

The Orca observer saw a SetSession call and its successful reply. It retained only headers. The distribution trace covered its first try. Its standard login succeeded on the second try, after the trace stopped. Missing headers do not prove a missing setter. Worker markers: unavailable; debug off.

Sudo refused the standard user in both profiles. The admin did not log in on the distribution profile after two tries. The admin login and record checks passed on the Orca profile. Sudo refused that wheel administrator by policy, with exit status 1. The evidence does not establish the inherited contents of sudoers. Second-login and reboot acceptance did not run.

Both fresh journeys remain failed, not qualified. Neither the persistence evidence nor the package provenance satisfies the missing gates. No production publication, image pin or merge follows from this run.

Actual candidate campaign 41262331704​

Run 41262331704 ran for PR #985 head 598ca553abc24f4b10bb335d22f99a5ff4e17a81. CI used merge commit a3d6334efd4176ff7082a6b4d1ddc5d9ad1fe1d3, with parents eed9aa95658f5d75496954b10142d50c807ee88f and 598ca553abc24f4b10bb335d22f99a5ff4e17a81. Package source stayed at 7b52bdf34be22bc60fef62b157be64b589b13f6a. The archives in CI had the same hashes as those in the previous run:

PackageSHA256
gdm 51.1.1-1.1 x86_648c67e6be0e707f157cd901d319dd2af522f9666c352af42958d4a8bbaca62f66
libgdm 51.1.1-1.1 x86_64a559fae8e6d3166294c6d11ff76b39258ed0ca2c5e62f4e9ccdd5bc5d24f9af4

For both profiles, real first logins passed for the standard user and administrator. Physical records held Session=tuna; read-only observations reported Saved=true. Both records survived the first logout. Standard sudo denial, administrator authorization, both manual second logins and all pre-reboot logouts passed.

Then the host reported literal timed out before V-FRESH-REBOOT. No actual new boot ID, post-boot retention or post-reboot login/logout received credit. Both fresh journeys failed.

The old CLI used systemctl reboot --no-block before it wrote JSON. Its host still awaited the child exit and output through the agent being torn down. That sequence can cause a teardown/reply race; the run does not identify the exact timed-out I/O or prove that the guest changed boots. The QGA request removes only that dependency; all gates stay intact.

The receipt for each profile contains nine cases for the strip. Both reported TimeoutError and cleanup_confirmed=true. This failure of a strip deadline is independent of policy binding. This run is not a PASS for the full suite. This run does not qualify reboot. No production publication, image pin or merge follows from it.

Actual candidate campaign 41267289849​

Run 41267289849 used PR #985 head c8554f32c87630ffc68afefcd0bb17e2d5f21ec3. CI used merge commit b50a3148e90e3d362bed2beb327b287ac281888a. Its parents are 017a7c6af924894f0ccd74f4d2bba46b3ac765f4 and that PR head. Package source stayed at 7b52bdf34be22bc60fef62b157be64b589b13f6a. The new factory build measured the same three archive hashes as the previous CI run.

Both native profiles passed all 24 unchanged gates for fresh accounts. Each profile has six login/logout cycles for UID 1001 and UID 1002. All sessions used gdm-password, Wayland VT2 and Tuna Desktop. During first login, the physical records existed before the read-only property queries. They had Session=tuna.

Sudo refused the standard user and authorized the administrator. Each account passed its manual second login.

The reboot proof includes actual distinct boot IDs and a ready GDM greeter:

ProfileBefore the requestAfter reboot
Distribution2eacd615-fc63-4dda-a7ab-09fbe78b83c0c7dafa91-fbde-451f-9a76-de691354a5b2
Orcadf390a84-f4e9-4bde-b0f1-9a7c59c5e0b2565b11aa-2ded-452e-8f37-a546bb4f7da6

Before the postboot login cycle, both accounts had physical records with Session=tuna. Both postboot login and logout cycles then passed. Each trace has a SetSession call for User1001 and a return with the same serial. There is no message body. Worker markers remain unavailable; no debug toggle occurred. Their absence gives no negative evidence.

The final images, after all upgrades and the verified pair installation, were:

ProfileImmutable image ID
Distributionsha256:226fd57b804f944c9f7d80dbb3e4756c7a679e166739517407ce1c6f58d13968
Orcasha256:503888154f04fd42a03ded3b1f85c6e2a056f5e5d8c971f12c0d0a1d24c39512

The package artifact is 13067948020. The distribution and Orca artifacts are 13068871310 and 13069002236. Public hashes from the initial guest boot matched the final overlay's payload seal. The probe did not read those policy files again after the fresh journey's reboot.

This qualifies the native fresh-account behavior, not the full CI run or production. Both ordinary strip journeys still failed at 10-column-height for beta. Their diagnostics show 21 tries and last reason fixture-color-mismatch. The local deadline of 10 seconds expired, with more than 81 seconds left on the total deadline. Cleanup passed.

No timeout, pixel, identity or acceptance guard changed. PR #985 remains blocked by the required strip gate; there is no merge or publication.

Bounded strip capture failure metadata​

Run 41262331704 placed the strip failure inside the capture for case 10. That run did not establish whether the local or total deadline caused the failure. The new diagnostic does not change either deadline or qualify a failed capture.

A failed capture adds one capture_failure object to the existing receipt. It contains the source-owned case label and role, plus the attempts count. Its fixed fields are:

  • capture_begin_monotonic_ns and failure_monotonic_ns;
  • local_deadline_monotonic_ns, capture_deadline_monotonic_ns and total_deadline_monotonic_ns;
  • stage: png-acquire, ppm-acquire, ppm-pixels, png-pixels, retry-refresh or post-frame-read;
  • last_not_ready_reason: fixture-body-not-visible, fixture-color-mismatch, other-not-ready or null;
  • ppm_rgb_verified and png_rgb_verified: separate booleans for the latest try. Null means that the capture did not reach the color check.

The object contains no frames, inventory, exception text or bus payload. The capture re-raises its original exception; cleanup stays unchanged. The local 10-second and total 120-second limits stay unchanged. Both pixel checks and the identity check after the frame remain mandatory.

The existing code does not guard the final write of receipt.json. An OSError from that write can mask an earlier failure. This diagnostic does not change that existing limit.

Normal notification preparation for strip proof​

The public frames show a notification over Beta. It covered two of the 25 color points in both PNG and PPM. The fix uses normal keyboard input, not a change to the pixel tolerance.

Strip preparation uses Super+N, observes key release, then uses Tab and observes release again. Before Return, a second observation must find one Notifications Close control. That control must belong to the original shell process and its AT-SPI bus. It must report FOCUSED. A layer namespace alone is not proof of that owner.

After the normal Close action, the banner must be absent and all keys released. Original windows, focus, workspace and session mode must remain unchanged. Both PNG and PPM still need all original 25 color checks. No pointer coordinates or empirical offsets select the control. There is no DND change, backend notification clear or authentication-dialog action.

The controller refuses preparation for held buttons, gestures, authentication UI or unknown or ambiguous overlays. All key commands share the unchanged 10-second capture and 120-second total deadlines.

Actual nested UI proof​

A private smoke of the nested UI used the exact package from run 41267289849, artifact tuna-arch / 13067334148. Its package hash and both executable hashes match the receipts from both native profiles. The original checks of 25 color points failed for each format with a test banner of critical urgency. The normal focused Close action removed it. The first post-close PNG and PPM pair passed the original checks. Window geometry, focus and mode stayed unchanged; post-frame identity and cleanup passed.

The smoke used separate actual PNG and PPM acquisitions through private Xvfb/scrot. It did not use native QMP acquisition or exercise an installed VM. There was no native compilation. Both required locks, the 40-GiB floor and the 8-GiB/no-swap/200-percent caps applied. Its HOME/XDG directories were fresh.

Focused checks passed: 57 acceptance/stationary tests and 20 existing navigation tests. The keyboard path has approval from two independent reviews. Receipt: 985-c855-public-notification-close/receipt.json in the retained finish evidence.

The keyboard path still needs qualification in CI with the actual installed image. The completed fresh-account proof above does not qualify that separate path.

Private image and input corrections​

The old controller used one or two forward Tabs. Each user has a focusable button. Hover can also move focus. The number of Tabs depends on the list and current focus. The retained run does not reveal where the UI had focus.

The corrected input parks the pointer outside the list. Both Escapes share the original two-second budget. Five Up presses reach the first item without vertical wrap. Two backward Tabs then reach Not listed? through the bottom accessibility button. The fixture has at most four human accounts and those two buttons. The old per-Tab sleeps are gone.

GNOME's focus handler lets backward Tabs wrap, but not Up by default. The controller still allows at most two login tries. Its 60-second login deadline and 90-second trace bound stay unchanged. Real UI qualification still needs the complete native journeys.

The pinned installer gives administrators the wheel group. Its user setup does not grant sudo authorization. The result shows a gap in authorization for the administrator. It does not establish the inherited sudoers bytes.

The shared image uses tuna-vm-admin-policy for its global wheel role. Its rule is %wheel ALL=(ALL:ALL) PASSWD: ALL. It names no test account or probe command. The installed file has owner 0:0 and mode 0440. The account helper and sudo probe remain unchanged. Standard accounts outside wheel must still fail by policy.

An explicit include precedes inherited rules. The .conf suffix prevents a second include through @includedir. This order keeps exceptions for specific commands, such as bootc upgrade. The image uses real visudo to check syntax and effective inclusion.

The cvtsudoers tree proves the role is first and unique. It rejects alternate or indirect duplicate includes. Setup keeps one explicit include. Tests parse the actual role with sudo's tools, without PAM or privileges.

The image appends hashes for the policy and root sudoers file. It does not regenerate the existing GDM payload entries. Later Orca and GDM layers must preserve that complete receipt. These source and parser checks do not qualify the runtime gates.

Why a first login need not create a cache file​

PR #931 incorrectly required a physical cache file after the first login. The direct GNOME reference receipt boots the same pinned published Marlin GNOME base as failed diagnostic run 41223633281: ghcr.io/tuna-os/marlin:gnome@sha256:8d71660f78120a141e9ae8e620bd66645c6bfb52f1ab4cd663f9672aaa0b0b94. It ran on an owned Corral VM on karnataka, without a Tuna package or the VM-lane layer. useradd -m created separate standard and wheel accounts; neither had a cache record, .dmrc, preselected session or autologin. Both first successful GDM password logins started stock GNOME Shell 51.1 with gnome-session 51.0, GDM 51.1.1 and AccountsService 26.27.3, in active Wayland gdm-password sessions on VT2. Both cache files remained absent; AccountsService exported empty Session and Saved=false. The monitored system bus contained lookups but no SetSession or CacheUser calls before the explicitly labelled post-observation intervention. The receipt retains actual package/template/session identities, navigation distinctions and successful VM/disk deletion plus exact-name inventory absence.

This matches upstream, rather than a missing Tuna session-start method:

  • AccountsService 26.27.3 user_update_from_template applies the template to in-memory/exported properties without saving a file. Its user_change_session_authorized_cb writes extra data only when the normalized new session differs from the current property.
  • GDM 51.1.1 session settings explicitly loads properties even without saved state. Its gdm_session_settings_save only sends non-NULL session/type/language values, and caches only non-local users. Loading a local account is not CacheUser. Tuna's existing templates supply Session=tuna; setting that same value is not a reason for AccountsService to materialize a cache file.
  • GNOME Shell 51.1 can request SelectSession when the session menu's active item changes; this is not a guarantee that every default login changes an account property. Its user/session UI reads AccountsService users; the observed GNOME login does not cache them. gnome-session 51.0 has no AccountsService mutation API.
  • GNOME Initial Setup is a different account-creation/property-writing path, not installed in this Marlin reference. Its 51.0 account page calls CreateUser and sets an icon; its language page writes the language. AccountsService's CreateUser path calls cache_user, unlike an installer using useradd. A record produced there must not be misattributed to the first ordinary GDM login.

The corrected proof does not write a record, call SetSession, preseed a choice or add product-side cache writes. It reads the effective AccountsService session with FindUserByName/Get(Session) and requires tuna after every login and after reboot, while still rejecting a saved GNOME choice. The record-free preflight, seeded negative control, real authentication, standard/admin sudo policy, logout and reboot remain gates.

Original #931 merge audit​

#931 did not have green exact-head CI. Head d62ec870c0f402c22fbae6719afd227cd0353b81 was merged by hanthor at 2026-10-31T15:42:38Z as cba3cdbd5cd629d42c2512124f3fe7e35c117b55, while its exact-head run 41215443626 was unfinished. check failed at 16:00:23Z, marlin-vm at 16:27:14Z, marlin-vm-orca51 at 16:30:35Z and ci-gate at 18:14:39Z. An earlier PR comment already documented the missing-record failure as a merge blocker. At investigation time the branch protection endpoint returned “Branch not protected”; the effective organization ruleset required only deletion/non-fast-forward protection, not successful status checks. These observations explain why an ordinary merge was possible; they do not prove an administrator bypass was used.

The Debian package journey must retain the original beta package's share/pam.d/tuna-lock conffile bytes, including its comment. Run 41151873648's journey failed on dpkg's conffile prompt before the upgrade could qualify. PR #930 restores those bytes; the fresh-account change must include that fix, not force a conffile answer or change PAM policy to bypass the upgrade failure. The Python lane.boot API keeps its original four-value return and original VM device profile by default. Lifecycle callers explicitly request guest_agent=True to add the QGA transport and receive its socket as a fifth value; existing performance callers keep working without a guest agent.

Paired stock-GNOME switcher reference (P-WM01)​

The independent DEV profiles --profile gnome-reference and --profile tuna-reference use the existing scripts/tuna-vm-lane.boot(), raw-disk installer, QMP/QGA transport, native KVM requirement and unchanged 1280×800 GPU topology. They do not replace either ordinary five-boot Tuna lifecycle/Orca row, its tour, fresh-account journey or any existing gate. Whenever lane selection includes marlin-vm, both paired DEV members and their artifact-context consumer are required by the normal aggregate gate. There is no dispatch-only, repository-variable or credential skip for them. Neither profile is a shipping image or a performance-lane substitute.

packaging/marlin/vm-lane/Containerfile.reference derives both members from the actual published GNOME 51.1 amd64 manifest sha256:104870f1a05fa0a75649ee4ed9952027b29ab12e53af83931ae69e8636e44c3f, which belongs to index sha256:29097ba660d6f9f7c35c2063b83cf7bbe830202a497852b9b1869bb84dfd6aa5 (tunaOS publication 41254872557, source 07022c9fb853fcd36aa27098d50c38a8be0d18ac). Both add the same current-source Tuna package and fixture dependencies using the image's retained package databases, never a rolling repository refresh. The build retains the original package inventory and GNOME/GDM/session/PAM/application file hashes, then refuses any original package-version or file-byte change. Only the CI fixture's tuna-test UID 1000 AccountsService session selection differs: gnome versus tuna. No automatic login, GNOME keybinding write, GDM theme replacement, hidden-session rewrite or Shell.Eval admission is involved.

Before either image build or boot, the normal CI path must verify the exact platform digest under the existing TunaOS cosign subject/issuer policy and retain its real platform-cosign.json output, actual platform-verification.json process witness and exact index.json bytes, including the unique linux/amd64 membership. The witness binds the pinned Cosign 3.1.4 executable/version, exact subject/issuer-constrained argv, actual exit status, command bounds, stderr and original stdout hash. Its successful cryptographic invocation remains mandatory; the consumer never treats copied JSON as signature verification. Modern Sigstore output names the exact repository-at-platform digest and https://sigstore.dev/cosign/sign/v1; its empty optional object is not fabricated into extracted certificate claims. Supply all three files as --reference-base-evidence. The published wheel-policy receipt identifies the image but does not itself contain a signature-verification result. CI acquires the pinned cosign binary through the existing Sigstore installer action; it does not assume Go is installed on the native runner.

The ordinary platform path uses the same verifier via scripts/lib/vm_gnome_reference.py verify-platform --platform-digest sha256:… --out DIR: it retains the two platform files for the supplied strict digest in ghcr.io/tuna-os/marlin, with one actual verification invocation and the same certificate policy. The shared validate_platform_verification consumer binds that modern envelope to its process witness without imposing the stock index. In contrast, verify-base --out DIR and validate_base_bytes remain strict stock-reference operations: they additionally require exact index29097 bytes and unique platform104870 linux/amd64 membership. A different ordinary platform witness cannot qualify the paired stock reference.

The fixed route GuestAgent.run("gnome-reference", OP, JSON) executes only /usr/libexec/tuna-vm-gnome-reference OP JSON; OP is one of admit, prepare, observe, file or stop, with the existing transport deadline capped at 30 seconds. Original stock-GDM and logind/native session admission comes from this observer. The GNOME path branches before all Tuna DRM-log and TUNA_COMPOSITOR_STATE readiness loops, without a Tuna fallback. After password submission it waits for the same session's actual compositor, bus and DRM admission within the existing login deadline. A partially started session never causes another password to be typed into the desktop. If that unchanged greeter deadline expires, greeter-before-login.json retains the last actual admission observation and latest transport error, including missing original-principal or payload-integrity prerequisites. The diagnostic is failure evidence, never admission or a deadline extension. The observer queries each logind session property with its own --property= option, as the existing lifecycle probe does. loginctl treats a comma-joined value as one property name and returns no properties; it is not systemctl's multi-property syntax. Original active-VT, native package and session-owner admission stay required; an empty observation is never accepted.

The paired consumer owns its real two Files windows and one Settings window, original process/ELF identities, read-only settings, public AT-SPI snapshots, physical QMP held-key/Down actions and bracketed framebuffer captures. wm-reference/member.json and the paired context receipt describe those actual observations. AT-SPI bus/object identities are not WM or kernel window identities. Context readiness is not a visual-parity verdict and does not close P-WM01; the actual paired visual assessment remains required. Source changes and controlled transport tests establish no native runtime pass. Missing admission, evidence, held-phase observations or strict owned cleanup must fail rather than silently skip a member.

Native lifecycle observations​

QEMU's guest agent executes a fixed probe with a root UID, while the compositor, shell and application clients must belong to the non-root session owner. Artifacts record window IDs, app IDs, rectangles, process PIDs/start ticks and redacted session metadata. They contain no process arguments, passwords or application titles.

The last of the existing five pristine boots runs these checks after the feature tour:

GateActual operation and evidence
V-LIFECYCLE-OWNERRoot probe, non-root gdm-password login session, three application surfaces/processes
V-NORMAL-INPUT / V-NORMAL-INPUT-FAIL-CLOSEDA real GTK client receives ordinary QMP keys while unlocked, then receives none while locked; artifacts contain only counts and PID, never key values
V-MOUSE-HANDEDNESSNine physical QMP clicks pass through native libinput to the same non-root GTK client while the real GNOME mouse setting changes false→true→false; left/right swap, middle stays unchanged, exact before/after counters retained in lifecycle/mouse-handedness.json
V-POINTER-LIVEGNOME mouse speed/natural-scroll, acceleration profile (flat, then adaptive), middle-click emulation (on, then off) and scroll-wheel emulation (button 2 → on-button-down with BTN_MIDDLE and button lock, then off) reach live libinput device state and restore, asserted on the compositor's per-device readback. Each option needs at least one mouse offering it; wheel emulation is asserted on mice offering button scrolling and BTN_MIDDLE in pointer_buttons (libinput rejects a button the device lacks). Both source-built and published guests must report profile/button state and provide GNOME 51's wheel-emulation schema; absence fails rather than skipping proof. Snapshots: lifecycle/pointer-live.json (#343)
V-POINTER-HOTPLUGA QMP-hotplugged USB mouse arrives with the current settings already applied (DRM DeviceAdded path); unplug drops it while survivors keep theirs; bus discovery, arrival and survivor snapshots retained in lifecycle/pointer-hotplug.json (#343)
V-VTVT away/back, DRM pause/activate events, unchanged process identities and app rectangles, matching body plus a fresh visible count update from the real GTK client
V-SLEEP-DELAY-INHIBITOROn the unlocked desktop the compositor's snapshot records its logind sleep delay inhibitor held, and logind's inhibitor list (systemd-inhibit --list) shows a delay inhibitor on sleep from "Tuna Desktop" for the session user's compositor PID, as GNOME 51's screen shield holds one while unlocked (lifecycle/sleep-delay-before.json)
V-SUSPEND-UNLOCKED (with -GPU-PM / -GPU-UPLOAD)The first real suspend starts from the unlocked desktop; the same checks as V-SUSPEND, so the first snapshot after wake must already be locked
V-SUSPEND-LOCKED-FRAMEAcross that suspend the compositor's sleep_delay record shows exactly one sleep whose inhibitor it released only once every lit output had scanned out a frame drawn locked, no release without that, no overrun of logind's InhibitDelayMaxUSec, and the inhibitor taken again after wake (lifecycle/sleep-delay-after.json). QEMU's display while suspended is logged as a non-gating observation
V-SUSPENDThe second real suspend, from a locked session: real logind suspend reaches QEMU suspended; the first guest snapshot after system_wakeup must already be locked, with all application identities intact; the real varied lock mask must repaint within five seconds, matching the pre-suspend mask; QEMU inactive-output placeholders fail
V-SLEEP-AUTH-FAIL-CLOSED / V-SLEEP-FRESH-AUTHThe second real suspend interrupts a correct password attempt delayed by actual pam_exec; the real PAM success from the old generation is refused, then fresh correct authentication succeeds
V-SLEEP-CLIENT-REPAINTThe same GTK client receives a fresh key and visibly repaints after wake/unlock
V-VT-FAIL-CLOSEDVT away/back while locked keeps the mask and original application processes
V-AUTH-FAIL-CLOSEDReal PAM service temporarily uses pam_deny; fresh submitted/refused milestones prove the correct test password was attempted and denied; original service restored afterward
V-CAPTURE-INPUT-FAIL-CLOSEDReachable real untrusted grim/wtype clients are denied capture/virtual-keyboard interfaces while locked; a successful Wayland connection is required and timeouts do not count as refusal; this does not assert that these interfaces are available while unlocked
V-SHELL-FAIL-CLOSEDStop the portal and kill the lock UI; original app/compositor identities survive and the supervised replacement stays locked
V-GSD-PLUGINSEach of the thirteen gnome-settings-daemon plugins the session target wants, MediaKeys included, is the packaged gsd-* binary, running as the session user as its unit's main PID and owning its bus name; org.gnome.SettingsDaemon.Color belongs to the compositor; A11ySettings has no owner; the session targets are active and the compositor holds gnome-session-leader-fifo (lifecycle/gsd-plugins.json, #355, #589)
V-GSD-POWER-IDLEWith idle-delay set to 20 s and no input, dbus-monitor records gsd-power's own SetDimming(true) before its PowerSaveMode=3; the compositor snapshot reports locked and power_save_mode 3; Return wakes it to 0 and gsd-power undims; the original delay is restored (lifecycle/gsd-power-idle.json, #355)
V-LOCKED-HINTlogind's LockedHint for the login session reads no; the session user's loginctl lock-session locks the compositor and the hint turns yes; the real password unlocks and it reads no; Super+L locks with yes; loginctl unlock-session unlocks with no, as GNOME 51's screenShield.js does (lifecycle/locked-hint.json, #846). Required on both acquisitions
V-POWER-BUTTONgsd-media-keys (its unit's main PID, the session user) holds logind's block inhibitor on handle-power-key:handle-suspend-key:handle-hibernate-key with Why "GNOME handling keypresses". On a fixture desktop chassis, after each gsd restart a new gsd-media-keys must own its name, hold that inhibitor, have XF86PowerOff granted by GrabAccelerators and the shell must log that it yielded the bindings. QMP system_powerdown (ACPI power button) with power-button-action interactive: AcceleratorActivated reaches gsd, gsd calls org.gnome.SessionManager.Shutdown, the compositor calls EndSessionDialog.Open(1, …) and the tuna-shell-end-session layer shows within 15 s; 5 s after the press QEMU is running with the same compositor process and logind session and no logind power request; Escape (always sent, far inside the 60 s countdown) removes the layer, the shell emits Canceled and logs its D-Bus Power Off dialog. nothing: gsd receives the key, and for at least 8 s no Shutdown, dialog call, layer or logind request follows. On the image's vm chassis with a fixture shutdown block inhibitor held by root: the key reaches gsd and gsd's own process calls logind PowerOffWithFlags(0); gsd sends it without a reply callback (NO_REPLY_EXPECTED), so logind's refusal has no bus reply, and 5 s later logind has emitted no PrepareForShutdown/PrepareForSleep, PreparingForShutdown is still false with the fixture inhibitor held, no other logind request or non-error reply followed, no dialog opened and QEMU, compositor and session are unchanged. Setting, chassis and /etc/machine-info are restored (lifecycle/power-button.json, #353, #589)
V-RESTART-DIALOGBoth current-build and published guests must exercise the shell's own power menu by physical keys on the unlocked desktop after V-POWER-BUTTON (#589): Super+S opens quick settings; physical Tab moves until the shell's AT-SPI tree (tuna-vm-shell-focus end-session mode, observation only) reports Power Off Menu focused, Return opens the menu with Suspend, Restart…, Power Off… and Log Out… showing, and Tab/Return chooses Restart…. Within 15 s the shell (the org.gnome.Shell owner, not the compositor) has called org.gnome.SessionManager.Reboot and the compositor's manager has answered with the shell's EndSessionDialog.Open(2, 0, 60, …), with the tuna-shell-end-session layer up; the dialog reads "The system will restart automatically in 60 seconds" with Cancel focused, as GNOME 51's endSessionDialog.js. It steps to 50 seconds with QEMU running, the session and compositor identity unchanged, no logind power request, PrepareFor* signal or PreparingForShutdown, and the request still unanswered; Return on the focused Cancel then emits the shell's Canceled, fails Reboot with gnome-session's G_IO_ERROR_CANCELLED and takes the layer down, still with logind untouched. Any failure presses Escape so no countdown is left running (lifecycle/end-session-menu.json). An older published shell does not waive the gate
V-POWER-OFF-DIALOGAs V-RESTART-DIALOG for Power Off…: org.gnome.SessionManager.Shutdown, Open(1, 0, 60, …), "The system will power off automatically in 60 seconds" stepping to 50, Cancel by Return, Shutdown failed as cancelled, logind untouched
V-LOGOUTBoth lanes use the shell's power menu by physical keys (#589), as V-RESTART-DIALOG runs Restart…: the shell must ask org.gnome.SessionManager.Logout(0), the compositor's manager must answer it once its Open(0, 0, 60, …) has put the dialog up, counting down from 60 with Cancel focused, and the session must outlive that countdown, whose expiry confirms the logout (the greeter may only return at least 55 s after the dialog opened). Within 30 s of the logout taking effect (90 s from the menu dialog's opening) GDM's non-root gnome-shell greeter in its active gdm-launch-environment session repaints the greeter frame captured before login and no original compositor survives. A logind Session.Terminate substitute or historical published build does not pass

The original tour's V-LOCK still requires exactly unlocked → locked → unlocked. Its observations end before lifecycle operations add their own lock transitions. A QEMU pause is never a substitute for guest suspend. The repaint comparison excludes the panel clock and tolerates up to 20% changed body pixels for applications such as System Monitor; it also requires exact window and process identity preservation.

V-POWER-BUTTON runs on the unlocked desktop after V-SHELL-FAIL-CLOSED and before V-LOGOUT. gnome-settings-daemon 51 reads hostnamed's Chassis once at startup and, on vm, powers off for every power-button-action but nothing. QEMU guests report vm, so the fixture (root, through QGA) records /etc/machine-info and Chassis, calls hostnamed SetChassis with desktop and restarts org.gnome.SettingsDaemon.MediaKeys.target as the session user. The service refuses manual start and stop, but a restart of its PartOf target propagates; if refused, the fixture kills the main PID and Restart=on-failure brings it back. Afterwards the fixture restores the recorded chassis and file bytes and restarts gsd again, so the rest of the lane sees the image default. The vm case stays safe because gsd 51 sends PowerOffWithFlags without SD_LOGIND_SKIP_INHIBITORS, and since systemd 257 logind refuses that request outright under a block inhibitor. The fixture refuses to press the button on an older logind. power-button-action suspend is not exercised. It would add a second real S3 cycle, and the user-authorized single-S3 contract below forbids that. hibernate, the lock-screen mode, and an unrestrained VM power-off are not exercised either.

The compositor also listens to the trusted logind system-bus sleep signal: preparation invalidates pending authentication generations and engages its own lock; wake restores KMS connector/plane state and retires lost flip buffers without marking them presented. After the blocking device reset, bounded nonblocking reads drain obsolete kernel completions before any new frame is submitted. A valid modeset completion can carry the preceding vblank's timestamp, so timestamps alone do not identify abandoned frames. A failed reset or drain keeps scanout blocked until recovery succeeds. Wake recovery does not depend on a VT change; an off-seat wake waits for seat activation.

While the session is active and the lock screen is allowed, the compositor holds a logind delay inhibitor on sleep, as GNOME 51's screenShield.js _syncInhibitor does. On PrepareForSleep(true) it locks and closes the inhibitor only once every lit output has completed the page flip of a frame drawn locked, so the frame left on screen through sleep is the lock. It takes the inhibitor again on PrepareForSleep(false). When logind's InhibitDelayMaxUSec runs out first, logind sleeps anyway and the session is still locked; the record counts an overrun. A lockdown that disables the lock screen, or a session off its VT, releases the inhibitor at once.

These new gates require their own actual CI results before acceptance. Output hotplug and loss of every compositor-adjacent service remain unproven, as do physical hardware and 24-hour soak. This document does not claim those parts of issue #62 are complete.

Protocol references: QEMU guest agent and QEMU wake-up.

The shared GPU profile virtio-vga-pcie-pm-preserved-v1 keeps the same virtio-vga, resolution, CPU, RAM and workloads. It places that GPU behind one PCIe root port and advertises PM no-soft-reset. All native GNOME capture and Tuna Desktop/performance boot callers use the same GPU arguments; manifest/report fields record them. Fresh GNOME comparison evidence must use this profile; old-profile latency or cadence does not qualify it. The actual guest PCI capability chain must include PCIe, PM and the no-soft-reset bit, and each real resume must produce fresh host resource creation, backing attachment and nonzero scanout commands as well as the painted lock mask. A configured flag alone never passes the gate. The shipping AWS Bochs GPU is a separate, unmeasured suspend profile.

The diagnosed fixture runs kernel 7.2.8-arch1-2 and QEMU 10.2.1. In Linux 7.2.8 virtio GPU the driver has no freeze/restore callbacks. The PCI PM path skips reset when the device advertises no-soft-reset; otherwise virtio restore resets the device. QEMU 10.2 virtio PCI provides that capability only behind a PCIe port and preserves the device in D3hot. A GPU reset destroys its resources and clears the display. The prior actual trial 37211055978/111464650400 reached real S3, stayed locked with apps intact and processed guest pageflips, but showed inactive host output and no fresh resource uploads. It remains a failed qualification. The supported PM profile requires its own strict actual run; no kernel/package override, extra VM matrix cell or physical hardware claim is introduced. QEMU 10.2 is required, so the existing performance runner uses ubuntu-26.04.

The performance lane keeps the user-authorized single-S3 contract: the session is prelocked and a real PAM attempt is pending before sleep. Its fail-closed wake and stale-success/fresh-authentication checks share that cycle. Varied mask repaint is retained as an observation; fresh host GPU uploads, client input/repaint and unchanged application identity remain mandatory. The prior native lifecycle qualification (#265) separately demonstrated locking from an unlocked session and two real suspend cycles on the preserved-resource VM profile; that evidence does not certify a later candidate automatically.

Wake-time process inventory classification​

Run 41490837174, job 125453770508, on candidate b8feffef reached real S3 and fresh host GPU resource creation, backing attachment and nonzero scanout, but failed V-SUSPEND. Original artifact 13170520388 has SHA256 a30fbd6ab110c3a863fa92ec079bf47c38a11646f9062ea5b989bb17299f6206. Its lifecycle/first-resume.json is already locked and preserves the ordered window records and all five original compositor/shell/application process identities. The failing identity comparison instead sees one additional tuna-compositor process, PID 33679, start ticks 60527; that process is absent from the later repaint inventory. The exact retained PPM body changes are 800742/942080 pixels on first resume and 800653/942080 on the last locked frame, both above the unchanged 25% fail-closed threshold.

The probe formerly classified private PAM helpers in one /proc enumeration and then enumerated owned processes separately. A helper born between those enumerations could therefore be counted as another desktop compositor. [INFERENCE] The transient PID is a fresh wake-time PAM helper: the serial trace records new password/fingerprint conversations in that interval, but the failed artifact does not retain that PID's exact helper flag, parent, process group or executable inode. It cannot prove that classification retrospectively.

Inventory now reads each scanned compositor's fixed private-helper metadata. Exclusion requires the exact private argument shape, live helper state, original user's UID, direct parent, own process group, start-time ordering and matching executable device/inode. Its parent is reread and must still match the unique initial desktop identity, including start ticks; neither a numeric PID match nor a reused parent PID is sufficient. Unrelated or impostor compositor processes remain identity mismatches. Controlled /proc regressions cover a helper born between enumerations, unrelated parent/group, older start time, wrong arguments/inode, ambiguous desktop identity and PID reuse. No application/window comparison, frame threshold, GPU requirement, suspend deadline or first-resume lock requirement changes.

The failed run remains failed. A fresh owner-authorized native qualification with the corrected installed lifecycle probe must still demonstrate the actual first locked wake, exact application geometry/process preservation, fresh GPU uploads, stale-auth retirement, fresh authentication and client repaint; fixture regressions alone cannot qualify suspend.

Wake-time guest-agent transport (guest Linux 7.3)​

PR runs for #1083, #1089, #1090 and #1092 (e.g. 43385429967) passed V-SUSPEND-GPU-PM and reached real S3. The compositor reset scanout and repainted after wake. The run then failed V-LIFECYCLE with a bare socket timed out. The guest's qemu-ga logged guest-exec ... inventory at 501.22 s and no guest-exec-status afterwards. Its next request came at 530.45 s, on the reconnected diagnostics transport. The host therefore never received the guest-exec acknowledgement. The inventory probe itself did not block. The green main run 43369611171 polled status 42 ms after the same request.

The difference is the guest kernel, not the PRs. pacman -Syu during the image build moved the guest from linux 7.2.15 (main) to 7.3.4 (all four PRs). Upstream commit 952261953cf7 ("virtio_console: fix control queue race during restore", in 7.3.4 but not 7.2.15) changed virtcons_restore() to fill the control receive queue only after sending PORT_READY. QEMU 10.2 answers PORT_READY with the host-open event before it returns the control-out buffer. send_control_msg() drops that event when the receive queue is empty. Freeze had set the guest port to host-disconnected, so after restore every qemu-ga write returns EAGAIN until the host chardev reconnects.

After system_wakeup, the lane now calls GuestAgent.reopen_after_wake (2 s per connection, 20 s total). It closes and reopens the QGA chardev, which makes QEMU send PORT_OPEN 0/1 again, until one guest-sync-delimited round trip completes. The step runs no guest-exec and reads no guest state, so the following inventory is still the first state observation after wake. The number of connections and the elapsed time go to lifecycle/wake-transport.json. The inventory keeps its 30 s deadline, and no suspend gate changes. A lost acknowledgement or status reply now names its QGA step instead of a bare timed out. This is controlled-socket evidence only. A fresh actual run must still qualify V-SUSPEND on the 7.3 guest.

Date & Time restoration and published capability failures​

Published-flavor run 41151873652 observed the new Auckland panel time, then restored timedated to UTC without observing the panel's UTC repaint. G-DATETIME-CLOCK-SET immediately read the Auckland label after the manual clock jump (39,610 seconds from UTC), so that assertion did not have a restored-panel baseline.

The timezone round trip now retains panel_revert dumps and panel_revert_verified in lifecycle/datetime-evidence.json, waiting within the existing ten-second repaint deadline. The clock timezone gate requires both outbound and return repaint. A failed restoration or dump remains visible in the probe error, fails the clock timezone gate, and prevents the manual-time probe from changing the clock against an unverified baseline. The polkit timezone probe runs after the manual-clock probe, so its own timezone round trip cannot invalidate that baseline.

Both published panel-change assertions stay strict. A released image that cannot perform them is reported as unqualified/unsupported with its actual failure evidence, never converted to a pass or credited with another lane's result. No additional wait extends the capability deadline.

Locked password speech​

Run 41151873648, job 124396764753 passed eight speech gates, then failed G-ORCA-SPEECH-LOCK: the retained orca-speech artifact 13025680053 contains Lock Screen, Password, password text. followed by an empty lock-type speech section. Its sanitized state records an active password conversation, locked: true, and an owned, watched KeyboardMonitor. The AT-SPI registry was already active. This is not evidence of bus startup failure.

The missing route was committed masked input to speech, not raw-key delivery. The compositor then suppressed raw KeyboardMonitor events while locked; it no longer does (see below). In the installed Orca 50.3 insertion classifier, _get_insertion_text_key_reason relies on a last printable key/space to classify password insertion as typing echo. Without that cue, _get_insertion_fallback_reason can return UNKNOWN or an unpresentable auto-insertion reason; the default insertion handler does not speak those insertions. The retained artifact does not contain Orca's internal event-reason trace, so no particular fallback branch is claimed as an observed runtime result.

The native Orca consumer needs a package-level correction for the supported 50.3 and 51 versions so committed masked password insertions are presented without raw locked-session keys. That correction and its package/runtime qualification are separate and pending; this change does not claim the masked-input speech failure is fixed. No shell-generated mask announcement or replacement for Orca's native text-event handling is included.

scripts/tuna-orca-speech-proof --artifacts orca-speech-artifacts remains the end-to-end native nested proof: after a D-Bus lock and after an ordinary Super+L it physically types into the lock prompt, requires exactly one spoken mask per typed character, rejects plaintext speech, requires the owned trusted reader's KeyboardMonitor to receive the lock-screen keys as Mutter 51.1 does (screen reader keys while locked), then unlocks through PAM. No injected Orca customization or relaxation of the bullet assertion is used. The failed run is not retroactively qualified.

The native proof isolates XDG_STATE_HOME and XDG_CACHE_HOME under its artifact root as well as HOME/config/data. A local native attempt stopped before the speech gates because inherited state selected ~/.local/state/roost-shell for upgrade, where the ownership/ancestor guard refused migration. user_upgrade::pairs_from and the shell's xdg::state_home_from prefer an absolute XDG_STATE_HOME over HOME; changing HOME alone does not isolate that route. The proof now exports private state/cache roots explicitly. It does not touch the real user paths or bypass the migration guard.

The first local attempt used host Orca 46.1, which could not satisfy the KeyboardMonitor contract. A later isolated Ubuntu 26.04 run with native Orca 50.3-0ubuntu0.1+tuna1 passed all 11 speech assertions in 141 seconds. It typed nine characters, observed nine spoken bullets without plaintext, unlocked through PAM, and confirmed that the toggle stopped Orca. The retained report records source 69e9a04fd20aa2bc38427c79aec043cb3174da3b and the actual package and binary identities. This proves the nested speech path, not the separate GDM fresh-account journey or ISO installation.