style: rustfmt the abandoned-devnode sweep

fix(host,audio): the registry stamp route reached for the Render hive even for capture endpoints
write_stamps falls back to a raw-registry write when the property store denies it. That fallback built its path from MMDEV_RENDER_PATH unconditionally, so stamping the minted microphone's CAPTURE endpoint reached for ...\MMDevices\Audio\Render\{capture-guid}\Properties - a key that cannot exist. RegOpenKeyExW then failed, write_stamps returned the error, and stamp_identity degraded to 'keeps the driver's default name'. Invisible to the pad program, whose endpoints are render-only, and invisible on any box where the property store route succeeds (both field logs show registry=[] on every stamp line, so neither reporter ever took this path). It only bites where the property store is denied - exactly the boxes the ACL repair exists for. The hive now follows the direction the endpoint id encodes, with render as the default for anything unrecognised. Unit-tested.
2026-08-23 09:49:49 +02:00 · 2026-08-23 09:40:01 +02:00 · 2026-08-23 09:08:57 +02:00 · 2026-08-23 09:03:52 +02:00 · 2026-08-22 23:38:23 +00:00 · 2026-08-23 01:02:27 +02:00
15 changed files with 1125 additions and 65 deletions
@@ -50,7 +50,10 @@ on:
      - 'crates/pf-vaadec/**'
      - 'packaging/flatpak/**'
      - 'Cargo.lock'
+      # Both halves of this job's correctness, not of the bundle's content: a change to either
+      # can only be proven by a real run, and there is no other trigger that would give it one.
      - '.gitea/workflows/flatpak.yml'
+      - 'scripts/ci/flatpak-deps-present.sh'
    tags: ['v*']
  workflow_dispatch:

@@ -270,12 +273,13 @@ jobs:

      - name: Prefetch deps + sources (retried — the network phase, split off the build)
        run: |
+          set -euo pipefail
          # All of the job's heavy network I/O happens HERE, retried, so a dropped DNS lookup
          # or TCP dial costs a backoff-retry instead of the whole (long) compile:
-          #   1) --install-deps-only pulls everything the manifest declares from Flathub: the
-          #      GNOME 50 runtime/SDK + the rust-stable (//25.08, rustc 1.96) and llvm20 SDK
-          #      extensions. (No codec extension: the client links no FFmpeg — see the
-          #      manifest header.)
+          #   1) the Flathub deps the manifest declares — the GNOME 50 runtime/SDK + the
+          #      rust-stable (//25.08, rustc 1.96) and llvm20 SDK extensions — but ONLY the ones
+          #      genuinely MISSING; see the block below. (No codec extension: the client links no
+          #      FFmpeg — see the manifest header.)
          #   2) --download-only fetches every source (all crates in cargo-sources.json) into
          #      the .flatpak-builder state dir. Both are resumable/idempotent, so re-running
          #      after a partial failure is safe and cheap.
@@ -288,9 +292,40 @@ jobs:
          # for the mechanism.
          # 10 attempts (~9min budget), matching the remote-add bootstrap above — same shared,
          # load-sensitive runner, same flathub.org resolution path.
-          bash scripts/ci/retry.sh 10 flatpak-builder --user --force-clean --disable-rofiles-fuse \
-            --install-deps-from=flathub --install-deps-only \
-            "$PWD/build-dir" "$MANIFEST"
+          #
+          # WHY THIS IS NOT AN UNCONDITIONAL `--install-deps-only` ANY MORE (2026-08-22):
+          # that flag does not install what is missing, it UPDATES what is present.
+          # builder_manifest_install_dep() branches on `flatpak info --show-commit <ref>` succeeding
+          # and runs `flatpak update` for every dep already installed — with no fallback to a
+          # plain install when that update fails — and ci/flatpak-ci.Dockerfile bakes
+          # the entire runtime set, so on a healthy run it was a pure no-op that nonetheless made
+          # every build depend on Flathub being healthy at that minute. It bit on 2026-08-22:
+          #     Updating runtime/org.freedesktop.Sdk.Extension.rust-stable/x86_64/25.08
+          #     Error: Failed to update org.freedesktop.Sdk.Extension.rust-stable: While pulling …
+          #       .filez: Server returned HTTP 404
+          # dl.flathub.org served a 404 for one object of the then-current rust-stable//25.08
+          # commit, deterministically — all 10 retry.sh attempts died on the SAME object over
+          # ~9 min — and flatpak-builder SEGFAULTED on its own error path (rc=139), so retry.sh
+          # saw a crash rather than a clean "this will never work" either. The build never wanted
+          # that newer commit: the manifest pins a runtime VERSION, not a commit, and the baked
+          # one satisfies it. Updating bought nothing and imported an upstream outage.
+          #
+          # So: assert what the image already has, and reach for Flathub only on a real miss —
+          # the same "guard, don't install on top of a stale image" doctrine as the Tooling step.
+          # The check lives in scripts/ci/flatpak-deps-present.sh (run its --self-test after
+          # touching it): a bug in it that reports "satisfied" when it is not would build against
+          # whatever runtime happened to be lying around, which is worth more than an inline
+          # if-statement. It deliberately fails OPEN — anything it cannot parse takes the slow
+          # install path below.
+          if bash scripts/ci/flatpak-deps-present.sh "$MANIFEST"; then
+            echo "deps satisfied by the baked image — not touching Flathub"
+            flatpak list --user --columns=ref
+          else
+            echo "::warning::$MANIFEST declares deps punktfunk-flatpak-ci does not have — pulling from Flathub (~1.5 GB). Bump GNOME_VERSION/FREEDESKTOP_VERSION in ci/flatpak-ci.Dockerfile so this stays off the hot path."
+            bash scripts/ci/retry.sh 10 flatpak-builder --user --force-clean --disable-rofiles-fuse \
+              --install-deps-from=flathub --install-deps-only \
+              "$PWD/build-dir" "$MANIFEST"
+          fi
          bash scripts/ci/retry.sh 10 flatpak-builder --user --force-clean --disable-rofiles-fuse \
            --download-only --disable-updates \
            "$PWD/build-dir" "$MANIFEST"
@@ -298,7 +333,17 @@ jobs:
      - name: Build the flatpak (offline — deps + sources prefetched above)
        run: |
          # Everything is already local (state dir warmed by the prefetch step), so this long
-          # step needs no network; --install-deps-from stays as a no-op safety net.
+          # step needs no network.
+          #
+          # --install-deps-from=flathub USED to sit here, commented as "a no-op safety net". It
+          # was neither. builder-main.c calls builder_manifest_install_deps() whenever that flag
+          # is set — --install-deps-only only decides whether it EXITS afterwards — so this step
+          # re-ran the same `flatpak update` of the runtimes that killed the prefetch step on
+          # 2026-08-22 (Flathub HTTP 404 on a rust-stable//25.08 object; see there). A live pull
+          # of multi-GB runtimes is a strange thing to call a safety net in the step whose whole
+          # design is to be offline, and it could only ever fire if the prefetch step above had
+          # already failed the job. Dropped: the prefetch step is the one place that talks to
+          # Flathub, and it is the one place with retries.
          #
          # --disable-updates is LOAD-BEARING, not tidiness: without it this step was never
          # actually offline. flatpak-builder runs the DOWNLOAD PHASE again as part of every
@@ -326,7 +371,6 @@ jobs:
          flatpak-builder --user --force-clean --disable-rofiles-fuse \
            --default-branch="$FLATPAK_BRANCH" \
            --disable-updates \
-            --install-deps-from=flathub \
            --repo="$PWD/repo" \
            "$PWD/build-dir" "$MANIFEST"

@@ -515,8 +515,21 @@ class MainActivity : ComponentActivity() {
        else -> KeyEvent.KEYCODE_DPAD_RIGHT
    }

-    /** Resolve the panel's highest-refresh mode (same resolution) once, for [setConsoleHighRefreshRate]. */
+    /**
+     * Resolve the panel's highest-refresh mode (same resolution) once, for [setConsoleHighRefreshRate].
+     *
+     * NEVER on a TV, which leaves the id at `0` and makes every [setConsoleHighRefreshRate] call a
+     * no-op. The pin exists for phone refresh governors that cap third-party apps at 60 Hz; a TV has
+     * no such governor, and there it does active harm. `display.mode` is what [nativeDisplayMode]
+     * reads to resolve "Native" refresh at connect, so a menu-time pin makes the session negotiate
+     * the PINNED rate rather than the TV's real HDMI output — and [StreamScreen] then releases the
+     * pin on TV (the decoder's own mode switch governs there), dropping the panel back to 60 while
+     * the host is already serving 120. Every frame then waits out that mismatch, which is the
+     * "latency explodes unless I set the refresh by hand" field report: picking a refresh explicitly
+     * is precisely what bypasses the corrupted `nativeDisplayMode` answer.
+     */
    private fun resolveHighRefreshMode() {
+        if (isTvDevice(this)) return
        @Suppress("DEPRECATION")
        val disp = if (Build.VERSION.SDK_INT >= Build.VERSION_CODES.R) display else windowManager.defaultDisplay
        highRefreshModeId = disp?.supportedModes?.maxWithOrNull(
@@ -478,7 +478,12 @@ fun nativeDisplayMode(context: Context): Triple<Int, Int, Int> {
    val mode = display.mode
    val w = mode.physicalWidth
    val h = mode.physicalHeight
-    val hz = mode.refreshRate.toInt().coerceAtLeast(1)
+    // ROUNDED, not truncated: TVs report the fractional NTSC rates over HDMI (59.94, 29.97,
+    // 23.976), and `toInt()` turns 59.94 into 59 — a rate no display mode anywhere has, which the
+    // host then serves by clamping DOWN to the highest mode it advertises at or below it. Rounding
+    // also keeps this agreeing with `MainActivity.streamPanelFps`, which already rounds; the two
+    // describe the same panel and must not disagree.
+    val hz = kotlin.math.round(mode.refreshRate).toInt().coerceAtLeast(1)
    return Triple(maxOf(w, h), minOf(w, h), hz)
 }

@@ -367,6 +367,38 @@ fn takeover_state_is_live(state: &TakeoverState) -> bool {
        || state.forced_screen_env
 }

+/// Restart the box's own autologin gaming session(s) after a leftover idle drop-in was swept off
+/// a host that died holding one ([`restore_takeover_on_startup`]).
+///
+/// Gated on the box actually being dark ([`box_session_live`]): if the user is already in game mode
+/// or on a desktop, the drop-in we removed was inert and bouncing their session would be the bug.
+/// Only an ACTIVE instance is restarted — under a just-removed idle drop-in, active means "running
+/// the sleep"; an inactive one is a leftover the display manager will handle on its own.
+fn hand_back_idled_units_after_crash() {
+    if box_session_live() {
+        return; // something is already drawing — the drop-in was inert
+    }
+    let units: Vec<String> = listed_autologin_units()
+        .into_iter()
+        .filter(|(_, active)| active == "active")
+        .map(|(unit, _)| unit)
+        .collect();
+    if units.is_empty() {
+        return;
+    }
+    tracing::warn!(
+        ?units,
+        "gamescope: the box's Game Mode is running the dead host's idle placeholder and its panel \
+         is dark — restarting it"
+    );
+    for unit in &units {
+        if let RestoreVerb::Failed(why) = issue_restore_verb(&["restart", unit]) {
+            tracing::error!(unit, status = %why, "gamescope: could not restart it");
+        }
+    }
+    ensure_box_session_or_escalate(&units);
+}
+
 /// On host startup, restore the TV's gaming session if a previous host instance took it over and
 /// crashed before restoring (`design/gamemode-and-dedicated-sessions.md` A3). Loads the persisted
 /// [`TakeoverState`] into the statics and schedules a restore after a short reconnect grace (so a
@@ -399,6 +431,13 @@ pub fn restore_takeover_on_startup() {
            "gamescope: removed a leftover idle drop-in from a previous host instance — the box's \
             own Game Mode session would have started and then done nothing"
        );
+        // Removing the FILE does not touch the unit RUNNING under it. That unit's `ExecStart` was
+        // replaced with a sleep, so it is `active` and drawing nothing, and nothing below will
+        // restart it: the takeover file may be absent, unparseable, or not `takeover_state_is_live`
+        // — and all three of those exits used to leave the box sitting on a dark panel with its
+        // Game Mode "running". A host killed mid-stream (SIGKILL, OOM, a yanked update) lands
+        // exactly there, and on glass it is indistinguishable from broken hardware. Hand it back.
+        hand_back_idled_units_after_crash();
    }
    let Ok(bytes) = std::fs::read(takeover_state_path()) else {
        return; // no takeover file — clean start
@@ -2984,6 +3023,48 @@ fn replay_switch_under_restored_dm(dm: &str) {
    }
 }

+/// The box's autologin gaming instances and their ACTIVE state, as `(unit, active)` pairs — the
+/// `--plain` columns are UNIT LOAD ACTIVE SUB DESCRIPTION, so the state is the third.
+///
+/// An unanswered query reads as "none listed", which is the safe direction for both callers: the
+/// takeover then frees nothing rather than killing a session it could not see properly, and the
+/// crash hand-back restarts nothing rather than bouncing one.
+fn listed_autologin_units() -> Vec<(String, String)> {
+    let Ok(out) = crate::proc::output_within(
+        Command::new("systemctl").args([
+            "--user",
+            "list-units",
+            "--type=service",
+            "--all",
+            "--no-legend",
+            "--plain",
+            "gamescope-session-plus@*.service",
+        ]),
+        UNIT_QUERY_BUDGET,
+    ) else {
+        return Vec::new();
+    };
+    parse_listed_units(&String::from_utf8_lossy(&out.stdout))
+}
+
+/// [`listed_autologin_units`]'s parser (the unit-testable core). Which column the ACTIVE state is
+/// in decides whether the takeover can tell a live gaming session from a dead leftover, and
+/// getting that wrong is silent in both directions — a live session read as dead leaves Steam
+/// holding the instance our own launch then collides with, and a dead one read as live idles a
+/// session nobody was in.
+fn parse_listed_units(stdout: &str) -> Vec<(String, String)> {
+    stdout
+        .lines()
+        .filter_map(|l| {
+            let mut cols = l.split_whitespace();
+            let unit = cols.next()?;
+            let active = cols.nth(1).unwrap_or("");
+            (unit.starts_with("gamescope-session-plus@") && unit.ends_with(".service"))
+                .then(|| (unit.to_string(), active.to_string()))
+        })
+        .collect()
+}
+
 /// Stop every autologin gaming-mode session (`gamescope-session-plus@*.service`) so its
 /// single-instance Steam is free for our own host-managed session. Records the units so
 /// [`schedule_restore_tv_session`] can restart them on disconnect. Our own session is the transient
@@ -3011,33 +3092,9 @@ fn replay_switch_under_restored_dm(dm: &str) {
 /// The ORDER is therefore load-bearing and not a style choice: stop the DM, bail if it did not
 /// land, and only then mask. A mask laid before a stop that never arrives is the storm.
 fn stop_autologin_sessions() -> Result<()> {
-    let Ok(out) = crate::proc::output_within(
-        Command::new("systemctl").args([
-            "--user",
-            "list-units",
-            "--type=service",
-            "--all",
-            "--no-legend",
-            "--plain",
-            "gamescope-session-plus@*.service",
-        ]),
-        UNIT_QUERY_BUDGET,
-    ) else {
-        return Ok(());
-    };
-    // `(unit, ACTIVE state)` — the `--plain` columns are UNIT LOAD ACTIVE SUB DESCRIPTION.
-    let listed: Vec<(String, String)> = String::from_utf8_lossy(&out.stdout)
-        .lines()
-        .filter_map(|l| {
-            let mut cols = l.split_whitespace();
-            let unit = cols.next()?;
-            let active = cols.nth(1).unwrap_or("");
-            (unit.starts_with("gamescope-session-plus@") && unit.ends_with(".service"))
-                .then(|| (unit.to_string(), active.to_string()))
-        })
-        .collect();
+    let listed = listed_autologin_units();
    if listed.is_empty() {
-        return Ok(()); // nothing autologged in — Steam is already free
+        return Ok(()); // nothing autologged in (or the query failed) — Steam is already free
    }
    let dm = display_manager_unit();
    // Only a LIVE instance holds Steam / justifies touching the DM. A loaded-but-inactive
@@ -3439,7 +3496,13 @@ pub fn restore_takeover_now() {
    }
    *PENDING_RESTORE.lock().unwrap_or_else(|e| e.into_inner()) = None; // doing it right here
    tracing::info!("gamescope: host is shutting down — restoring the box's own session first");
-    do_restore_tv_session();
+    // `verify: false` — the escalation ladder waits up to a minute, and this runs inside
+    // `native.rs`'s 20 s `SHUTDOWN_RESTORE_GRACE`, after which `exit(0)` runs no destructors.
+    // Spending that grace watching instead of restoring would COST the hand-back, not check it.
+    // The next host start is what covers a shutdown that left the box dark
+    // ([`restore_takeover_on_startup`], which now hands the box back rather than only sweeping the
+    // drop-in off it).
+    do_restore_tv_session(false);
 }

 /// What a bounded `systemctl --user` lifecycle verb on the RESTORE path actually did. Three states,
@@ -3503,11 +3566,168 @@ fn connected_connector_under(base: &std::path::Path) -> bool {
    })
 }

+/// How long a hand-back waits for the box to show something on its own panel before it starts
+/// escalating. Generous on purpose: the unit's `ExecStart` is a whole gamescope + Steam start, and
+/// on a cold box that is not quick — while a false escalation costs the user a session bounce.
+const HANDBACK_GRACE: Duration = Duration::from_secs(25);
+
+/// How long each escalation rung gets. Shorter than [`HANDBACK_GRACE`]: by the time a rung runs,
+/// the ordinary start has already had its full grace and not delivered.
+const HANDBACK_RUNG_GRACE: Duration = Duration::from_secs(15);
+
+/// Poll slice for the two waits above.
+const HANDBACK_POLL: Duration = Duration::from_millis(500);
+
+/// Is ANYTHING driving the box's own panel right now — its game mode, or a desktop it switched to?
+///
+/// [`super::detect_active_session`] answers precisely the question the symptom asks: it reports the
+/// running compositor of our uid, and [`super::ActiveKind::None`] means nothing is drawing
+/// anywhere. Only sound AFTER `stop_session(SESSION_UNIT)` has killed our own managed session —
+/// that kill is a synchronous SIGKILL ([`kill_unit`]), so by the restore's escalation point our
+/// gamescope cannot still be answering for the box.
+fn box_session_live() -> bool {
+    super::detect_active_session().kind != super::ActiveKind::None
+}
+
+/// Poll [`box_session_live`] until it is true or `grace` runs out. [`HandbackWait::Superseded`]
+/// means a client reconnected and took the box over again — the hand-back we were checking is moot,
+/// and every remedy below would now be fighting a live stream for the box's session.
+enum HandbackWait {
+    Live,
+    Superseded,
+    TimedOut,
+}
+
+fn wait_for_box_session(grace: Duration) -> HandbackWait {
+    let deadline = Instant::now() + grace;
+    loop {
+        if takeover_live() {
+            return HandbackWait::Superseded;
+        }
+        if box_session_live() {
+            return HandbackWait::Live;
+        }
+        if Instant::now() >= deadline {
+            return HandbackWait::TimedOut;
+        }
+        std::thread::sleep(HANDBACK_POLL);
+    }
+}
+
+/// **The hand-back's last line of defence for a dark panel**, and the only part of this file that
+/// checks whether the restore it just performed actually WORKED.
+///
+/// Everything above issues a lifecycle verb and reports what systemd said about the JOB. That is
+/// not the same question as "does the box show a picture again", and the gap between the two is
+/// where every "my screen stays black after disconnecting" report lives — including ones whose
+/// trigger nobody has reproduced. So stop inferring the outcome and measure it: if nothing is
+/// driving the panel a full [`HANDBACK_GRACE`] after the hand-back, climb a ladder of remedies,
+/// each of which is a mechanism measured on both distro families (Bazzite `44.20260818`, Nobara
+/// f44, 2026-08-22), and say loudly at every rung what is happening.
+///
+/// 1. **`stop` the autologin unit.** Its login session's script is parked on
+///    `systemctl --user --wait start <unit>` (verified on both images), so stopping the unit
+///    releases that wait, the session exits, and `Relogin=true` logs straight back in — starting
+///    the unit inside a fresh login session with a seat. `stop`, never `restart`: a restart does
+///    NOT release the parked waiter (measured), which is exactly why it cannot rescue a box the
+///    ordinary restart already failed to bring back.
+/// 2. **Restart the display manager.** What the pre-0.31.0 takeover did on every disconnect, and
+///    proven to return the box to game mode. Needs privilege, so it can honestly fail.
+/// 3. **`PUNKTFUNK_RECOVER_SESSION_CMD`**, then an ERROR naming the command a human must run.
+///
+/// **Detached**, and that is not incidental. The restore runs under [`RESTORE_FLIGHT`], which a
+/// reconnecting client must take before it can re-take the box; watching for up to a minute while
+/// holding it would put that whole wait in front of every reconnect. So the caller fires this and
+/// returns, and the watcher stands down by itself the moment [`takeover_live`] says a new takeover
+/// armed — the box belongs to that stream now, and a remedy fired into it would be the bug.
+/// Call it AFTER `clear_takeover()`, or the very first poll reads our own finished takeover as a
+/// new one and stands down immediately.
+///
+/// A box that was already fine costs one [`box_session_live`] call and the thread exits.
+fn ensure_box_session_or_escalate(units: &[String]) {
+    let units: Vec<String> = units.to_vec();
+    std::thread::spawn(move || handback_watch(&units));
+}
+
+fn handback_watch(units: &[String]) {
+    match wait_for_box_session(HANDBACK_GRACE) {
+        HandbackWait::Live => {
+            tracing::info!(
+                "gamescope: the box is driving its own panel again — hand-back complete"
+            );
+            return;
+        }
+        HandbackWait::Superseded => return,
+        HandbackWait::TimedOut => {}
+    }
+    tracing::warn!(
+        secs = HANDBACK_GRACE.as_secs(),
+        units = ?units,
+        "gamescope: NOTHING is driving the box's panel {}s after the hand-back — its screen is \
+         dark. Escalating: stopping the autologin unit so the display manager relogins into a \
+         session with a seat",
+        HANDBACK_GRACE.as_secs()
+    );
+    // Rung 1 — release the login session's parked `--wait start` and let the DM relogin.
+    for unit in units {
+        if let RestoreVerb::Failed(why) = issue_restore_verb(&["stop", unit]) {
+            tracing::warn!(unit, status = %why, "gamescope: could not stop the autologin unit");
+        }
+    }
+    match wait_for_box_session(HANDBACK_RUNG_GRACE) {
+        HandbackWait::Live => {
+            tracing::info!(
+                "gamescope: the display manager relogged the box into its own session — panel back"
+            );
+            return;
+        }
+        HandbackWait::Superseded => return,
+        HandbackWait::TimedOut => {}
+    }
+    // Rung 2 — put the display manager itself through a restart.
+    if let Some(dm) = display_manager_unit() {
+        tracing::warn!(
+            %dm,
+            "gamescope: the box is still dark — restarting its display manager"
+        );
+        match restore_display_manager(&dm) {
+            Ok(()) => match wait_for_box_session(HANDBACK_RUNG_GRACE) {
+                HandbackWait::Live => {
+                    tracing::info!(%dm, "gamescope: the display manager brought the box back");
+                    return;
+                }
+                HandbackWait::Superseded => return,
+                HandbackWait::TimedOut => {}
+            },
+            Err(why) => tracing::warn!(
+                %dm,
+                shape = why.shape(),
+                reason = %why,
+                "gamescope: could not restart the display manager"
+            ),
+        }
+    }
+    // Rung 3 — the operator's own escape hatch, then say what is left to do by hand.
+    if crate::try_recover_session() {
+        tracing::warn!(
+            "gamescope: fired PUNKTFUNK_RECOVER_SESSION_CMD to bring the box's session back"
+        );
+        return;
+    }
+    tracing::error!(
+        units = ?units,
+        "gamescope: the box has NO session driving its panel and every automatic remedy failed — \
+         its screen stays dark until someone runs `systemctl --user restart <unit>` for one of \
+         these, or `sudo systemctl restart display-manager.service`. Set \
+         PUNKTFUNK_RECOVER_SESSION_CMD to let the host do this itself"
+    );
+}
+
 /// Tear down our host-managed session (freeing Steam) and restart the autologin gaming session(s)
 /// we stopped on connect — so the TV returns to gaming mode when no one is streaming. Invoked by
 /// [`start_restore_worker`] once the debounce deadline passes; takes the stopped-unit list so a
 /// cancelled+reconnected window keeps the list for a later real restore.
-fn do_restore_tv_session() {
+fn do_restore_tv_session(verify: bool) {
    // SteamOS: we reconfigured `gamescope-session.target` headless via a drop-in. Restore = remove
    // the drop-in + restart the target (back to the physical panel) — unless the user switched to a
    // desktop session meanwhile, in which case drop the override and leave the desktop alone.
@@ -3574,6 +3794,9 @@ fn do_restore_tv_session() {
                ),
            }
            clear_takeover(); // A3: consumed — after the restart, not before it
+            if verify {
+                ensure_box_session_or_escalate(&[STEAMOS_SESSION_TARGET.to_string()]);
+            }
            return;
        }
    }
@@ -3699,14 +3922,14 @@ fn do_restore_tv_session() {
    }
    // (The idle drop-in is already gone — removed above every early return, so the restarts
    // below bring the box's real session back rather than another idle one.)
-    for unit in units {
+    for unit in &units {
        // Checked, not discarded: this call and the SteamOS `restart` above were the two places
        // that logged an unconditional success over a thrown-away exit status. A `--user start`
        // fails for reasons an operator can act on (the unit is masked, its start limit tripped),
        // and the DM branch thirty lines up already shows the shape — say what happened.
        // `restart`, not `start`: the idle takeover leaves the unit ACTIVE, and `start` on an
        // active unit is a no-op that would report success over a session still running nothing.
-        match issue_restore_verb(&["restart", &unit]) {
+        match issue_restore_verb(&["restart", unit]) {
            RestoreVerb::Done => tracing::info!(
                unit,
                "restored the TV's autologin gaming session (debounce elapsed, no client)"
@@ -3731,6 +3954,12 @@ fn do_restore_tv_session() {
        }
    }
    clear_takeover(); // A3: consumed — and only now, with the restarts actually issued
+                      // …and CHECK that the restart above actually put a picture back on the box's panel, rather
+                      // than trusting the job status to mean that. AFTER `clear_takeover`, which is what makes a
+                      // later `takeover_live()` mean "a client reconnected" — see [`ensure_box_session_or_escalate`].
+    if verify {
+        ensure_box_session_or_escalate(&units);
+    }
 }

 /// Host-lifetime worker that fires a pending [`schedule_restore_tv_session`] once its debounce
@@ -3767,7 +3996,10 @@ pub fn start_restore_worker() -> std::sync::Arc<()> {
                        }
                    };
                    if still_due {
-                        do_restore_tv_session();
+                        // The disconnect restore: verified. This is the path the field reports
+                        // are about, it is on a worker thread with no deadline over it, and a box
+                        // left dark here stays dark until someone walks up to it.
+                        do_restore_tv_session(true);
                    }
                }
            }
@@ -5334,12 +5566,12 @@ mod tests {
        classify_output_size, connected_connector_under, display_manager_unit_under, dm_plan,
        game_hz, gamescope_output_size, hdr_args, idle_dropin_body, idle_dropin_path,
        install_idle_dropin, is_steam_launch, mask_unit, missing_flags, mode_mismatch,
-        nested_wrapper_script, our_wsi_layer_dir, plan_bind, release_autologin_mask,
-        remove_idle_dropin, script_hardcodes_gamescope, sentinel_advanced, shape_dedicated_command,
-        switch_ends_mask_window, takeover_state_is_live, unmask_unit, xwayland_refusal_marker,
-        BindOff, BindPlan, BoxOutputSize, DmHelperError, SessionBind, TakeoverState, WsiPlan,
-        AUTOLOGIN_MASKED, DISTRO_GAMESCOPE_PATH, PENDING_RESTORE, RESTORE_FLIGHT,
-        STOPPED_AUTOLOGIN, WSI_OFF_ENV, X11_SOCKET_DIR,
+        nested_wrapper_script, our_wsi_layer_dir, parse_listed_units, plan_bind,
+        release_autologin_mask, remove_idle_dropin, script_hardcodes_gamescope, sentinel_advanced,
+        shape_dedicated_command, switch_ends_mask_window, takeover_state_is_live, unmask_unit,
+        xwayland_refusal_marker, BindOff, BindPlan, BoxOutputSize, DmHelperError, SessionBind,
+        TakeoverState, WsiPlan, AUTOLOGIN_MASKED, DISTRO_GAMESCOPE_PATH, PENDING_RESTORE,
+        RESTORE_FLIGHT, STOPPED_AUTOLOGIN, WSI_OFF_ENV, X11_SOCKET_DIR,
    };
    use std::time::{Duration, Instant};

@@ -5538,6 +5770,39 @@ mod tests {
    /// drop-in APPENDS the sleep to the box's own session command and both run — the takeover
    /// would then be fighting the very Steam it set out to free, and nothing on the box would say
    /// why. Pins the reset, its order, and that the resolved `sleep` is the one that gets run.
+    /// The `--plain` column the ACTIVE state lives in, pinned against real `systemctl --user
+    /// list-units` output from both distro families. Read the wrong column and a live gaming
+    /// session looks dead (Steam stays held, and our launch collides with it) or a dead leftover
+    /// looks live (the takeover idles a session nobody was in) — both silent on glass.
+    #[test]
+    fn listed_units_take_the_active_column_not_the_load_column() {
+        // Bazzite 44.20260818 and Nobara f44, verbatim (unit / LOAD / ACTIVE / SUB / description).
+        let out = "gamescope-session-plus@ogui-steam.service loaded active running Gamescope Session Plus\n\
+                   gamescope-session-plus@steam.service loaded inactive dead Gamescope Session Plus\n";
+        assert_eq!(
+            parse_listed_units(out),
+            vec![
+                (
+                    "gamescope-session-plus@ogui-steam.service".to_string(),
+                    "active".to_string()
+                ),
+                (
+                    "gamescope-session-plus@steam.service".to_string(),
+                    "inactive".to_string()
+                ),
+            ]
+        );
+        // `loaded` is the LOAD column and must never be mistaken for the state — that is the
+        // off-by-one this pins.
+        assert!(parse_listed_units(out).iter().all(|(_, a)| a != "loaded"));
+        // Anything that is not one of our template's instances is not ours to touch.
+        assert!(
+            parse_listed_units("plasma-plasmashell.service loaded active running Shell\n")
+                .is_empty()
+        );
+        assert!(parse_listed_units("").is_empty());
+    }
+
    #[test]
    fn idle_dropin_replaces_exec_start_rather_than_appending() {
        let body = idle_dropin_body("/usr/bin/sleep");
@@ -26,17 +26,26 @@

 use super::{audio_control, audio_probe, minted, pad_endpoint as pe};
 use anyhow::Result;
-use windows::Win32::Devices::DeviceAndDriverInstallation::SetupDiEnumDeviceInfo;
+use windows::Win32::Devices::DeviceAndDriverInstallation::{
+    SetupDiEnumDeviceInfo, SPDRP_HARDWAREID,
+};

 /// The `Device Parameters` REG_DWORD each punktfunk-minted devnode family stamps on itself. The
 /// VALUE is what differs per family; presence of the NAME is "this one is ours", which is all a
 /// sweep needs.
-const OWNER_MARKERS: [&str; 3] = [
+pub(crate) const OWNER_MARKERS: [&str; 3] = [
    pe::PAD_INDEX_VALUE,
    minted::ROLE_MARKER,
    audio_probe::PROBE_MARKER,
 ];

+/// The Steam streaming hardware ids every audio devnode this product mints is created with —
+/// the second half of the ABANDONED-devnode test in [`owned_devnodes`].
+const MINTED_HWIDS: [&str; 2] = [
+    "ROOT\\SteamStreamingSpeakers",
+    "ROOT\\SteamStreamingMicrophone",
+];
+
 /// What one sweep removed. `endpoint_records` is counted separately from `devnodes` because the
 /// registry half is best-effort by design — see [`delete_endpoint_record`].
 #[derive(Debug, Default, Clone, Copy, PartialEq, Eq)]
@@ -117,11 +126,91 @@ fn owned_devnodes() -> Result<Vec<String>> {
            .any(|m| pe::read_devparam_dword(&set, &did, m).is_some())
        {
            out.push(inst);
+            continue;
+        }
+        // ABANDONED: `ROOT\MEDIA\NNNN` carrying one of our minting hardware ids but no marker at
+        // all — a devnode registered by a host that died before the marker write landed. It is
+        // still bound and still serving endpoints, so leaving it behind is the "uninstalling
+        // punktfunk left Sound settings full of Punktfunk devices forever" report all over again.
+        //
+        // The instance prefix is what makes this safe, and it is NOT redundant with
+        // [`is_removable_instance`]: Steam's own devnodes carry these very hardware ids and are
+        // ROOT-enumerated too, but live under `ROOT\SteamStreamingSpeakers\*` /
+        // `ROOT\SteamStreamingMicrophone\*`. Only `ROOT\MEDIA\*` can have come from our
+        // `SetupDiCreateDeviceInfoW(… DICD_GENERATE_ID)`.
+        if is_abandoned_mint(
+            &inst,
+            &pe::devnode_multi_sz_prop(&set, &did, SPDRP_HARDWAREID),
+        ) {
+            out.push(inst);
        }
    }
    Ok(out)
 }

+/// The ABANDONED-devnode test, split out from the PnP enumeration so the rule that keeps this
+/// sweep off VALVE'S OWN devices is checkable without a live devinfo set. See [`owned_devnodes`].
+fn is_abandoned_mint(instance_id: &str, hwids: &[String]) -> bool {
+    instance_id
+        .to_ascii_uppercase()
+        .starts_with("ROOT\\MEDIA\\")
+        && MINTED_HWIDS
+            .iter()
+            .any(|want| hwids.iter().any(|h| h.eq_ignore_ascii_case(want)))
+}
+
+#[cfg(test)]
+mod abandoned_tests {
+    use super::is_abandoned_mint;
+
+    fn hw(s: &str) -> Vec<String> {
+        vec![s.to_string()]
+    }
+
+    #[test]
+    fn adopts_our_own_unmarked_devnodes() {
+        // What a host that died mid-mint leaves behind, either role.
+        assert!(is_abandoned_mint(
+            r"ROOT\MEDIA\0004",
+            &hw(r"ROOT\SteamStreamingMicrophone")
+        ));
+        assert!(is_abandoned_mint(
+            r"ROOT\MEDIA\0002",
+            &hw(r"ROOT\SteamStreamingSpeakers")
+        ));
+        // PnP casing is not guaranteed on either half.
+        assert!(is_abandoned_mint(
+            r"root\media\0009",
+            &hw(r"root\steamstreamingspeakers")
+        ));
+    }
+
+    #[test]
+    fn never_matches_valves_own_devices() {
+        // THE safety rule: Steam's devnodes carry the very same hardware ids and are ROOT-
+        // enumerated too — only the instance prefix separates them from ours.
+        assert!(!is_abandoned_mint(
+            r"ROOT\STEAMSTREAMINGMICROPHONE\0000",
+            &hw(r"ROOT\SteamStreamingMicrophone")
+        ));
+        assert!(!is_abandoned_mint(
+            r"ROOT\STEAMSTREAMINGSPEAKERS\0000",
+            &hw(r"ROOT\SteamStreamingSpeakers")
+        ));
+    }
+
+    #[test]
+    fn never_matches_other_vendors_or_real_hardware() {
+        // VB-Cable mints ROOT\MEDIA devnodes too — a different hardware id is all that saves it.
+        assert!(!is_abandoned_mint(r"ROOT\MEDIA\0000", &hw("VBAudioVACWDM")));
+        assert!(!is_abandoned_mint(
+            r"HDAUDIO\FUNC_01&VEN_10EC&DEV_0897",
+            &hw(r"ROOT\SteamStreamingSpeakers")
+        ));
+        assert!(!is_abandoned_mint(r"ROOT\MEDIA\0001", &[]));
+    }
+}
+
 /// A devnode this sweep is allowed to remove: ROOT-enumerated, i.e. software-created.
 ///
 /// Every devnode we mint comes from `SetupDiCreateDeviceInfoW(… DICD_GENERATE_ID)` on the MEDIA
@@ -275,13 +275,22 @@ fn ensure_role(role: Role) -> Result<(String, String, Option<String>)> {
    let (hwid, inf) = discover_driver(role.needle(), role.inf_name())?;
    let devnode = match find_role_devnode(role)? {
        Some(inst) => inst,
-        None => {
-            let inst = pe::create_media_devnode(role.desc(), &hwid, |set, did| {
-                pe::write_devparam_dword(set, did, ROLE_MARKER, role.value())
-            })?;
-            tracing::info!(role = role.label(), devnode = %inst, "minted an audio devnode");
-            inst
-        }
+        // Before minting a SECOND devnode, reclaim an abandoned one. Minting is two PnP steps
+        // (register, then mark), and a host that dies between them — the 0.30.0 teardown abort
+        // did exactly this, five times on one box — leaves a registered, driver-bound, endpoint-
+        // serving devnode that carries no marker. Nothing then resolves it: the next pass mints
+        // a fresh one and the orphan lingers as a duplicate "Punktfunk Speakers"/"Punktfunk
+        // Microphone" in the Sound zoo, invisible to the marker-matched uninstall sweep.
+        None => match adopt_orphan_devnode(role, &hwid)? {
+            Some(inst) => inst,
+            None => {
+                let inst = pe::create_media_devnode(role.desc(), &hwid, |set, did| {
+                    pe::write_devparam_dword(set, did, ROLE_MARKER, role.value())
+                })?;
+                tracing::info!(role = role.label(), devnode = %inst, "minted an audio devnode");
+                inst
+            }
+        },
    };
    pe::bind_driver(&hwid, &inf)?;

@@ -531,6 +540,61 @@ fn find_role_devnode(role: Role) -> Result<Option<String>> {
    Ok(None)
 }

+/// Reclaim an ABANDONED punktfunk devnode for `role`, re-marking it so it resolves normally from
+/// here on; `None` when there is nothing to adopt (the ordinary first-mint path).
+///
+/// The shape adopted is `ROOT\MEDIA\NNNN` + the role's Steam hardware id + NO owner marker.
+/// That triple can only be ours: `ROOT\MEDIA\NNNN` is what
+/// `SetupDiCreateDeviceInfoW(… DICD_GENERATE_ID)` on the MEDIA class yields, and STEAM'S OWN
+/// devnodes are enumerated under `ROOT\SteamStreamingSpeakers\*` /
+/// `ROOT\SteamStreamingMicrophone\*` — they carry the same hardware id but never that instance
+/// prefix, which is precisely what keeps this from adopting (and later sweeping) Steam's devices.
+/// A marker of ANY family is left alone: it is a live devnode, ours but spoken for.
+///
+/// Which family the orphan came from does not matter. Every one is a plain instance of the same
+/// Valve driver; roles are ours to assign, and re-marking it here is what makes the assignment
+/// stick across restarts.
+fn adopt_orphan_devnode(role: Role, hwid: &str) -> Result<Option<String>> {
+    use windows::Win32::Devices::DeviceAndDriverInstallation::{
+        SetupDiEnumDeviceInfo, SPDRP_HARDWAREID,
+    };
+    let set = pe::media_class_devs()?;
+    for i in 0.. {
+        let mut did = pe::devinfo_data();
+        // SAFETY: live set; `did` is a live out-param with cbSize set.
+        if unsafe { SetupDiEnumDeviceInfo(set.0, i, &mut did) }.is_err() {
+            break; // ERROR_NO_MORE_ITEMS
+        }
+        let Some(inst) = pe::instance_id(&set, &did) else {
+            continue;
+        };
+        if !inst.to_ascii_uppercase().starts_with("ROOT\\MEDIA\\") {
+            continue;
+        }
+        if !pe::devnode_multi_sz_prop(&set, &did, SPDRP_HARDWAREID)
+            .iter()
+            .any(|h| h.eq_ignore_ascii_case(hwid))
+        {
+            continue;
+        }
+        if super::devnode_cleanup::OWNER_MARKERS
+            .iter()
+            .any(|m| pe::read_devparam_dword(&set, &did, m).is_some())
+        {
+            continue;
+        }
+        pe::write_devparam_dword(&set, &mut did, ROLE_MARKER, role.value())?;
+        tracing::warn!(
+            role = role.label(),
+            devnode = %inst,
+            "adopted an abandoned audio devnode — one of ours whose owner marker never landed \
+             (a host that died mid-mint). Re-marked and reused instead of minting a duplicate"
+        );
+        return Ok(Some(inst));
+    }
+    Ok(None)
+}
+
 /// Find the (exact hardware id, INF path) for one of Steam's streaming drivers: prefer any
 /// installed devnode whose hardware-id list contains `needle` (its `oemNN.inf` is the driver
 /// Windows already trusts), else fall back to Steam's driver directory. Shared with the
@@ -1192,6 +1192,17 @@ fn grant_system_full_control(subkey_path: &str) -> Result<()> {
    result
 }

+/// The MMDevices hive an endpoint's record lives in, chosen by the direction its id encodes
+/// (`{0.0.1.…}` = capture, anything else = render). Render is the safe default: it is what every
+/// non-capture id resolves to, and the pad program only ever has render endpoints.
+fn mmdev_path_for(endpoint_id: &str) -> &'static str {
+    if endpoint_id.starts_with(CAPTURE_ENDPOINT_ID_PREFIX) {
+        MMDEV_CAPTURE_PATH
+    } else {
+        MMDEV_RENDER_PATH
+    }
+}
+
 /// The raw-registry stamp route: repair the Properties key ACL, then write the serialized
 /// values (see [`reg_registry_value`]). Values written here are STORED but possibly not
 /// SERVED until an AudioEndpointBuilder restart — the caller's read-back decides.
@@ -1199,7 +1210,14 @@ fn registry_stamp(endpoint_id: &str, stamps: &[&Stamp]) -> Result<()> {
    use winreg::enums::HKEY_LOCAL_MACHINE;
    use winreg::RegKey;
    let guid = endpoint_guid_part(endpoint_id)?;
-    let path = format!(r"{MMDEV_RENDER_PATH}\{guid}\Properties");
+    // The hive follows the endpoint's DIRECTION. This was hardcoded to Render, which is
+    // invisible for the pad program (its endpoints are render-only) but wrong for the minted
+    // provider, which stamps the virtual microphone's CAPTURE endpoint through the same
+    // writer: the fallback then reached for `…\Render\{capture-guid}\Properties`, a key that
+    // cannot exist, so every registry-route stamp of a capture endpoint failed on a box where
+    // the property store was denied — silently, since the caller degrades to "keeps the
+    // driver's default name".
+    let path = format!(r"{}\{guid}\Properties", mmdev_path_for(endpoint_id));
    grant_system_full_control(&path)
        .with_context(|| format!("make {path} writable (registry stamp route)"))?;
    let key = RegKey::predef(HKEY_LOCAL_MACHINE)
@@ -2109,6 +2127,23 @@ fn pad_capture_thread(
 mod tests {
    use super::*;

+    /// The registry stamp route must reach for the hive matching the endpoint's DIRECTION —
+    /// it was hardcoded to Render, so a capture endpoint's fallback stamp could never land.
+    #[test]
+    fn registry_stamp_hive_follows_the_endpoint_direction() {
+        assert_eq!(
+            mmdev_path_for("{0.0.1.00000000}.{2753f927-2093-4ab4-aa90-9d880e959128}"),
+            MMDEV_CAPTURE_PATH,
+            "the minted microphone's capture endpoint records under Capture"
+        );
+        assert_eq!(
+            mmdev_path_for("{0.0.0.00000000}.{5da9b5c9-8a10-4b54-8cf6-ce02b8354f16}"),
+            MMDEV_RENDER_PATH,
+        );
+        // Anything unrecognised keeps the old behaviour rather than inventing a hive.
+        assert_eq!(mmdev_path_for("nonsense"), MMDEV_RENDER_PATH);
+    }
+
    /// The serialized container blob for pad 0 must be byte-for-byte the on-glass-measured
    /// value, and byte 23 must be the pad index.
    #[test]
@@ -590,6 +590,9 @@ fn watch(

    // ---- Phase 1: wait for the game to show up. ----
    let start_deadline = spawned_at + START_GRACE;
+    // How long the scan has *continuously* seen something for this title — the scan-side twin of
+    // [`SHIM_WINDOW`]. See `scan_settled` below for what it is protecting against.
+    let mut seen_since: Option<Instant> = None;
    loop {
        if cancelled() {
            return;
@@ -713,12 +716,39 @@ fn watch(
            && (child.is_some() || spawned.is_some())
            && spawned_at.elapsed() >= SHIM_WINDOW;
        let live = scanner.find(&shared.spec, shared.launch_stamp);
+        // The same rule for what the *scan* finds, and for the same reason. A store's launch is a
+        // chain of process trees, and the ones that run before the game carry the signals the game
+        // carries: Steam wraps its shader pre-caching and its Proton prefix work in the very
+        // `reaper SteamLaunch AppId=<appid>` the game gets, so the first poll of a launch can match
+        // a tree that was never the game.
+        //
+        // Latching on one poll is what costs, because the two phases are patient in opposite ways.
+        // This one waits [`START_GRACE`] — five minutes — and ending it never ends the session.
+        // Phase 2 waits [`EXIT_CONFIRM`] — three seconds — and ending it *does*. A single sighting
+        // flips the lease from the first to the second, permanently; when that tree then exits with
+        // the real game not yet started, the stream drops mid-launch. On Linux that ended a Rocket
+        // League session 10 s after launch, while Steam was still compiling its shaders, and the
+        // player had to launch a second time to get one that stayed up (field report 2026-08-22).
+        //
+        // Requiring the sighting to persist buys that back for a few seconds of `GameRunning`
+        // latency and nothing else — exit detection is untouched. ⚠ It is a window, not a proof: a
+        // pre-launch tree that outlives the window still latches. Signals sharp enough to tell one
+        // from the other belong in [`crate::procscan`] (where Steam's shader job is already excluded
+        // by name); this bounds what no signal caught.
+        let scan_settled = if live.is_empty() {
+            seen_since = None;
+            false
+        } else {
+            seen_since.get_or_insert_with(Instant::now).elapsed() >= SHIM_WINDOW
+        };
        // A provider saying so is as good as seeing it — better, for a title there is nothing to
        // see: it is the launcher that started the game telling us it did. This is the only way a
        // [`LeaseKind::Reported`] lease ever leaves this phase, and for a `Matched` one it just
-        // gets there sooner than the scan would.
+        // gets there sooner than the scan would. Not gated by the window above: a report is the
+        // launcher's own statement about the game, not an inference from a process that resembles
+        // it, so there is nothing to wait out.
        let said_running = reported().is_some_and(|l| l.running);
-        if !live.is_empty() || child_alive || said_running {
+        if scan_settled || child_alive || said_running {
            known = live.clone();
            publish(&live);
            shared.was_running.store(true, Ordering::Relaxed);
@@ -731,6 +761,8 @@ fn watch(
                title = %shared.game.title,
                kind = kind.as_str(),
                procs = live.len(),
+                // Which processes, not just how many: see [`crate::procscan::names`].
+                names = ?crate::procscan::names(&live),
                "the launched game is running"
            );
            break;
@@ -2019,6 +2051,78 @@ mod tests {
        );
    }

+    /// 🛑 The 2026-08-22 field report: a **pre-launch** process tree must not be mistaken for the
+    /// game.
+    ///
+    /// Steam wraps its shader pre-caching in the same `SteamLaunch AppId=` reaper the game itself
+    /// gets, so the first poll of a launch matches a tree that was never the game. What shipped
+    /// latched on that single sighting: the lease left the start phase immediately, and when the
+    /// compile finished and that tree exited — with Rocket League still starting — the exit watch
+    /// called it the game exiting and closed the session with `APP_EXITED`, 10 s after launch. On
+    /// the player's screen the stream dropped mid-"Processing Vulkan shaders"; their workaround was
+    /// to launch the game twice.
+    ///
+    /// The scanner now knows Steam's replayer by name ([`crate::procscan`]). This pins the bound
+    /// behind that: a matched process that does not outlive [`SHIM_WINDOW`] never arms the exit
+    /// watch, whatever it was — which is what covers the pre-launch trees nobody has named yet.
+    ///
+    /// Ignored by default: it outlives the shim window and then waits out [`EXIT_CONFIRM`], ~11 s.
+    #[cfg(target_os = "linux")]
+    #[test]
+    #[ignore = "drives a real process for ~11s (shim window + exit confirmation)"]
+    fn a_pre_launch_tree_that_exits_never_ends_the_session() {
+        use std::sync::atomic::AtomicUsize;
+
+        // The stand-in has to keep the name `sleep`: coreutils is a multi-call binary that
+        // dispatches on `argv[0]`, and under any other name it exits instantly — which would pass
+        // this test for entirely the wrong reason. (Same trap as the live matcher test in
+        // [`crate::procscan`].)
+        let td = tempfile::tempdir().expect("tempdir");
+        let stand_in = td.path().join("sleep");
+        std::fs::copy("/bin/sleep", &stand_in).expect("copy a stand-in pre-launch binary");
+        let launch_stamp = launch_clock();
+
+        // Alive for less than the shim window — Steam's shader job, in miniature.
+        let mut child = std::process::Command::new(&stand_in)
+            .arg("3")
+            .spawn()
+            .expect("spawn the fake pre-launch tree");
+        // Reaped on its own thread: a zombie keeps its `/proc` entry with an unchanged start time,
+        // so the scan would call it alive forever and the exit under test never happen.
+        std::thread::spawn(move || {
+            let _ = child.wait();
+        });
+
+        static PRE_EXITS: AtomicUsize = AtomicUsize::new(0);
+        PRE_EXITS.store(0, Ordering::SeqCst);
+        let lease = open(
+            LeaseRequest {
+                launch_stamp,
+                // No child and no pid: the scan is the only signal, which is the field-report shape
+                // (`steam steam://rungameid/…` had already handed off and exited).
+                ..req("steam:pre-launch", DetectSpec::dir(td.path()), false)
+            },
+            Box::new(|| {
+                PRE_EXITS.fetch_add(1, Ordering::SeqCst);
+            }),
+        );
+        let shared = lease.shared();
+        assert!(matches!(shared.kind(), LeaseKind::Matched));
+
+        std::thread::sleep(SHIM_WINDOW + EXIT_CONFIRM + Duration::from_secs(3));
+        assert_eq!(
+            PRE_EXITS.load(Ordering::SeqCst),
+            0,
+            "a tree that ran before the game must not end the session when it exits — this is the \
+             field report"
+        );
+        assert_ne!(
+            shared.state(),
+            GameState::Exited,
+            "the game never started, so nothing of it can have exited"
+        );
+    }
+
    /// The whole point of the module, against a real process: a `Child` lease sees its game running,
    /// notices when it exits, and reports that exit exactly once.
    ///
@@ -1111,6 +1111,21 @@ fn spawn_sender(

 use crate::send_pacing::percentile;

+/// How long to ignore further keyframe requests after emitting one.
+///
+/// The window bounds IDR emission in TIME, so it needs an absolute floor rather than a frame
+/// count: it has to outlast the round trip in which the client receives and decodes the IDR it
+/// already asked for. The original `frame_interval * 2` closes long before that at high refresh —
+/// 16.7 ms at 120 fps, while a Moonlight client under loss re-asks every ~30 ms — so every request
+/// passed the gate and the stream became ~32 full IDRs/s, whose bulk causes the very loss that
+/// prompts the next request. That storm sustains itself and reads as stutter at a flat latency
+/// (field log, AMD RX 7800 XT / Bazzite 44 HEVC, 2026-08-22: 1118 requests, 1115 honoured, 3
+/// coalesced). 100 ms matches the encoder-reset backoff below and is about one IDR's service time
+/// on a saturated link.
+fn keyframe_coalesce_window(frame_interval: Duration) -> Duration {
+    (frame_interval * 2).max(Duration::from_millis(100))
+}
+
 /// The encode → packetize loop, over a borrowed capturer. Sending runs on a dedicated thread
 /// (see [`spawn_sender`]) so a send spike can never stall capture/encode.
 #[allow(clippy::too_many_arguments)]
@@ -1194,6 +1209,11 @@ fn stream_body(
    // also fails safe when nobody tells it, but pass the REAL depth: `idd_depth` is configurable
    // and a deeper ring is free pipelining the fallback would forfeit.
    enc.set_input_ring_depth(capturer.pipeline_depth().max(1));
+    // What `enc` was opened against. The capture source can change size/format UNDER this loop with
+    // nothing negotiating it (see the follow-the-source guard below); tracked so the loop can notice.
+    // Both sites that swap `enc` re-bind `frame` with it, so this is always
+    // `(frame.format, frame.width, frame.height)` right after one.
+    let mut enc_src = (frame.format, frame.width, frame.height);
    // FEC overhead percent (Sunshine default 20). Override with PUNKTFUNK_FEC_PCT (0 = data-only).
    let fec_pct: u8 = std::env::var("PUNKTFUNK_FEC_PCT")
        .ok()
@@ -1273,9 +1293,9 @@ fn stream_body(
    // RFI (VAAPI/AMD — `supports_rfi=false`) each one becomes a full IDR, so an un-coalesced request
    // stream turns EVERY frame into a 4K IDR, saturates the send path, and collapses the session
    // instead of recovering. One fresh IDR already resolves all pending loss, so after emitting one
-    // we ignore further keyframe requests for a short in-flight window (~2 frames). NVENC
-    // ref-invalidation (cheap, no IDR spike) is never rate-limited — only full keyframes are.
-    let keyframe_coalesce = frame_interval * 2;
+    // we ignore further keyframe requests for the in-flight window below. NVENC ref-invalidation
+    // (cheap, no IDR spike) is never rate-limited — only full keyframes are.
+    let keyframe_coalesce = keyframe_coalesce_window(frame_interval);
    let mut last_keyframe: Option<Instant> = None;
    // A frame dropped at the pipeline head (below) breaks the reference chain for the following
    // P-frames: the client never receives it, but the encoder advanced its references past it, and —
@@ -1362,6 +1382,7 @@ fn stream_body(
                .context("reopen encoder after rebuild")?;
                // A rebuilt encoder starts unconfigured — same reason as the first open above.
                enc.set_input_ring_depth(capturer.pipeline_depth().max(1));
+                enc_src = (frame.format, frame.width, frame.height);
                supports_rfi = enc.caps().supports_rfi;
                enc.request_keyframe();
                last_keyframe = Some(Instant::now());
@@ -1375,6 +1396,82 @@ fn stream_body(
            }
        }
        let t_cap = tick.elapsed();
+        // Follow an AUTONOMOUS source mode change — one nothing negotiated. The IDD-push capturer
+        // re-opens its ring on a confirmed display-descriptor change (a fullscreen game mode-setting
+        // the virtual display, or an HDR flip changing the format), and the encoder is the one
+        // component that cannot follow a resolution change in place. Every `submit` below then
+        // refuses the frame ("captured WxH != encoder AxB"), and the submit ladder only rebuilds the
+        // encoder IN PLACE — at the SAME configured size — which cannot fix a size the source has
+        // already left, so all five resets burn on it and the stream ends (native/stream.rs carried
+        // the identical gap; a 2026-08-22 field report hit it there at 4K→1080p).
+        //
+        // GameStream has no mid-stream mode-change message, so the client is NOT told: Moonlight
+        // decodes a bitstream that disagrees with the resolution it configured its decoder from.
+        // That is the same bargain the first open above already takes whenever the captured size
+        // differs from the negotiated one (the monitor-mirror case) — tolerant decoders re-init off
+        // the SPS and scale, a strict one (Media Foundation on Xbox) may stall and drop the session.
+        // Taking it here too is strictly better than the alternative, which is ending every stream
+        // the moment a game changes mode.
+        if enc_src != (frame.format, frame.width, frame.height) {
+            match encode::open_video(
+                cfg.codec,
+                frame.format,
+                frame.width,
+                frame.height,
+                cfg.fps,
+                cfg.bitrate_kbps as u64 * 1000,
+                frame.is_cuda(),
+                // Derived from the delivered format, so an HDR flip re-opens at the right depth.
+                gs_bit_depth(frame.format),
+                encode::ChromaFormat::Yuv420, // GameStream stays 4:2:0 — see the first open
+                cursor_blend,                 // same capture cursor mode — see the first open
+                cfg.slices,                   // client slicing ceiling — see the first open
+            ) {
+                Ok(e) => {
+                    tracing::info!(
+                        from = %format!("{}x{} {:?}", enc_src.1, enc_src.2, enc_src.0),
+                        to = %format!("{}x{} {:?}", frame.width, frame.height, frame.format),
+                        negotiated = ?(cfg.width, cfg.height),
+                        "gamestream: the capture source changed mode mid-stream — reopened the \
+                         encoder at the delivered size (the client is not told; a strict decoder \
+                         may not follow — see the note at this guard)"
+                    );
+                    enc = e;
+                    enc_src = (frame.format, frame.width, frame.height);
+                    // A rebuilt encoder starts unconfigured — same reasons as the first open.
+                    enc.set_input_ring_depth(capturer.pipeline_depth().max(1));
+                    supports_rfi = enc.caps().supports_rfi;
+                    enc.request_keyframe();
+                    last_keyframe = Some(Instant::now());
+                    // The old encoder died with its in-flight submissions — their AUs will never
+                    // arrive, so the numbering prediction restarts at `au_seq` (same reasoning as
+                    // the capture rebuild above). Restart the stall clock for the fresh encoder and
+                    // give it the full reset budget.
+                    enc_inflight = 0;
+                    encoder_resets = 0;
+                    last_au_at = Instant::now();
+                }
+                Err(e) => {
+                    // Don't spend the stream on the FIRST failed open: the mode-set that triggered
+                    // this is exactly the kind of event that leaves the driver settling, which is
+                    // what the submit ladder's backoff exists for. Spend the shared reset budget at
+                    // the same exponential pace, re-entering this guard each round — the old encoder
+                    // stays installed and mismatched meanwhile, so it simply keeps failing submit.
+                    encoder_resets += 1;
+                    if encoder_resets > MAX_ENCODER_RESETS {
+                        return Err(e).context("reopen encoder at the source's new mode");
+                    }
+                    let backoff = frame_interval
+                        .max(Duration::from_millis(100u64 << (encoder_resets - 1).min(4)));
+                    tracing::warn!(error = %format!("{e:#}"), reset = encoder_resets,
+                        max = MAX_ENCODER_RESETS,
+                        "gamestream: reopening the encoder at the source's new mode failed — retrying");
+                    next_frame = Instant::now() + backoff;
+                    std::thread::sleep(backoff);
+                    continue;
+                }
+            }
+        }
        // Honor a client recovery request. Prefer reference-frame invalidation (the encoder
        // re-references an older still-valid frame — no costly IDR spike); if the encoder can't
        // invalidate (range too old, or no NVENC RFI) it returns false and we force a keyframe.
@@ -1716,6 +1813,27 @@ mod tests {
        assert_eq!(t.game.title, "/opt/game/run");
    }

+    /// The coalesce window must bound forced IDRs in time, not in frames. A frame-scaled window
+    /// vanishes exactly where it matters most — at high refresh, where a client's recovery spam
+    /// arrives far slower than two frame intervals and so passes the gate every time.
+    #[test]
+    fn keyframe_coalesce_window_outlasts_a_clients_request_cadence() {
+        // The observed storm: a 120 fps session against a client re-asking every ~30 ms. The
+        // pre-floor window was 16.7 ms, so every request became a full IDR.
+        let at_120 = keyframe_coalesce_window(Duration::from_secs_f64(1.0 / 120.0));
+        assert!(
+            at_120 >= Duration::from_millis(100),
+            "120 fps window {at_120:?} does not outlast a ~30 ms request cadence"
+        );
+        // 60 fps was under the floor too (33.3 ms), which is why this is not a 120-only fix.
+        assert!(keyframe_coalesce_window(Duration::from_secs_f64(1.0 / 60.0)) >= at_120);
+        // A slow stream keeps the frame-scaled window — the floor only ever raises it.
+        assert_eq!(
+            keyframe_coalesce_window(Duration::from_millis(200)),
+            Duration::from_millis(400)
+        );
+    }
+
    /// End-to-end check of the send thread: batches pushed on the channel arrive, complete and
    /// byte-identical, at a peer socket via the paced sendmmsg path.
    #[test]
@@ -55,7 +55,14 @@ pub struct DetectSpec {
    /// Steam appid, for titles Steam itself installed (never for non-Steam shortcuts, whose reaper
    /// appid semantics differ — those carry an [`exe`](Self::exe) instead). On Linux this is the
    /// sharpest signal available: Steam wraps every launch — native or Proton — in
-    /// `reaper SteamLaunch AppId=<appid>`, whose lifetime is exactly the game's.
+    /// `reaper SteamLaunch AppId=<appid>`.
+    ///
+    /// ⚠ That reaper is the *appid's*, not the game's. Steam wraps its **pre-launch** work for a
+    /// title in one too — shader pre-caching most visibly — so a launch is a chain of reaper trees
+    /// and only the last of them is the game. Reading the first as the game is what dropped a
+    /// stream 10 s into a Rocket League launch, mid-shader-compile (field report 2026-08-22); the
+    /// shader job is excluded by name in [`crate::procscan`], and [`crate::gamelease`] waits out a
+    /// window before believing any of them.
    pub steam_appid: Option<u32>,
    /// A launcher-stamped environment marker.
    pub env_marker: Option<EnvMarker>,
@@ -1875,6 +1875,14 @@ pub(super) fn virtual_stream(ctx: SessionContext, prepared: Option<PreparedDispl
        mut cur_display_gen,
        built_bitrate,
    ) = pipe;
+    // What `enc` was opened against. The capture source can change format/size UNDER this loop with
+    // no client `Reconfigure` at all — the IDD-push capturer re-opens its ring on a confirmed
+    // display-descriptor change (a fullscreen game mode-setting the virtual display, an HDR flip) —
+    // and every backend's `submit` then refuses the frame. Tracked so the loop can FOLLOW the
+    // source (see the guard in the submit path) instead of dying against an error no in-place
+    // encoder reset can fix. Every site below that swaps `enc` re-binds `frame` with it, so this is
+    // always `(frame.format, frame.width, frame.height)` immediately after one.
+    let mut enc_src = (frame.format, frame.width, frame.height);
    // The display exists now, so the portal has answered: settle the cursor plan against what it
    // actually negotiated rather than what this session asked for (see `settle_portal_cursor`).
    // `mut`: every capture-loss rebuild re-runs `create`, hence re-negotiates.
@@ -2613,6 +2621,7 @@ pub(super) fn virtual_stream(ctx: SessionContext, prepared: Option<PreparedDispl
                );
                cur_mode = new_mode;
                next = std::time::Instant::now();
+                enc_src = (frame.format, frame.width, frame.height);
                // H2/H3: the backend may have honored a different mode than requested — KWin caps
                // a virtual output's refresh, or Windows pf-vdisplay rejects a resolution its
                // running monitor doesn't advertise and the host falls back to the actual display
@@ -2695,6 +2704,7 @@ pub(super) fn virtual_stream(ctx: SessionContext, prepared: Option<PreparedDispl
                    trace.as_ref(),
                    true,
                ) {
+                    enc_src = (frame.format, frame.width, frame.height);
                    // The owed AUs died with the old encoder — same bookkeeping as a resize.
                    inflight.clear();
                    last_au_at = std::time::Instant::now();
@@ -3388,6 +3398,7 @@ pub(super) fn virtual_stream(ctx: SessionContext, prepared: Option<PreparedDispl
                interval = new_interval;
                cur_node_id = new_node_id;
                cur_display_gen = new_display_gen;
+                enc_src = (frame.format, frame.width, frame.height);
                // The rebuild re-ran `create`, so the portal answered again — possibly a different
                // backend's portal (the retarget above), possibly with a different verdict. Settle
                // the cursor plan against THIS display, exactly as bring-up did: the retarget arm
@@ -3650,6 +3661,106 @@ pub(super) fn virtual_stream(ctx: SessionContext, prepared: Option<PreparedDispl
        // exactly that volume, so host apps already tone-mapped the content into it and the honest
        // mastering description IS the client's panel. (The IDD capturer only knows the generic
        // baseline; if the driver ever forwards per-content IDDCX_HDR10_METADATA, prefer that here.)
+        // Follow an AUTONOMOUS source change — one no client `Reconfigure` announced. The IDD-push
+        // capturer re-opens its ring on a confirmed display-descriptor change: a fullscreen game
+        // mode-setting the virtual display (2026-08-22 field report: a 4K60 HEVC session, the game
+        // switched the display to 1080p mid-play), or an HDR flip changing the frame format. The
+        // encoder is the one component that cannot follow that in place (same note as
+        // `try_inplace_resize`), so every `submit` below refuses the frame — and the submit-error
+        // path only rebuilds the encoder IN PLACE, at the SAME configured size, which cannot fix a
+        // size the source has already left. All five resets burn on it and the session ends while
+        // audio keeps running. Reopen at what the source actually delivers instead; the client
+        // learns the new mode from the `Reconfigured` below and its decoder from the opening IDR.
+        if enc_src != (frame.format, frame.width, frame.height) {
+            let actual = delivered_mode(frame.width, frame.height, interval);
+            // Same per-mode pin the client-initiated resize re-resolves: PyroWave's Automatic rate
+            // IS a function of the mode, so carrying the old one across a source-driven mode change
+            // hands it the wrong operating point. H.26x rates are mode-independent (ABR owns them),
+            // and an explicit client rate is never second-guessed.
+            let src_kbps = if bitrate_auto && plan.codec == crate::encode::Codec::PyroWave {
+                resolve_bitrate_kbps_for(plan.codec, 0, &actual, plan.chroma, plan.bit_depth)
+            } else {
+                bitrate_kbps
+            };
+            let opened = crate::encode::open_video(
+                plan.codec,
+                frame.format,
+                frame.width,
+                frame.height,
+                actual.refresh_hz,
+                src_kbps as u64 * 1000,
+                frame.is_cuda(),
+                bit_depth,
+                plan.chroma,
+                plan.cursor_blend,
+                plan.max_slices,
+            )
+            .with_context(|| {
+                format!(
+                    "the capture source changed to {}x{} {:?} mid-session and the encoder could not \
+                     be reopened at it",
+                    frame.width, frame.height, frame.format
+                )
+            });
+            let mut new_enc = match opened {
+                Ok(e) => e,
+                Err(e) => {
+                    // Don't spend the session on the FIRST failed open. The mode-set that triggered
+                    // this is exactly the kind of event that leaves the driver settling — the same
+                    // transient the submit path's backoff exists for ("NVENC session open failing
+                    // after a codec switch", 2026-07) — so spend the shared reset budget on it at
+                    // the same exponential pace, re-entering this guard each round. The old encoder
+                    // is still installed and still mismatched; it simply keeps failing submit until
+                    // an open succeeds or the budget runs out.
+                    encoder_resets += 1;
+                    if encoder_resets > MAX_ENCODER_RESETS {
+                        return Err(e).context("encoder reopen at the source's new mode");
+                    }
+                    let backoff = std::cmp::max(
+                        interval,
+                        std::time::Duration::from_millis(100u64 << (encoder_resets - 1).min(4)),
+                    );
+                    tracing::warn!(error = %format!("{e:#}"), reset = encoder_resets,
+                        max = MAX_ENCODER_RESETS,
+                        "reopening the encoder at the source's new mode failed — retrying");
+                    next = std::time::Instant::now() + backoff;
+                    std::thread::sleep(backoff);
+                    continue;
+                }
+            };
+            if let Some(c) = plan.wire_chunk {
+                new_enc.set_wire_chunking(c);
+            }
+            // A rebuilt encoder starts with the ring bound unset — re-report it, as every other
+            // rebuild site does, or an in-place backend can encode a texture the capturer has
+            // already rotated and overwritten.
+            new_enc.set_input_ring_depth(capturer.pipeline_depth().max(1));
+            tracing::info!(
+                from = %format!("{}x{} {:?}", enc_src.1, enc_src.2, enc_src.0),
+                to = %format!("{}x{} {:?}", frame.width, frame.height, frame.format),
+                "the capture source changed mode mid-session with no client reconfigure — reopened \
+                 the encoder at the delivered size"
+            );
+            enc = new_enc;
+            enc_src = (frame.format, frame.width, frame.height);
+            adopt_built_bitrate(&mut bitrate_kbps, src_kbps, &live_bitrate, &retarget_tx);
+            // The owed AUs died with the old encoder — same bookkeeping as a resize.
+            inflight.clear();
+            last_au_at = std::time::Instant::now();
+            encoder_resets = 0;
+            // A fresh encoder opens on an IDR — anchor the cooldown.
+            last_forced_idr = Some(std::time::Instant::now());
+            // The client's mode slot still says the old size, and its stats/aspect follow it.
+            // Publish what it is really decoding now, exactly as an accepted resize does.
+            live_mode.store(
+                pack_mode(actual.width, actual.height, actual.refresh_hz),
+                Ordering::Relaxed,
+            );
+            let _ = reconfig_result_tx.send(Reconfigured {
+                accepted: true,
+                mode: actual,
+            });
+        }
        let hdr_meta = capturer.hdr_meta().map(|m| client_hdr.unwrap_or(m));
        enc.set_hdr_meta(hdr_meta);
        let mut resend_meta = hdr_meta != last_hdr_meta;
@@ -87,6 +87,26 @@ pub fn resolve(pid: u32) -> Option<ProcRef> {
    }
 }

+/// Short names for the processes a lease adopted, in `procs` order.
+///
+/// Diagnostics only — nothing decides anything on these, and they are deliberately not part of
+/// [`ProcRef`], which is compared for equality. They exist because a launch that adopted the game
+/// and a launch that adopted a *pre-launch* tree logged identically (`procs=1`), which is what left
+/// the 2026-08-22 field report unclosable from its log: the one question worth asking of that line
+/// is which process the lease latched onto.
+pub fn names(procs: &[ProcRef]) -> Vec<String> {
+    #[cfg(any(target_os = "linux", windows))]
+    {
+        let scanner = Scanner::system();
+        procs.iter().map(|p| scanner.name_of(*p)).collect()
+    }
+    #[cfg(not(any(target_os = "linux", windows)))]
+    {
+        let _ = procs;
+        Vec::new()
+    }
+}
+
 /// An out-of-band opinion on whether a spec's game is still running, independent of the process scan.
 ///
 /// Consulted **only to veto** declaring a game gone — never to declare it running, and never as the
@@ -126,6 +126,14 @@ impl Scanner {
        Some(ProcRef { pid, start })
    }

+    /// This process's `comm` — its short name, as `ps` shows it. Diagnostics only (see
+    /// [`super::names`]); `?` for a process that has already gone, which is routine.
+    pub fn name_of(&self, p: ProcRef) -> String {
+        std::fs::read_to_string(self.root.join(p.pid.to_string()).join("comm"))
+            .map(|s| s.trim().to_string())
+            .unwrap_or_else(|_| "?".into())
+    }
+
    /// Which of `procs` are still the same live processes — pid present **and** start time unchanged,
    /// so a recycled pid is never reported alive (rule 2).
    pub fn alive(&self, procs: &[ProcRef]) -> Vec<ProcRef> {
@@ -180,16 +188,29 @@ impl Scanner {
            if let Some(tok) = steam_tok {
                // Both tokens together, exact-matched, so `AppId=57` never satisfies appid 570 and
                // Steam's own (non-reaper) helper steps aren't mistaken for the game.
+                //
+                // …with one exception, because the reaper is *not* only the game's: Steam wraps its
+                // shader pre-caching for a title in the same `SteamLaunch AppId=<appid>` reaper it
+                // wraps the game in, so that job satisfies this recipe exactly while the game has
+                // not started yet. Adopting it points the lease at a tree that exits when the
+                // compile finishes, which reads as the game exiting — on Linux that dropped a
+                // Rocket League stream 10 s into a launch, mid-"Processing Vulkan shaders", and the
+                // player had to launch a second time to get a session that stayed up (field report
+                // 2026-08-22). The payload names itself: `fossilize_replay` is Steam's replayer and
+                // is never a game.
                let mut launch = false;
                let mut appid = false;
+                let mut shader = false;
                for arg in cmdline.split(|&b| b == 0) {
                    if arg == b"SteamLaunch" {
                        launch = true;
                    } else if arg == tok.as_bytes() {
                        appid = true;
+                    } else if program_name(arg) == b"fossilize_replay" {
+                        shader = true;
                    }
                }
-                if launch && appid {
+                if launch && appid && !shader {
                    return true;
                }
            }
@@ -247,6 +268,15 @@ impl Scanner {
    }
 }

+/// The last `/`-separated component of an argv entry — the program's own name, when the entry is a
+/// path to one. Bytes rather than `str` because an argv entry is not required to be UTF-8.
+fn program_name(arg: &[u8]) -> &[u8] {
+    match arg.iter().rposition(|&b| b == b'/') {
+        Some(i) => &arg[i + 1..],
+        None => arg,
+    }
+}
+
 /// Read a `/proc` blob with a hard size cap (see [`MAX_PROC_BLOB`]). `None` when the process vanished
 /// or the file is unreadable — both routine during a scan.
 fn read_capped(path: &Path) -> Option<Vec<u8>> {
@@ -472,6 +502,42 @@ mod tests {
        assert_eq!(pids(s.find(&DetectSpec::steam(57), None)), vec![31]);
    }

+    /// The 2026-08-22 field report: Steam's **shader pre-caching** runs under the game's own
+    /// `SteamLaunch AppId=` reaper, so it satisfies the appid recipe while the game has not started.
+    ///
+    /// Adopting it is what dropped a Rocket League stream 10 s into a launch — the lease called that
+    /// tree the game, and its exit (the compile finishing) the game exiting. The reaper's payload is
+    /// the whole tell, and it is only ever Steam's replayer.
+    #[test]
+    fn steam_shader_pre_caching_is_not_the_game() {
+        let td = fake_proc_root(
+            1000.0,
+            &[
+                // The shader job for this very appid — the game is still being brought up.
+                FakeProc::new(35, 50_000).cmdline(&[
+                    "/home/p/.steam/ubuntu12_32/reaper",
+                    "SteamLaunch",
+                    "AppId=252950",
+                    "--",
+                    "/home/p/.steam/steamapps/common/SteamLinuxRuntime/fossilize_replay",
+                    "/home/p/.steam/steamapps/shadercache/252950/fozpipelinesv6/steamapprun_pipeline_cache.foz",
+                ]),
+                // The game itself, same appid, same reaper. This one IS the game.
+                FakeProc::new(36, 50_000).cmdline(&[
+                    "/home/p/.steam/ubuntu12_32/reaper",
+                    "SteamLaunch",
+                    "AppId=252950",
+                    "--",
+                    "/home/p/.steam/steamapps/common/Proton/proton",
+                    "waitforexitandrun",
+                    "/home/p/.steam/steamapps/common/rocketleague/RocketLeague.exe",
+                ]),
+            ],
+        );
+        let s = scanner(td.path());
+        assert_eq!(pids(s.find(&DetectSpec::steam(252_950), None)), vec![36]);
+    }
+
    #[test]
    fn matches_env_marker_by_exact_value_or_presence() {
        let td = fake_proc_root(
@@ -120,6 +120,14 @@ impl Scanner {
        Some(ProcRef { pid, start })
    }

+    /// This process's image file name. Diagnostics only (see [`super::names`]); `?` for a process
+    /// that has already gone or cannot be opened, which is routine.
+    pub fn name_of(&self, p: ProcRef) -> String {
+        process_start_and_image(p.pid)
+            .and_then(|(_, image)| image.file_name().map(|n| n.to_string_lossy().into_owned()))
+            .unwrap_or_else(|| "?".into())
+    }
+
    /// Which of `procs` are still the same live processes — pid present **and** creation time
    /// unchanged, so a recycled pid is never reported alive (rule 2). Windows reuses pids briskly, so
    /// this check is what makes signalling a remembered pid safe at all.
@@ -0,0 +1,111 @@
+# shellcheck shell=bash
+# Does this box already have every Flathub dep a flatpak manifest declares?
+#   bash scripts/ci/flatpak-deps-present.sh <manifest.yml>   -> exit 0 = yes, 1 = no
+#   bash scripts/ci/flatpak-deps-present.sh --self-test      -> run the asserts below
+#
+# WHY THIS EXISTS: flatpak.yml used to prefetch deps with `flatpak-builder --install-deps-only`,
+# which does NOT mean "install what is missing". builder_manifest_install_dep() branches on
+# `flatpak info --show-commit <ref>` succeeding and runs `flatpak update` for every dep that IS
+# installed (a failed update is fatal there — it never falls back to install) — and
+# ci/flatpak-ci.Dockerfile bakes the whole runtime set, so on a healthy run that flag did nothing
+# except make the build depend on Flathub being up at that minute. On 2026-08-22 it took the job
+# down: dl.flathub.org returned HTTP 404 for one .filez object of the then-current
+# rust-stable//25.08 commit, identically on all 10 retry.sh attempts (~9 min), and flatpak-builder
+# segfaulted on its own error path (rc=139) so the retry wrapper could not tell a dead end from a
+# blip. Nothing about the build wanted that newer commit: the manifest pins a runtime VERSION, not
+# a commit, and the baked one satisfies it.
+#
+# So the workflow asks this first and only reaches for Flathub on a real miss.
+#
+# FAILS OPEN, deliberately: an unreadable/unexpected manifest reports "not present" (1), so the
+# caller does the full install. Silently skipping the install on a manifest we stopped
+# understanding is how you build against the wrong runtime.
+set -uo pipefail
+
+deps_present() {
+  local manifest="$1" runtime rt_ver sdk exts e
+
+  runtime=$(sed -n 's/^runtime: *//p'         "$manifest" | head -1)
+  rt_ver=$(sed -n  's/^runtime-version: *//p' "$manifest" | tr -d "\"'" | head -1)
+  sdk=$(sed -n     's/^sdk: *//p'             "$manifest" | head -1)
+  exts=$(sed -n '/^sdk-extensions:/,/^[^ #-]/p' "$manifest" | sed -n 's/^ *- *//p')
+
+  [ -n "$runtime" ] && [ -n "$rt_ver" ] && [ -n "$sdk" ] && [ -n "$exts" ] || return 1
+
+  flatpak info --user "$runtime//$rt_ver" >/dev/null 2>&1 || return 1
+  flatpak info --user "$sdk//$rt_ver"     >/dev/null 2>&1 || return 1
+  # Extensions are checked for PRESENCE, not version: flatpak-builder resolves their version from
+  # the SDK's own metadata (it prints "Dependency Extension: … 25.08"), never from the manifest.
+  # Any bump that moves them moves runtime-version too, which the two checks above already catch.
+  for e in $exts; do
+    flatpak info --user "$e" >/dev/null 2>&1 || return 1
+  done
+}
+
+self_test() {
+  local rc fails=0 full
+  # NOT `local`: the EXIT trap fires after this function has returned.
+  SELFTEST_TMP=$(mktemp -d) || return 1
+  trap 'rm -rf "$SELFTEST_TMP"' EXIT
+  local tmp="$SELFTEST_TMP"
+
+  cat > "$tmp/ok.yml" <<'YML'
+runtime: org.gnome.Platform
+runtime-version: '50'
+sdk: org.gnome.Sdk
+sdk-extensions:
+  - org.freedesktop.Sdk.Extension.rust-stable
+  - org.freedesktop.Sdk.Extension.llvm20
+command: punktfunk-client
+YML
+  # A manifest this script cannot read (the fail-open case).
+  printf 'app-id: io.unom.Punktfunk\n' > "$tmp/unparseable.yml"
+
+  # Stub `flatpak`: $INSTALLED is the newline-separated set of refs it admits to having.
+  mkdir -p "$tmp/bin"
+  cat > "$tmp/bin/flatpak" <<'STUB'
+#!/usr/bin/env bash
+# only `flatpak info --user <ref>` is exercised here
+# args are: info --user <ref>
+[ "$1" = info ] || exit 0
+printf '%s\n' "$INSTALLED" | grep -qxF "$3"
+STUB
+  chmod +x "$tmp/bin/flatpak"
+  PATH="$tmp/bin:$PATH"
+
+  check() { # <expected rc> <label> <installed set> <manifest>
+    INSTALLED="$3" deps_present "$4"; rc=$?
+    if [ "$rc" != "$1" ]; then
+      echo "FAIL: $2 (expected rc=$1, got $rc)" >&2; fails=$((fails + 1))
+    else
+      echo "ok: $2"
+    fi
+  }
+
+  full='org.gnome.Platform//50
+org.gnome.Sdk//50
+org.freedesktop.Sdk.Extension.rust-stable
+org.freedesktop.Sdk.Extension.llvm20'
+
+  check 0 "everything baked -> skip Flathub"     "$full"                              "$tmp/ok.yml"
+  check 1 "cold box -> install"                  ""                                   "$tmp/ok.yml"
+  check 1 "runtime missing -> install"           "${full/org.gnome.Platform\/\/50/x}" "$tmp/ok.yml"
+  check 1 "sdk missing -> install"               "${full/org.gnome.Sdk\/\/50/x}"      "$tmp/ok.yml"
+  # The regression that started all this: llvm20 fine, rust-stable not.
+  check 1 "one sdk-extension missing -> install" "${full/*.rust-stable/x}"            "$tmp/ok.yml"
+  # A runtime installed at ANOTHER version must not pass just because the name matches.
+  check 1 "runtime at the wrong version"         'org.gnome.Platform//51
+org.gnome.Sdk//51
+org.freedesktop.Sdk.Extension.rust-stable
+org.freedesktop.Sdk.Extension.llvm20'                                                "$tmp/ok.yml"
+  check 1 "unreadable manifest -> fail open"     "$full"                       "$tmp/unparseable.yml"
+
+  [ "$fails" = 0 ] || { echo "$fails check(s) failed" >&2; return 1; }
+  echo "all checks passed"
+}
+
+case "${1:---help}" in
+  --self-test) self_test ;;
+  --help|-h)   sed -n '2,4p' "$0"; exit 2 ;;
+  *)           deps_present "$1" ;;
+esac