feat(wallpaper): OCS browser for importing wallpaper packs

[perlowja]

Depends on #25

Stacked on the wallpaper source-selector branch in #25.

Summary

Updated 2026-09-12 after maintainer review: this branch no longer uses libadwaita at all. Every row/group/dialog is a libsingularity widget, Adw.init() and the Adw.StyleManager colour-scheme forcing are gone from src/core/main.vala, and dependency('libadwaita-1') is gone from meson.build. grep -rn 'Adw' src tests is empty and ldd singularity-desktop | grep -i adwaita returns nothing.

Earlier update 2026-09-07 after live-testing and direct maintainer feedback (use the real *Row widgets with the Group/Page convention; the browser should be an extension of the control panel, not a separate window). The browser was substantially redesigned from the original window-based version:

• WallpaperOcsBrowser is now a SettingsPage (not a Gtk.Window) pushed via the same navigation stack every other control-panel page uses (view.navigate_to(...)), with standard back navigation -- not a standalone modal.
• Category and Tag filters are SelectionRow dropdowns (previously chip pills and an expander-based control that had a real click bug).
• Import/Pin actions are a visible button below each card's thumbnail, not overlaid on the image.
• The wallpaper attribution overlay (title/author shown on the live desktop background, bottom corner) is wired up end-to-end for both Bing and real-OCS-imported sources, reading from the existing per-image .json sidecar the ncz-wallpaper-ocs helper already writes.
• Removed a duplicate provider entry (opendesktop.org and pling.com are the same content network under two names).
• WallpaperOcs (core): a pure JSON adapter over the already-shipped ncz-wallpaper-ocs CLI -- invoked via argv + async GLib.Subprocess, never a shell command string. Parses providers/index/browse/import output; validates every JSON node's type before calling a typed getter, and fails inline (never fabricates a license/author/preview) on malformed or unsupported output.
• WallpaperOcsImports (core): tracks import state via per-image sidecar files so re-browsing the same pack doesn't re-trigger a duplicate import.

The sensors-popover work that was briefly on this branch has been reverted out of it and split into its own PR, so this branch is wallpaper-only.

Validation evidence

Built via the real production path (build-singularity.sh, no SINGULARITY_SOURCE_DIR override, fresh clone from this branch) on an aarch64 build host, deployed and live-tested on real Sky1/labwc hardware -- not inferred from a clean compile:

• Live UI walkthrough confirmed: embedded settings page with back navigation, Category/Tag dropdowns actually filter results, Import/Pin buttons visible below thumbnails, provider list shows gnome-look/kde-look/pling/Bing (no opendesktop duplicate).
• A real Pling item (Mochis, OCS ID 2188591) was imported as the active desktop user; sidecar correctly recorded title/artist; after restarting singularity-desktop, the browser rediscovered the import and showed "Added" (persistence across restarts confirmed, not just in-session state).
• tests/wallpaper_ocs_test.vala and tests/wallpaper_sidecar_test.vala cover the JSON/sidecar contracts (providers/index/browse/import response shapes, missing/malformed fields, OCS vs Bing sidecar schemas, missing-sidecar-clears-attribution).
• After the libadwaita removal, re-verified on aarch64: meson setup configures with no libadwaita in core_deps, meson compile builds all 363 targets, meson test is 14/14 OK.

Not yet verified with a real pointer (synthetic input events were rejected as outside-clicks by the layer-shell surface): the visual attribution-overlay rendering itself (applying a Bing/OCS image and confirming the on-screen overlay text) needs a manual click-through to fully close out -- the underlying data path (sidecar -> GSettings keys -> overlay properties) is verified, just not the final visual render via a real click.

Compatibility

No daemon/backend changes, no new settings/state files beyond the two new background-attribution-title/background-attribution-author GSettings keys (schema in the singularity-desktop superproject) -- an import is only "Added" once it's backed by a readable registered .collection and image payload; the existing wallpaper stays active until the user explicitly picks the imported one.

Rollback

Revert this branch's commits.

---

Assisted-by: Claude Code:claude-opus-5
AI scope: assisted with the OCS/Bing JSON adapter and sidecar parsing and their unit tests, the browser page layout and filter wiring, and the libadwaita removal; the design decisions, the live hardware testing and the review responses are mine.

AI assistance: disclosed

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, you can upgrade your account or add credits to your account and enable them for code reviews in your settings.

[perlowja]

Update (2026-09-09): this branch has moved substantially since the description above. Summary of what's new:

• Crash fix: the attribution write introduced here (background-attribution-title/-author) was hitting an undeclared GSettings key and aborting the whole desktop process with SIGABRT. Root-caused via a real coredump on hardware and fixed in the schema -- see fix(schema): add missing background-attribution GSettings keys singularity-desktop#251, which this PR depends on.
• Provider-boundary refactor: extracted a WallpaperProvider interface + WallpaperProviderRegistry so OCS/Bing (and Openverse/Unsplash, present but currently unregistered) each implement one shared contract instead of the browser calling out to each provider's specifics directly. ACTIVE_PROVIDER_IDS = { "ocs", "bing" } for now -- re-enabling another provider is a one-line change to that list, not a re-port. This was driven by a real defect-class concern: a new provider (or a new field in an existing provider's response) should not be able to reach GSettings/UI code unsanitized again.
• Delete capability: pack-level and individual-image-level delete, with a real protection rule (origin set + both the content directory and registry file under the user's home) that keeps shipped/first-party wallpaper packs non-deletable without needing per-provider special-casing.

All three are real, hardware-tested, with meson test coverage (wallpaper-collections, wallpaper-ocs, settings-safety) and a real production build (ninja -C build singularity-desktop) on top of the existing state described above.

[perlowja]

One more note on scope/stability, to be explicit: Bing and OCS are the stable, production-ready providers in this PR -- both are hardware-tested, registered by default (ACTIVE_PROVIDER_IDS), and what we'd consider ready for review/merge consideration.

Openverse and Unsplash are real, working code behind the same interface, but still in progress -- not registered by default, not part of what we're asking to be reviewed here. They're structured so enabling either later is a one-line change, but please don't block this PR on their state; happy to bring them forward as a separate follow-up once they're fully settled.

[perlowja]

Update 2026-09-11 — 6 new commits, real operator-reported bugs found and fixed on live O6N hardware

All found via live use after the redesign above landed, each root-caused with hard evidence (not inferred from a clean build) before fixing:

• 9b5d4e7 / b3a3c51 — provider display name renamed OCS → OCS Network (the aggregate spans gnome-look/kde-look/opendesktop/pling, "OCS" alone was ambiguous with any one of those sites), and the collection-label join rephrased as attribution (%s — by %s) instead of a bare em-dash, so a raw OCS personid handle (e.g. dilaniweerasingha2000, which is the only submitter identifier the OCS content API provides — there's no separate display-name field) doesn't read as garbled text jammed against the photo title.
• 790ecde — every OCS import was accumulating into one growing "OCS Network" pack correctly (via a matching fix on the ncz-wallpaper-ocs CLI helper, downstream repo), but the WallpaperOcsBrowserPage — cached and reused by SettingsView across every visit, never reconstructed — only ever called imports.discover() once, at construction. Deleting the pack and re-importing the same images left the in-memory "already imported" tracker stale, so the re-imported cards rendered greyed out as if still present. Fixed by re-running discover() and forcing a re-browse every time the page is shown again (this.map signal), not just on first construction.
• 7d15a0c — deleting a single image inside a pack silently did nothing (the only way to remove it was deleting the whole pack). add_wallpaper_card() resolved the delete target via find_collection(rotation_state.get_selected_collection("ncz")) — always the active rotation source, not necessarily the collection that actually owns that card's uri (the grid legitimately mixes in "recent" images from other collections). A mismatched collection made delete_image()'s contains_uri() check fail, throwing IOError.PERMISSION_DENIED — swallowed by a bare warning(), zero user-visible feedback. Fixed with a real find_owning_collection(uri) lookup.
• 9bbbbfc — a real defensive improvement: WallpaperManager.reload() now debounces (200ms) on rapid background-picture-uri changes, so clicking through several thumbnails quickly coalesces into one decode+GPU-texture-upload instead of racing several. Kept even though it turned out not to be the root cause of the crash below — it's still correct, cheap insurance against overlapping GPU work on hardware with known driver fragility.
• 468c556 — the conclusive fix for a real, reliably-reproducible crash: clicking "Select Picture..." (the Current Wallpaper preview's file-picker button) crashed the whole shell (which then auto-restarted and correctly showed the picked image, via the session wrapper's own crash-restart loop — explaining why it looked self-healing). Root-caused with WAYLAND_DEBUG=1 on live hardware (NCZ-OS, labwc compositor): the last protocol request before death is zxdg_exporter_v2.export_toplevel (xdg-foreign-v2), triggered by passing a parent window to Gtk.FileDialog.open() so GTK can hand the portal a transient-parent handle. labwc advertises zxdg_exporter_v2 but disconnects the client instead of replying with zxdg_exported_v2.handle — confirmed via a live gdb backtrace showing the resulting exit() call originates entirely inside libgtk-4.so.1/libglib-2.0.so.0, no application code anywhere in the crashing frames. Fixed by passing null instead of a parent window — a real, supported GtkFileDialog usage pattern, at the minor cost of losing window-stacking/transiency on compositors where xdg-foreign works correctly. (Will file the underlying protocol bug against labwc separately.)

All 6 commits build clean (684/684 targets), were rebuilt via a real container build each time, ABI-verified (ldd), and deployed+re-tested live on O6N (Sky1/labwc) after every fix — including the crash reproduction and its resolution, confirmed live by the operator.

[mirkobrombin]

Please replace every direct libadwaita widget with libsingularity components and split the unrelated sensors changes into a separate PR.

[perlowja]

Addressed both points:

libadwaita widgets replaced. Every direct Adw.* widget this branch's OCS/Bing browser code introduced is now a libsingularity component, matching how the rest of this codebase already does it (desktop_page.vala's grid_group/SwitchRow, calendar_view.vala's PreferencesGroup, ush_portal.vala/system_view.vala's ConfirmDialog):

• wallpaper_ocs_browser.vala: Adw.PreferencesGroup -> PreferencesGroup, Adw.ActionRow -> ActionRow, Adw.PreferencesRow -> PreferencesRow, Adw.EntryRow -> EntryRow, Adw.ComboRow -> SelectionRow for the provider/category/tag pickers.
• provider_credential_group.vala: now extends PreferencesGroup instead of Adw.PreferencesGroup; uses EntryRow/PasswordRow instead of Adw.EntryRow/Adw.PasswordEntryRow.
• desktop_page.vala: the Bing Markets picker moves from Adw.ComboRow to SelectionRow; the Bing markets multi-select dialog and both wallpaper delete-confirmation dialogs move from Adw.AlertDialog/Adw.MessageDialog to Singularity.Widgets.ConfirmDialog.

No behavioural change intended -- verified with a real meson setup/meson compile/meson test on a clean checkout (14/14 tests pass).

Sensors changes split out. The sensors-popover work (native-widget conversion + responsiveness) that had ended up bundled into this branch touched only src/components/panel/panel.vala and nothing else, so it's now cleanly extracted into its own branch and PR: #27. panel.vala on this branch is back to main's content.

Re-requesting review.

[mirkobrombin]

Direct libadwaita initialization and dependency still remain :D Also please remove Claude attribution and change it to:

Assisted-by: <tool>:<model-version>
AI scope: what the AI generated in this commit, and the prompt used (or a short summary of it)

[perlowja]

Both addressed in c2c0e20. Thanks for catching the leftovers — you were right, the widget swap in 5684deb left the initialization and the dependency behind.

1. Direct libadwaita initialization and dependency

Grepped the whole branch diff against main rather than just the last commit's files. Three sites, all of them introduced by this PR:

• src/core/main.vala:33 — Adw.init() in SingularityApp.startup(). Removed. It only existed to back the raw Adw rows this PR used to construct, and every one of those is now a libsingularity widget.
• src/core/main.vala:1183-1184 — Adw.StyleManager.get_default().color_scheme = …FORCE_DARK/FORCE_LIGHT in update_theme_mode(). Removed. It only affected libadwaita widgets inside this process, and there are none left. The shell chrome already follows Singularity.Style.StyleManager.apply_color_scheme() on the line above it, and third-party apps already follow gtk_application_prefer_dark_theme plus the org.gnome.desktop.interface color-scheme write a few lines below, so nothing lost coverage.
• meson.build:26 — dependency('libadwaita-1', version: '>= 1.2'). Removed.

On whether that dependency was genuinely removable rather than just redundant — I checked before deleting the line, since a transitive need would have made it a no-op rather than a fix:

• core_deps was a plain dependency('gtk4') before this branch, so the libadwaita entry is entirely this PR's.
• libsingularity deliberately keeps libadwaita out of core (fix(review): keep libadwaita out of core), so PreferencesGroup/ActionRow/SelectionRow/ConfirmDialog are plain GTK4 and don't pull it back in.
• goa-backend-1.0 is not a declared dependency here (only goa-1.0), src/vapi/goa-backend-1.0.deps lists just goa-1.0 and gtk4, and add_account_page.vala already avoids Goa.BackendProvider specifically to not depend on libadwaita.

So it was a real direct dependency with no remaining consumer, not a redundant declaration of something arriving another way.

I also reworded two stale comments that still mentioned Adw.EntryRow/Adw.ComboRow by name, so grep -rn 'Adw' src tests is now completely empty.

Verified on an aarch64 build host, not inferred:

• meson setup configures with no libadwaita in core_deps
• meson compile builds all 363 targets
• meson test — 14/14 OK
• ldd build/singularity-desktop | grep -i adwaita — no output, so the binary no longer links libadwaita even transitively

2. Attribution format

Switched to your format. c2c0e20 uses:

Assisted-by: Claude Code:claude-opus-5
AI scope: located the remaining libadwaita initialization and dependency
sites across this branch's full diff, removed the three call sites and the
meson dependency line, reworded two stale comments, and ran the build and
test verification quoted above. Prompt, summarized: "finish the libadwaita
removal - find any remaining Adw.init/Adw.* usage and the lingering
libadwaita pkg-config dependency in this PR's diff, confirm the dependency
is genuinely removable rather than needed transitively through another
dependency, remove it, and verify with meson setup/compile/test on the
aarch64 build host."

I've also put a PR-level Assisted-by: / AI scope: block at the bottom of the description covering the whole PR, and I'll use this format on all singularity-* work going forward rather than the old trailer.

One thing I did not do unilaterally: seven earlier commits on this branch still carry the old Co-Authored-By: trailer. Rewriting those means a force-push of the branch, which would churn the PR and detach your review anchors, so I left it to you — say the word and I'll rewrite the trailers in one pass.

The sensors popover work is already reverted off this branch (3b98920) and split out, so this branch is wallpaper-only as requested in your first pass.

[perlowja]

Both halves of your review are done. Branch head is 61378b2.

Libadwaita removal

Three sites, all introduced by this PR, all gone:

• src/core/main.vala:33 — Adw.init() in SingularityApp.startup().
• src/core/main.vala:1183-1184 — Adw.StyleManager…color_scheme = FORCE_DARK/FORCE_LIGHT in update_theme_mode(). It only affected libadwaita widgets inside this process and there are none left; the shell chrome already follows Singularity.Style.StyleManager.apply_color_scheme() on the line above, and third-party apps already follow gtk_application_prefer_dark_theme plus the org.gnome.desktop.interface color-scheme write just below.
• meson.build:26 — dependency('libadwaita-1', version: '>= 1.2').

I checked the dependency was genuinely removable rather than needed transitively before deleting it: core_deps was a plain dependency('gtk4') before this branch, libsingularity deliberately keeps libadwaita out of core, and goa-backend-1.0 is not a declared dependency here (src/vapi/goa-backend-1.0.deps lists only goa-1.0 and gtk4, and add_account_page.vala already avoids Goa.BackendProvider for exactly that reason).

Verified on an aarch64 build host: meson setup configures with no libadwaita in core_deps, meson compile builds all 363 targets, meson test is 14/14 OK, and ldd build/singularity-desktop | grep -i adwaita returns nothing — so it isn't linked directly or transitively. Two stale comments naming Adw.EntryRow/Adw.ComboRow were reworded too, so grep -rn 'Adw' src tests is empty.

Attribution trailers in history

Rewrote the seven commits that carried the old trailer — bfa5f5c, c3f9970, 468c556, 9bbbbfc, 7d15a0c, 032bb6d, 4dc9d30 — replacing Co-Authored-By:/Claude-Session: with your format, keeping the scope to one sentence per commit — e.g. on the debounce fix:

Assisted-by: Claude Code:claude-opus-5
AI scope: generated schedule_reload() and its 200ms debounce, from a
prompt to diagnose the rapid-click Wayland disconnect from the app log
and coalesce the overlapping decode and upload reloads.

Verification after the rewrite:

• git log main..HEAD — Co-Authored-By: / Claude-Session: / noreply@anthropic.com / Generated with: 0 matches. No other bot footer anywhere.
• 54 commits before, 54 after; every one still authored Jason Perlow <jperlow@gmail.com>.
• The rewrite was message-only — git diff between the pre- and post-rewrite heads is empty, not a byte of tree content changed, so the build verification above still applies unchanged.

The PR description carries the same trailer pair, scoped to the PR as a whole. This is the format I'll use on singularity-* work from here on.

Sorry for the force-push churn on your review anchors — reaching the older messages required it. The sensors popover was already reverted off this branch in 3b98920 and split out, so this is wallpaper-only. Ready for another look whenever you have time.

[perlowja]

Pushed ba53644 onto this branch: the Bing source in the online browser now offers the same de-duplicated "combined" view the wallpaper theme picker already uses.

Bing serves the same photograph to several regional markets on a given day. This page crawls one listing per market and merges them into a single grid, so with every market enabled it was showing each photo once per market that served it — 213 cards for 41 distinct photographs on the board I tested, while the theme picker next to it showed the correct 41.

De-duplication has to happen across markets, which a per-market listing structurally can't do, so the provider asks the helper for the combined view rather than trying to reconcile the markets itself:

• The helper advertises a consolidated pseudo-market from markets, and only when the user's configured markets are the "all" sentinel. I deliberately didn't derive that on this side: desktop_page.vala's Bing Markets picker treats an absent config file as "all" while the helper treats it as en-US only, so the helper's own answer is the contract.
• When that choice is present it replaces the market list rather than joining it (WallpaperBing.combined_view). Since browse_all() crawls one listing per choice into one grid, offering both would put the de-duplicated set and every raw per-market set in the same grid and bring the duplication straight back. Picking specific markets in the Bing Markets picker is what restores per-market browsing, and is also what makes the helper stop advertising the combined view.
• BingWallpaperProvider.display_name becomes "Bing (Combined, All Markets)" when it's serving that view, matching the collection label on the theme-picker side so the two name one thing one way. The name isn't known until the helper replies, so the "Online source" row is rebuilt from the registry's live names (rebuild_provider_row) instead of being snapshotted in initialize().
• rebuild_category_row() hides the row when it holds a single choice — it filters nothing in that state, and an "Any category / Combined (All Markets)" dropdown above a grid that is already exactly that is just noise.

Tests cover both directions of the replacement rule. Built and ran the full suite on aarch64: 14/14 green, including the two new cases.

The helper half is in cix-installer (post-install/45-wallpaper-rotator.sh, list --consolidated); list <market> keeps its existing meaning there, so nothing else changes shape.

[mirkobrombin]

This branch still contains the pre-runtime-consumer wallpaper stack instead of the current #25 head, so please rebase it after #25 lands and resolve the overlapping files before re-review.

[perlowja]

Rebased and consolidated onto the current PR #25 branch tip (wallpaper-source-selector-clean at 8ddf523) because #25 has not merged yet. The overlap in meson.build and desktop_page.vala was resolved against #25's current runtime rotator/source-selector implementation (including its comment/ASCII cleanup), while retaining the refreshable source selector required by OCS imports/deletions.

The rewritten series now contains:

• c8a83f0 — consolidated OCS browser integration on the runtime source selector
• cbd0bb4 / 13f9d84 — native responsive sensors popover + restored severity colours
• 41e5481 — Bing helper version-skew documentation/warning
• caa4b7a — inline Bing region picker

Verified in a Debian forky Podman container: Ninja 365/365; Meson tests 14/14 passing.

This is necessarily an interim rebase onto #25's branch tip. I will do one final re-rebase onto upstream main after #25 actually merges.

[perlowja]

@mirkobrombin, please re-review the rebased series. GitHub returned 404 when I attempted the requested-reviewers API call from the fork-owner account, so I am explicitly tagging you here.

[perlowja]

@mirkobrombin the last review on this PR (pinned to 67eeec3) asked for exactly the rebase that already happened -- the branch was rebased onto the current PR #25 tip and the overlapping files resolved, as noted in the comment above requesting re-review of the rebased series (current head 9a628da, 16 commits ahead / 58 behind 67eeec3). I do not have permission to dismiss the review from this account (the dismissal API 404s the same way the reviewer-request API does) -- could you dismiss it or take another look when you have a moment?

[perlowja]

Brought this branch up to the current OCS quality bar with 6 commits: raised the aggregate crawl cap to the real measured feed size (3352 unique items, was silently truncating at 1500), added a disk-backed thumbnail cache with LRU eviction so revisits are instant, bounded thumbnail memory to roughly a viewport worth via scroll-driven load/evict instead of decoding the entire feed up front (was measured at 1.35 GiB resident for 3352 cards), added stale-while-revalidate so a cached crawl paints immediately while refreshing in the background, sanitized free-text OCS tags against Unicode bidi-override/control characters before they reach the tag dropdown, and added a Clear Filters action. Verified with a full ninja build (228/228, 0 errors) and meson test (11/11) in a debian:forky container.

[perlowja]

Addressed all three threads on this branch (now at c18a1f6):

• libadwaita: already fully removed as of a prior push. grep -rln 'Adw\.' <changed files> returns nothing, and no dependency('libadwaita...') line was added to meson.build. (This review comment predates that cleanup landing.)
• Attribution format: verified against the repo's current AGENTS.md — all commits on this branch already use Assisted-by:/AI-Scope:, no AI scope/Co-Authored-By remnants.
• Rebase after feat(wallpaper): source selector, per-source grid scoping, and rotation controls #25: rebased onto current origin/main (post feat(wallpaper): source selector, per-source grid scoping, and rotation controls #25 merge). git's rebase auto-skipped 11 commits that duplicated what feat(wallpaper): source selector, per-source grid scoping, and rotation controls #25 (source selector/rotation) and fix(panel): convert sensors popover to native Singularity widgets, make it responsive #27 (sensors popover, already merged separately) had already landed — confirming the sensors-panel split you asked for happened via fix(panel): convert sensors popover to native Singularity widgets, make it responsive #27, not left bundled here. The one real conflict was in desktop_page.vala's grid-setup code: kept feat(wallpaper): source selector, per-source grid scoping, and rotation controls #25's already-merged refresh_wallpaper_sources()/dynamic-source-container mechanism (it's the one actually wired to pack-delete refresh elsewhere in the file) rather than feat(wallpaper): source selector, per-source grid scoping, and rotation controls #25's older superseded inline source_row, and added this branch's "Browse Online Wallpapers" entry point on top of it.

Compile verification: static checks above are solid; a full build hit an environment gap (libsingularity not available as an installed dependency or fetchable subproject on my available build hosts) unrelated to this diff — noting honestly rather than claiming something I didn't verify.

Assisted-by: Claude Code:claude-sonnet-5

[perlowja]

Pushed 6c24ca4 (was c18a1f6): the OCS wallpaper browser simplification requested by the operator -- category is now the only filter (tag filtering and the "Clear Filters" action are removed), and browsing is lazy per-category instead of the previous eager crawl that pulled every usable category into one merged aggregate before the user picked anything (measured against pling: 45 categories, 3485 results, 3300+ items in a single load). Selecting a provider now only loads its category list; picking a category loads (and, on a repeat visit or Refresh, re-validates) just that one category, reusing the existing cache-then-revalidate-if-stale flow -- now keyed by provider+category instead of provider alone. Verified via the wallpaper-ocs-test target: 34/34 pass, unchanged from before this change.

On the three open review threads, all pinned to commits well behind the current head -- I want to lay out precisely why each is resolved rather than just asserting it, since I don't have dismiss rights on this repo (confirmed: push: false) and this needs a maintainer action either way:

1. "replace every direct libadwaita widget... split the sensors changes" (bfa5f5c) -- resolved. The sensors-popover work was split out and merged separately (#27, now in main's history). No Adw.* widget remains in this branch's diff.
2. "Direct libadwaita initialization and dependency still remain... remove Claude attribution" (5684deb) -- resolved. grep -rn 'Adw' src tests on the current head matches only the literal string "Adwaita" as a GTK theme/cursor-scheme NAME (a real, selectable system theme a user can pick, e.g. desktop_page.vala's theme list) -- never a widget or namespace reference. meson.build has no libadwaita dependency. Commit messages use Assisted-by:/AI scope:, no Co-Authored-By.
3. "still contains the pre-runtime-consumer wallpaper stack... rebase after feat(wallpaper): source selector, per-source grid scoping, and rotation controls #25 lands" (67eeec3) -- resolved. #25 merged into main on 2026-09-14 (c10ad57); this branch is rebased on top of that merge commit (confirmed via git merge-base --is-ancestor).

Given I can't dismiss these myself, @mirkobrombin -- could you take a look and either dismiss what's stale or flag anything I've misjudged? Happy to address anything still live.

One more thing, flagged honestly rather than worked around: singularity-desktop (the actual app binary) currently fails to build on this branch, but for a cause entirely unrelated to wallpapers -- performance_page.vala/fan_controller.vala reference FanControlChannel and SensorMonitor.gpu_sampling, neither of which exist in the current libsingularity subproject. I confirmed this is pre-existing and not something this branch introduced: the same sensors.gpu_sampling = false; line is present verbatim on a fresh, independent clone of upstream main (the fan-monitoring feature, 91b5f55, merged same day as #25). I did not attempt to fix it -- it's a different subsystem than what I was asked to change, and I don't have enough context on the intended libsingularity/fan-control contract to risk a blind fix. Wanted to surface it rather than let a green wallpaper-only verification imply the whole app builds, which it currently does not.

---

Assisted-by: Claude Code:claude-opus-5
AI scope: wrote this status/verification comment summarizing the pushed commit, the stale-review analysis, and the unrelated pre-existing build break found while verifying.

[perlowja]

Amended to 6b64018 (was 6c24ca4) -- the AI-disclosure check correctly caught a real formatting bug in my own commit trailer (AI scope: with a space instead of the required AI-Scope: with a hyphen; verified the fix against the checkers exact regex before re-pushing). Check now passes. Everything else in the comment above still applies to this SHA.

Assisted-by: Claude Code:claude-opus-5
AI-Scope: wrote this follow-up noting the amended commit and the now-passing check.