Skip to main content

v1 variant qualification

No all-variant v1 qualification is claimed. Package publication, an enabled image build, and an image boot gate are not real-GDM/session/portal/lock proof. The v1 checklist and ADR 0008 remain unchanged. Marlin remains the GNOME 51 comparison target; the other variants still need their own immutable-image runtime evidence.

Owner decisions, 2026-11-14​

The owner decision of 2026-11-14, as revised the same day on #1050 after two of its premises were corrected, sets the v1 required set to 15 variants: all fourteen matrix variants plus stingray. That is yellowfin, albacore, skipjack, wahoo, bonito, bonito-rawhide, sailfin, grouper, marlin, flounder, flounder-sid, hummingbird, guppy, gurnard and stingray. None of them is qualified yet.

  • gurnard: required. Noble publishes no gtk4-layer-shell source (Launchpad query, no entries on 2026-11-14), and there is no noble factory target yet (packages #834). The factory will add one: gtk4-layer-shell 1.1 or later packaged for Noble, built with Noble's own versioned rustc-1.97/cargo-1.97. That compiler closes the earlier gap: the locked gtk4 0.11.5 needs Rust 1.92 or later, and rustc-1.97 and cargo-1.97 1.97.1 reached noble-updates and noble-security on 2026-11-06. TunaOS keeps gurnard unsupported until the target publishes.
  • stingray: required, added by the owner on 2026-11-14. It needs an ubuntu-rolling factory target, a tuna flavor and runtime runs.
  • hummingbird: required; closure work in progress. The TuneD rpm-ostree provider, libcanberra and build-chain publication of the current desktop (packages #982, #959) come first, then the image build is re-enabled.
  • guppy: publication approved under #577, pending qualification. TunaOS #3883 merges once its runtime run qualifies.
  • albacore: required; deferral lifted. tunaOS #3969 was closed as completed on 2026-11-01, and the owner lifted the deferral on 2026-11-14. Its tuna cell keeps building and publishing, and it is to be qualified like the others.

Current source inventory​

The primary inventory is TunaOS tunaos.toml at 124cc32623ce084580a2ef888350be07bd83508d. It declares 14 variants, eleven admitted amd64 Tuna image cells and three source-declared unsupported rows. No arm64 Tuna cell is enabled. All eleven Tuna cells have build_iso=false, build_qcow2=false, and luks_e2e=false; this matrix does not invent ISO/LUKS acceptance cells. TunaOS main has since added stingray as a fifteenth variant, recorded unsupported until the factory has an ubuntu-rolling target.

The native package inventory is tunaos-packages f853bc21d8316cb20bcfef42f4cc6d60ef7ec7d6 (tunaos-packages#1009, 2026-11-14). Its Tuna recipe/spec target beta.3 release 1, source desktop tag v1.0.0-beta.3, commit 721c55033dd5d2a1127f24473542e16646cb34e5. The beta.2 release 2 recipe (tunaos-packages c0998b94, tag v1.0.0-beta.2, commit 1e6ce28f) is what the 2026-10-31 rows below were measured on. That is not current desktop main. Recipe and vendor archive digests, actual Gentoo payload evidence, and publication attempts are recorded in package-evidence.json.

VariantBasePackage target / formatSupport and integration state
yellowfinAlmaLinux Kitten 10el10 / RPMadmitted; published image, not v1-qualified
albacoreAlmaLinux 10el10 / RPMadmitted; published image, not v1-qualified; required (2026-11-14), owner deferral lifted after tunaOS #3969 closed
skipjackCentOS Stream 10el10 / RPMadmitted; published image, not v1-qualified
wahooFedora ELN / EL11 previeweln / RPMadmitted, experimental; missing Tuna tag repaired through existing signed workflow; runtime still required
bonitoFedora 44fedora / RPMadmitted; published image, not v1-qualified
bonito-rawhideFedora Rawhiderawhide / RPMadmitted; published as bonito:tuna-rawhide, not a separate registry repository
sailfinopenSUSE Tumbleweedrollout label opensuse, recipe target opensuse-tumbleweed / RPMadmitted; published image, not v1-qualified
grouperUbuntu 26.04 Resoluteubuntu / DEBadmitted; source retains a portal session-start blocker note; needs actual digest-specific verdict
marlinArch Linuxarch / pacmanadmitted; real published-image journey has failures, below
flounderDebian 13 Trixietrixie / DEBadmitted; source retains a portal session-start blocker note; needs actual digest-specific verdict
flounder-sidDebian Siddebian / DEBadmitted; published as flounder:tuna-sid; same portal-note qualification limit
hummingbirdFedora Hummingbirdnative Hummingbird build chain / RPMrequired (2026-11-14); closure work in progress; not admitted on main; native package implementation exists; current served beta1 and power-provider closure block strict admission
guppyGentoogentoo / gpkg, amd64required (2026-11-14); publication approved under #577, pending qualification; not admitted on main; signed beta2-r1 publication and served-install verification pass; image admission/runtime remain open
gurnardUbuntu 24.04 Noblenoble / no implemented Tuna targetrequired (2026-11-14); unsupported on main until the factory gains a noble target with gtk4-layer-shell and rustc-1.97
stingrayUbuntu interim releaseubuntu-rolling / no implemented Tuna targetrequired (2026-11-14); unsupported on main until the factory gains an ubuntu-rolling target; flavor and runtime runs remain

Unsupported reasons are evidence, not a scope shortcut​

TunaOS main still says Hummingbird/Guppy await Tuna package publication. Those reasons are stale: both have native implementations, and Gentoo's new revision was actually published during this qualification. They must not be counted as satisfied unsupported exceptions merely because the old strings remain in TOML.

Existing admission branches were repaired rather than creating competing PRs:

  • TunaOS #3883: current-main integration, brltty[python] closure for Orca, and exact published =gui-wm/tuna-desktop-1.0.0_beta2-r1::tunaos pin. Signed package run 41213265948 passed build, publication and pristine served-binhost installation checks. Direct payload inspection finds the six current binaries without tuna-greeter. This is not a boot/login qualification.
  • TunaOS #3885: current-main integration and removal of unavailable redundant liberation-fonts, retaining Noto/DejaVu, strict GDM-only checks and the power-provider requirement. Observed native beta1 RPM includes the obsolete greeter and lacks tuna-wallpaper-svg. The native TuneD 2.28 RPM additionally requires (rpm-ostree if bootc) with no enabled rpm-ostree provider. Dropping the provider gate or inventing a Provides would hide the actual failure; neither was done. Targeted beta2 publication does not rebuild TuneD.

Both branches retain their manual workflow-path merge fence and promotion hold. Neither was merged or bypassed. Noble's missing target is tracked in packages #834. The 2026-10-31 primary-archive audit covers all Noble pockets and compiler names, not merely the exact rustc-1.92 name: published rustc sources and published Cargo sources still top out at native 1.91.1; rustc-web and cargo-web are not Noble escape hatches. The selected locked gtk4 0.11.5 metadata independently requires Rust at least 1.92. This is a confirmed lower bound, not an assertion about the whole lockfile's exact maximum MSRV. That audit is superseded: rustc-1.97 reached Noble on 2026-11-06, and gurnard is required for v1 (owner decisions above). A native Noble factory target, package build/install and immutable image runtime receipts are still required. No compiler from another target, third-party binary injection or target substitution is permitted.

Current required-cell accounting​

The exact-subject ledger, as of 2026-11-01 01:09 UTC, separates 0 qualified required cells, 1 failed required cell, 9 required cells with incomplete runtime measurement, 3 blocked required cells, and 1 owner-deferred/unqualified cell. All fourteen identities remain present; the ledger holds every untruncated image digest and the corresponding raw manifest path. Source-declared unsupported is not counted as qualified or substituted for an unwaived prerequisite.

  • Corrected Yellowfin 64e7a29…: 12 pass/1 fail; actual Screenshot and ScreenCast plus lock withdrawal in 0.19 seconds pass. The remaining gate failure is LockedHint=no. Reviewed portal/lock frames are additive evidence, not an all-gates pass.
  • Frozen unlocked sequence: retains all eight attempts, including Albacore completed before its deferral. It correctly stopped on unproven Grouper cleanup. Recovery proves later exact Grouper VM/VMI/PVC deletion and empty owned-prefix inventory.
  • Batched diagnostic sequence: Marlin 5ccc023… and Flounder Sid 81570c… installed and painted stock GDM, but full console stderr proves bare-name instance resolution is ambiguous. Flounder Trixie installed its disk/VM, then info timed out before readiness. These are not new product portal failures or completed login journeys.
  • The next runtime prerequisite is receipt-bound canonical instance routing, including actual backend/context/namespace/name and expected VM UID, plus independently owned Corral #674 canonical wait support. No bare-name fallback or alias/configuration change is accepted. A reviewed/merged exact-source signed normal-main CI tool may produce explicitly diagnostic receipts; final qualification requires the normal signed Corral release. No unreviewed PR tool or special trust waiver is used.
  • Hummingbird's provider closure and publication, Guppy's runtime qualification and Gurnard's Noble target remain distinct blockers. Guppy's release-approval fence was settled on 2026-11-14, and the Albacore deferral was lifted that day (see the owner decisions above).

Published immutable subjects​

The complete registry receipt records source revision, exact resolver command/output, image creation time and labels. An image's org.opencontainers.image.version is an image-source label, not the installed desktop package version. Native guest package queries and binary hashes belong in each runtime manifest.

All references below use ghcr.io/tuna-os/; digest prefixes are not truncated. The registry receipt predates the successful Wahoo publication, which has its own publication receipt.

VariantRepository and immutable digest
yellowfinyellowfin@sha256:355f66da7929059fc046527da6177e4958d97dab9546d62ca3d5ac53bc4daa4f
albacorealbacore@sha256:b68a195dbc6e079bbab6b330086e6420d860f5364d1f2232e5f377bc6cdf2a13
skipjackskipjack@sha256:30811ab4a8c8dbb4ee3cb1b77ef99fd312ec4835d05a7b40a3ed1aa410b7e5c3
wahoowahoo@sha256:865ec5320f25ae368316f07b4d3b49b5e0600308e5bc50b973ace90a5b53a46d
bonitobonito@sha256:e9a58bb671a71b3eaad4efe7c80de35eb62847bf721bc08ef37168d028273c65
bonito-rawhidebonito@sha256:dd2bc06ea05ea73fc06c7f208ac3ea2c7107cc4125024362a7d4a9ddfe18c2b3
sailfinsailfin@sha256:cbfba99f07a74a099de7366d5de2a2552d768a677cba9000476dacf0d7486605
groupergrouper@sha256:8cebaff1fbfe7d8f5076b0d09c7988cce804bbe193b12af733180f2eee3ef6fa
marlinmarlin@sha256:d0e65bce0ac55cb6c58bcf3739e30d6372fe4be978e6ea7c0c1310b3962bdfa5
flounderflounder@sha256:03a4224a791370669f5bf96228097397298fa5e39b566cd3c3d1c1fd11481e2e
flounder-sidflounder@sha256:81570c4609ad406ffa4ce5ac453c64f43708e74285ffc00aba7fc4006883cbc8
hummingbird, guppy, gurnard, stingrayno admitted Tuna image subject in the source inventory; no image pass inferred from native packages

Existing Marlin runtime verdict is not a v1 pass​

The retained 2026-10-31 real-GDM journey for the Marlin digest above installed tuna-desktop 1.0.0beta.2-2 and recorded 24 pass, 3 fail. Install, real-GDM login, session, refused wrong password, correct unlock and logout were exercised. Raw failures remain unchanged: absent VM backlight, notify-send missing GetServerInformation, and LockedHint=no while locked. Visual findings include the incorrect Dark Style tile, banner covering the date menu, and translucent power popover. These are published beta2 observations, not claims about later current source.

Actual Yellowfin runtime, 2026-10-31​

The complete immutable-image manifest ran 14:53:24–15:29:24 UTC through Corral 76b6dd52d35d, with 4 vCPUs, 6 GiB RAM, a 40 GiB disk, bochs-drm 1280×800 and real GDM on karnataka. The booted index matches 355f66…; the booted amd64 manifest is sha256:de704153960289d73a58e509d2f00d3713c8d5ee6f7ac23278b21e53cb85f120. The actual installed desktop is 1.0.0~beta.1-1.el10, not the beta2 release2 recipe now on main. GDM is 47.0, GNOME Shell 49.5 and Settings 47.7. No current-source qualification is inferred from a recent image date.

Raw result: 20 pass, 7 fail, 1 not exercised; all 28 noVNC canvas frames, hashes and guest checks are retained. The deleted VM/VMI/PVC and empty post-delete inventory are recorded in the manifest.

Required gateActual Yellowfin result
Install and booted-image identitypass for the exact published digest, bootc disk install
Real GDM loginpass; GDM greeter on seat0, password login through gdm-password
Tuna sessionpass; compositor and GTK shell run, no user GNOME Shell
Portalsfrontend Settings.Read, Screenshot version and ScreenCast source types respond; functional consent/capture/revocation not qualified
Lock/unlocklock gate fails: ScreenSaver active true but LockedHint=no; wrong password refusal and correct unlock pass
Visual reviewreviewed GDM, session, failed Settings-launch and lock frames; not a complete visual pass

Direct frame review: GDM, Tuna session, Settings attempt, lock curtain. The Settings attempt visibly contains repeated trailing s characters in the search field, not a Settings window. Consequently its timeout and downstream style/wallpaper checks are retained as failed journey observations, not independent proof that Settings itself cannot start. Other raw failures are Night Light false/no Color owner, notify-send missing GetServerInformation, and the lock hint. Sliders are not exercised because this VM has no relevant hardware device. An automated “settings-appearance” pass only records a key/ pointer action; the reviewed frame does not prove that UI appeared.

The subsequent Albacore attempt was interrupted by the outer 3600-second command deadline while Corral API reads were timing out. The interrupted matrix is intentionally not a completed matrix or a passing Albacore run. Host Tailscale/SSH and authenticated API /livez responded; Corral's default 5-second read timeout tripped its 15-second circuit breaker. This is access/ orchestration evidence, not a guest crash or an unsupported-variant reason. The access and cleanup receipt records successful deletion of only the interrupted Albacore VM/VMI/PVC and an empty post-delete inventory after using the documented CORRAL_KUBECTL_TIMEOUT=30s management budget. No cluster or desktop acceptance threshold was changed. Remaining journeys resumed sequentially with that transport setting; the original failure receipt is unchanged.

Final observed disposition​

The resumed discovery/interruption receipt resolved all admitted publications, including a later Skipjack digest sha256:36889e650e35893e4d3779deba2ddce7ce375f65a36271f86b77ca90d0c8e924. The resumed Albacore attempt did not produce a runtime/readiness manifest before interruption. No later variant VM was created by that sequence. The final disposition records deletion of that owned Albacore VM/PVC and an empty tuna-q-variant- inventory. Neither punjab nor the independent soak VM was touched.

Variant(s)Install / GDM / sessionPortalsLock/unlockQualification
yellowfin old 355f66… (beta1)pass in both retained runsavailability only; expanded run Screenshot Response 2 and ScreenCast Start timeout because generic Enter did not activate the visible Allow/Share controlsLockedHint=no fails; refusal/unlock passfail
yellowfin refreshed 64e7a29…trusted refresh run 41221855489 succeeded; guest reports beta2-2.el10; retained core run 10 pass/3 failearlier capture failures were locked-session denials; later untouched-image unlocked Screenshot/ScreenCast deliver real frames, with lock withdrawal observed in 0.147 secondsLockedHint=no; refusal/unlock/logout passnot qualified; full corrected core matrix receipt and visual review remain separate
marlinexisting exact beta2-2 journey passes these gatesnot qualified by that daily-driver receiptlock hint fails; refused password/unlock passfail, not a v1 pass
albacore b68a195…historical receipts retain actual beta1/GDM/session resultshistorical capture observations retained, not current admissionlock failures remain recordednot qualified; owner deferral lifted 2026-11-14 after Mesa crash #3969 closed; required, current-digest run pending
skipjack, wahoo, bonito, sailfinexpanded run: install/GDM/session passearlier receipts lack unlocked/timely-consent preconditions; not functional qualificationLockedHint=nofail on retained receipts; corrected preconditions pending
bonito-rawhidebuilder Secret label exceeded Kubernetes' 63-byte limit before the VM existedunprovenunprovenharness failure fixed by VM-name bound; rerun pending
grouperexpanded run interrupted during build; orphan builder VM/Secret/PVC deleted, inventory emptyunprovenunprovennot qualified
flounder, flounder-sidpublication resolved; not yet rununprovenunprovennot qualified
hummingbird, guppyadmission branches remain open; no admitted immutable-image runtime proofunprovenunprovenrequired; not qualified; stale package-unavailable reasons rejected
gurnardno native Tuna targetunprovenunprovenrequired (2026-11-14); not qualified; waits on a noble factory target
stingrayno Tuna target or flavorunprovenunprovenrequired (2026-11-14); not qualified

Owner scope override: Tuna on Albacore may be skipped temporarily while its Mesa/GDM crash is fixed elsewhere. This is an explicit deferred/unqualified row, not a pass or accepted fourteen-variant release. The current sequential run had already finished its Albacore attempt before that instruction arrived; its receipt stays unchanged, and no further Albacore attempt or publication is made. No other failed variant or gate is omitted by this decision. The machine-readable owner deferral keeps qualified:false distinct from required qualified, required failed and required unmeasured cells. The owner lifted this deferral on 2026-11-14, after tunaOS #3969 closed as completed on 2026-11-01; Albacore is required again and its retained receipts stay unchanged.

The expanded-run receipts are not retroactively converted into failures of the published desktop. The visible GNOME dialogs prove the harness pressed Enter while Allow/Share remained visible. They do prove portal qualification remained incomplete. The consent path now uses the existing AT-SPI helper with Corral's external capture; it does not spoof app IDs or replace backends.

The refreshed Yellowfin run at 19:19–19:50 UTC retained a black Screenshot prompt frame and a ScreenCast prompt showing the lock password screen. It did not assert ScreenSaver.GetActive=false at request/consent time. Its Screenshot Response 0 with empty URI and ScreenCast Response 2 must not be attributed to broken capture. On the same untouched 64e7a29… image, the retained InstalledSessionGaps proof records actual unlocked GetActive=false before/after timely Allow/Share: a real 1280×800 Screenshot file and a decoded ScreenCast frame from node 58. Explicit Session.Close removed that node in 0.013 seconds; that is not lock-triggered withdrawal. The deliberate locked negatives reproduce the earlier responses, with no new Screenshot image and no ScreenCast stream. Raw failed receipts remain unchanged. Its subsequent corrected-harness subset also proves actual Super+L withdrawal in 0.147 seconds, observer armed before the key under the unchanged two-second bound. Its separate lock check fails only LockedHint=no; unlock passes. This subset does not retroactively complete the full corrected image matrix.

The subsequent interrupted sequence completed Albacore and Skipjack before its coordinator stopped during Wahoo's builder phase. Albacore actually captured both portal images. Its lock observer expired with runner did not revoke the active grant: remote evidence collection used the client's unchanged 30-second .active→.revoke handshake before the host armed the lock. That receipt does not establish a product stream-withdrawal failure. Skipjack's capture attempts lacked the unlocked preconditions and remain unqualified. Recovery cleanup records deletion of only the interrupted Wahoo builder VM/Secret/disk and empty owned-prefix inventory. Coordinator loss is not an unsupported-variant reason.

Exact admission/publication gate snapshot:

  • Guppy PR #3883 head b1150929473ead1316427b8eae48f76832cf65d1. Earlier run 41215956243 selected no Guppy job because the migrated selector omitted it; that run is not Guppy evidence. In run 41219478905, Guppy job 124590901178 succeeded, including image build and PR boot verify; boot artifact 13054522656 records digest sha256:49b93f9e32579f03cb1f61978b080277ad8f2dd79d1dc463b59b0d96d8a1f662. That PR image is not a published immutable subject.
  • Exact Guppy publication blocker: the maintainer's tunaOS #3202 comment says: “Merge stays manual, because the change touches workflow paths and has the sentinel flag. The issue body's release gates still apply: no publication before the tuna-desktop v1 epic (tuna-desktop#577) approves it.” No candidate or signed registry image was published. Without that approval, Guppy GDM/session/ portal/lock qualification against a published digest is blocked. The owner approved Guppy publication under #577 on 2026-11-14 (decision); #3883 merges once its runtime run qualifies.
  • Hummingbird PR #3885 head 0de95641bfdd350e1f148b7f0b1eb600acf58913. Exact-head run 41213419115 failed the native TuneD transaction before desktop/boot gates.
  • Native Hummingbird publisher run 41213268502 built and signed beta2 in staging, then failed the pre-sync installability gate on a new libcanberra blocker. The served index remains beta1. Maintainer-owned packages #959 adds native rpm-ostree and libcanberra roots. Its first native build was cancelled by a later external commit, so no native closure result is claimed.

Focused regressions pass: 11 matrix tests, 17 unlocked-consent/ordering tests, 8 native coordinator tests, 11 scoped identity/cleanup tests and 8 rename-scanner tests (including actual CLI historical-vs-living inputs). They are not runtime qualification; the new canonical tool consumer remains blocked on the qualified Corral release.

Reusable actual-image path​

python3 scripts/tuna-variant-qualification --out NEW_DIRECTORY snapshots current primary configuration and resolves every admitted Tuna publication. --run additionally installs and exercises each reachable image sequentially through tuna-corral-journey; --variant NAME limits runtime work, never hides other inventory rows. The only node accepted by this lane is karnataka. Only newly owned tuna-q-variant-* VMs/disks are created/deleted. Existing punjab and the independent soak VM are never named. The Albacore deferral is lifted, so an unfiltered --run again covers every published variant; --variant still limits runtime work to a subset.

The default runtime scope is the required real-GDM/session/portal/ lock/unlock/logout core, using the journey's explicit --qualification-only mode. --full-journey also runs its app, Settings and notification tour. Neither mode upgrades portal availability into functional portal qualification. The matrix stops rather than creating another VM if owned cleanup is unproven.

~/.cache/tuna-journey-venv/bin/python scripts/tuna-variant-qualification \
--out /tmp/tuna-variants-NEW --run --variant grouper --variant flounder

Requirements and console behavior follow the Corral journey runbook: Corral bootc plugin/KubeVirt access, gh, skopeo, Python 3.11+, Pillow and Playwright Chromium. The script uses native pacman/RPM/dpkg package queries and the Gentoo VDB (/var/db/pkg, with each package's source repository), retains booted digest/package/binary provenance, drives the VM's own noVNC canvas, retains frame hashes, and deletes its VM/disks. Failure manifests also attempt a Corral framebuffer capture before deletion. No candidate compositor, shell or session payload is copied into the published-image test.

The portal availability step checks real frontend Settings, Screenshot and ScreenCast interface replies in the logged-in published session. Those probes alone remain availability-only. The functional Screenshot step reuses scripts/lib/tuna-portal-screenshot-client.py: a persistent native Gio requester that rejects a screenshot returned before a picker decision, awaits the actual Response, and validates real varied 1280×800 image pixels. The host retains that PNG and verifies its guest/host hashes. No app ID is spoofed and no backend or candidate desktop code is installed.

The ScreenCast step reuses tuna-portal-capture-client.py and tuna-pipewire-dump.py. It creates/selects/starts the real frontend session, opens its actual PipeWire FD, consumes a frame with native GStreamer, checks caller isolation, and holds its grant until the actual Super+L lock step. The unchanged two-second withdrawal observer is armed before the lock key, so remote transport latency cannot create a false pass. Actual response, requester logs, varied frame, hashes and the withdrawal snapshot are retained. Deferred frame geometry, pixel-variation and hash validation must all pass before the matrix calls ScreenCast functional; an active grant alone is not image proof. Requester exit/log diagnostics are retained even when an after-response unlocked precondition fails. Missing native Python/GObject/GStreamer/pw-dump tools remain explicit failures/prerequisites, not reasons to call the image unsupported or to install a replacement backend.

At a pending decision, the existing AT-SPI helper activates the actual visible GNOME Allow/Share control; Corral retains the console frame externally instead of using X11-only scrot. A grant test requires an unlocked session at request, consent and response; locked-session denial is a separate negative result, never a capture-product failure. Console frames still require visual review. The coordinator batches launch/readiness, then actual consent/unlocked/active/ arm bookkeeping beside the unchanged native requester; serialized SSH file reads must not consume the native 60-second consent or 30-second active-grant budget. ScreenCast may activate only Share, never a stale Screenshot Allow dialog. Its real noVNC lock key follows the arm response immediately. The existing two-second withdrawal observer remains unchanged. Full bootc provision stdout/stderr and console-start stderr are retained so orchestration failures are diagnosable.

New matrix runs freeze the host journey, native coordinator, shared identity helper, consent helper and native requester/parser sources into the output's harness/ directory before starting any VM, recording their hashes. Subsequent worktree edits cannot mutate those long-running tests. Only test requester code is placed in the fixture user's runtime directory; the installed desktop/provider payload stays untouched. The dated docs/variants/20xx-* receipt trees retain source snapshots, native package payload paths and their hashes byte-for-byte, as the existing dated performance evidence does. The rename tool preserves these historical bytes; living variant documentation and current implementation code remain scanned.

Earlier receipts identify their actual runner hash and do not gain new proof retroactively: Yellowfin's retained run predates the functional requesters. Full visual consent decisions remain required, never inferred from a successful frame. Lock-triggered withdrawal must be observed, not substituted with explicit Session.Close. Unreviewed frames, interface availability or incomplete portal flows do not become a v1 pass. Registry errors remain errors, not waivers.