MichaelScottsman--LiquidAss-Frame/docs/phase2/wp/C2b-review-R2.md

C2b Launcher ("+" popup): adversarial review R2

Reviewer: independent agent (review R2). Package card: PLAN §2.4 "C2b Launcher ("+" popup)" as amended by R2-15 (13:10); builder's log docs/phase2/wp/C2b.md (Status: "READY: wp.c2b", 09:56). Concept: docs/phase2/concepts/home-apps.md (HA) §4, §5, §11, §12.7, AT-5, AT-12. Steam build 11094443. Device clock = PC clock (EDT). Scratch evidence (probes, JSON): <scratchpad>/c2b_r2/ of this session (first session: <scratchpad>/r2/); kept shots are shots/p2_c2b_r2b_*.png (DOM captures of the popup only, no room).

Status

Done (2026-10-07 16:30 EDT). Sessions: first 11:07–12:13 (cut off), second 13:11–13:58, third 15:15–15:31 (a recheck after a usage-limit pause), fourth 16:22–16:30 (a short recheck after P10's session-5 gate changes). Everything below was measured in the second session unless it says "first", "third" or "fourth session". Verdict: fix.

• Fourth session: the files under test, the C2b log (11:42) and PLAN (13:08) are still unchanged (md5 as below, 16:22); REQ Coordinator->C2b is still open. With P10's 16:09 lab_gates.js, gates barpopup --only aud,motion --flags wp.c2b,c2bToggle --mode pad (16:22, native off) now matches the T3 records across the panel (47 re-matched) and FAILs on 7 real issues: the 5 cut labels and GONE Steam's "Liquid Glass" row and its label (M1, B1); MOTION PASS. glass.py motion barpopup now films the entry (16:24): one animation only, lgs-mat-large-in on the T3 panel (scale and opacity), no content channel (M6 confirmed on the filmstrip). Device after: /tmp/lgs/flags.json absent, no flag-steps records, popups closed by the lab's lock exit, native off.

• Files under test are unchanged across all three sessions and byte-identical on the PC and the Frame (md5 32-launcher.css 65fb3649, 32-launcher.js b232b5b6, theme/popups/32-launcher.json 836848cd; rechecked 15:15). The C2b log is unchanged since 11:42; PLAN is unchanged since 13:08.

• PLAN changed while the package sat at READY: R2-15 adds E-GRID and E-GRID (labels) to §1.16 and PLAN-2b-4 to the card. REQ Coordinator->C2b (R2-15) is open and unanswered.

• Third session: P10 changed lab/lab_gates.js at 14:51 (this review's REQ, label clause 3). gates --only aud was re-run (15:21–15:28, both modes, native off): PLAN-2b-1 now FAILs on the gate too, on 5 cut labels (M1 is corrected from 9 to 5), and T3 + toggle now FAILs with 49 GONE (M1). The lgs off half of G-REMOVE was added: a --stock step (15:20) and a leftovers check (15:17), both PASS.

• Every CSS-tier verdict below comes from a step whose step: line says native=off (steps that ran during another agent's native session were repeated).

• Device note (not caused by C2b's code; resolved by 15:15): from 13:13 to at least 13:58, /tmp/lgs/flags.json held c2bToggle: true and wp.c2b: true as a session override (flagSources: session; no running step held them). A lab step killed before its lock exit leaves its keys there until lgs off. While they were there, the live "+" popup showed the T3 toggle row, which only logs while actionsLive is off, so its Liquid Glass switch did nothing for a wearer. This review overrode them in every step (--flags wp.c2b=false,c2bToggle=false for T1) and did not edit the file. Third session (15:16): the file is gone, the runtime's flagsSet is {} and the launcher module is off ("flag wp.c2b is off"); someone cleared it. See "Requests".

• Device left (15:31): this review's last step (15:28) ended with the theme on, CSS only (lgs-shell native off). No flags-file entries or step flags of this review are left (/tmp/lgs/flags.json absent), every popup this review opened is closed, and the pointer is parked (L.unhover; the lab's lock exit). At 15:31 the theme reads off because another agent's locked pad-bfs --route /library/home --stock step (started about 15:27:30; "theme is off" at 15:28:47) holds the lab lock; the lab gives the theme back at that step's exit. No room imagery from this review is left on either machine (checked 15:16: no image files on the Frame from 13:40–14:00; the PC keeps only DOM captures of the popup).

Verdict

fix: 1 blocker, 7 majors, 10 minors. What passes: G-SIZE, G-TYPE, G-OUTLINE and G-MOTION's automatic part in every tier and both input modes; Reduce Motion; the laser and gamepad paths of "launch a program" and of Steam's own Liquid Glass row (spied, nothing ran); traversal and B; G-REMOVE (flag off and lgs off) and the leak check; the native cover and the popup's +25 mm. What fails on a gate: G-AUD (PLAN-2b-1) in T1, T3 and T3 + toggle, in both modes, with the current gate (M1; T3 + toggle also shows B1's missing row).

Findings

B1 (blocker) The pinned Liquid Glass toggle row fails open: the theme switch goes dead, and presses during the scan are dropped

device/rt/32-launcher.js. Whenever the toggle row is on (actionsLive: true, which V1 sets at release, S26; or the test flag c2bToggle), the wrap removes Steam's own working "Liquid Glass" row from the grid unconditionally (splitPrograms(list, keepLg=false), line 150), before anything is known about the replacement. The replacement Toggle acts only when both hold, and otherwise does nothing, with no feedback and no fallback:

• R.data.useNonSteamApps() resolved and contains the switch (entry, line 175). That hook rests on P2's optional finder (02-react.js line 1137: M.useNonSteamApps ? … : undefined, then null for good; REQUIRED at line 24 does not list it). onActivate is if (entry) runToggle(…) (line 187);
• R.actions.launchNonSteam() finds SteamClient.Apps.LaunchNonSteamApp; otherwise it returns refused and nothing runs.

Live proof (b1.js, --flags wp.c2b,c2bToggle --mode laser, 13:47, native off; Steam's launch call spied, P2 in test mode, nothing could run):

• Finder missing (P2's hook made to return null, exactly what it returns without the finder; restored in finally): 2.5 s after opening, the popup shows a green "Liquid Glass / Glass Shell is on" switch row with data-lgs-wait="scan" for good, 23 cells and no Liquid Glass cell; a click on the row logs no action and calls nothing. The theme cannot be switched off from the UI (HA-10: the switch lives only in "+").
• Real scan: the toggle row appears 48 ms after the click that opens the popup and stays inert until 268 ms; a click at 48 ms logged nothing (dropped), a click after the scan logged launchNonSteam … lgs "toggle".

PLAN R23 says a failed "+" patch falls back to the T1 grid with Steam's own row; here a partial failure (one optional finder, the risk PLAN §5.1 R1 names) removes the working switch and leaves a dead one. Fix: fail closed inside Toggle: render Steam's own row element (props.lg, which carries Steam's onActivate) in the pinned slot until entry is resolved, and whenever the finder is missing or the action comes back refused; or keep Steam's row element as the pinned row and only restyle it. Add a test that breaks the finder (rt.react.test.breakFinder('useNonSteamApps'), or the hook replacement above) and checks that a working switch row remains.

M1 (major) PLAN-2b-1 fails: 5 of 24 labels are cut mid-glyph (E-GRID (labels) clause 3), and G-AUD now FAILs in every tier and both modes

• The cut. Five names draw a broken last glyph and no "…": qBittorrent ("qBittorren" with half a "t"), Chromium ("Chromiun"), LXTerminal ("LXTermina"), RenderDoc ("RenderDo") and Hide/Show Screens ("Hide/Shov" over "Screens") (shots/p2_c2b_r2b_t1_laser.png, p2_c2b_r2b_t3tog_pad.png; zoomed crops in the scratch folder). Each has one unbreakable word wider than the 70 px label. The box computes text-overflow: ellipsis, but it is a display: -webkit-box; -webkit-line-clamp: 2 box, and Chrome draws no ellipsis across there. "Hide/Show" also does not break after "/" (HA §4 and §1.16 clause 3 allow a break there). Corrected in the third session: the second session counted 9 from scrollWidth > clientWidth (73–78 against 70). The four two-line names among them (Frametop Display Settings, Frametop Remote Access, KDE System Settings, Reset Screen Layout) are not cut: their shown lines fit and the clamp draws "…" at the end of line 2 (same shot). The current gate agrees (below).
• The gate, second session. gates barpopup (13:20–13:31, T1 and T3, both modes, native off) reported AUD PASS with 24 E-GRID and 10 "E-GRID (labels)" waivers, whose why said "(3) no glyph cut" for the five cut labels. P10's labelFacts accepted the computed text-overflow: ellipsis without checking that a line-clamped box can draw it.
• The gate, third session. P10 changed labelFacts at 14:51, answering this review's REQ: on a line-clamped box, an overflow across now counts as a cut whatever text-overflow says. gates barpopup --only aud (native off): T1 laser 15:21, T1 pad 15:27 and T3 (wp.c2b,c2bToggle=false) laser 15:24 all report AUD FAIL (exit 1). Each shows exactly 5 issues, "SHRUNK text %{Marquee>Content} "qBittorrent" 81x20 -> 70x17 (E-GRID label criterion not met: … (3) cut across at the run without an ellipsis (a line-clamped box draws none across) …)" and the same for the other four. Each also shows 24 E-GRID and 5 E-GRID (labels) waivers. The waived labels are the five two-line names that end in "…" (Frametop Display Settings, Frametop Input Settings, Frametop Remote Access, KDE System Settings, Reset Screen Layout), each "(3) no glyph cut", which matches the shot. PLAN-2b-1 FAILs on the gate as well as by hand.
• T3 + toggle, fourth session (P10's 16:09 gate, which re-matches records across the T3 move; r4_ga_t3tog_pad.out, 16:22, pad, native off): AUD FAIL, 7 issues: the 5 cut labels above, and "GONE ctl … "Liquid Glass"" and "GONE text … "Liquid Glass"" (Steam's row, removed by the toggle row: B1). 49 stock / 50 themed records, 47 re-matched, 28 waivers. This supersedes the 49-GONE reading below; the gate is no longer vacuous or keyed wrong.
• T3 + toggle, earlier. Second session (13:25 laser, 13:26 pad): AUD reported controls: 0, texts: 0, exempt: 0, so it compared nothing. Third session, with the 14:51 gate (15:21 laser, 15:28 pad): AUD FAIL, 49 GONE (all 49 stock records: the "LAUNCH PROGRAM" header and 24 program rows with their 24 labels). AUD keys a record by its index path from body (pathOf, lab/lab_helpers.js line 325). The T3 panel adds a wrapper div (.lgs-c2b-panel, no lgs- id) around Steam's list, which shifts every path. So 47 of the GONEs are key mismatches: the header and the other 23 cells are present and work (SIZE, TYPE, traversal, the launch spy). Two are real: Steam's "Liquid Glass" row and its label are not in the grid (B1). Either way the T3 + toggle gate fails, and the builder's "T3 with the toggle row passes all five gates" does not hold. The builder needs evidence the gate can read: for example a panel that keeps Steam's own element depth, or P10 matching records across a wrapper by kind and text (REQ below).
• REQ Coordinator->C2b (1) is open: "not built: the word-aware …" is not an accepted deviation (R2-15).

Fix: give single-word overflow its own box that can ellipsize (for example T2 tags data-lgs-fit="word" on labels whose longest word is wider than 70 px, and those labels get display: block; white-space: nowrap; overflow: hidden; text-overflow: ellipsis, two lines kept for the rest), allow the break after "/" (T3 can pass Steam's label with a <wbr> after "/", or T2 can insert one in the text node; the cell's aria-label keeps the stock name, clause 4).

M2 (major) PLAN-2b-4 fails in T1 and T3, in both input modes: the plate is drawn over its neighbours' labels, which stay at full opacity

Measured with glass.py focus barpopup on every cell (gamepad: a D-pad move into the cell, as a user does, because Steam paints .gpfocus in the popup only for gamepad moves; laser: a real CDP hover held 0.9 s with wp.p3):

Tier, mode	Cells	Whole stock name (no overflow)	Inside x 0–300	Overlapped labels ≤ .22	Back at rest 500 ms after	Time
T3 + toggle, gamepad	23	23 / 23	23 / 23 (6–294)	0 / 23: every overlapped label at 1.0	22 / 22	13:37
T3 + toggle, laser	the 10 shortened	10 / 10	10 / 10	0 / 10	9 / 9	13:39
T1, gamepad	24	24 / 24	24 / 24	0 / 24	23 / 23	13:41
T1, laser (wp.p3)	the 8 shortened	8 / 8	8 / 8	0 / 8	7 / 7	13:42

In the shots (p2_c2b_r2b_pf3_14.png gamepad, p2_c2b_r2b_lf3_3.png laser, "Frametop Display Settings") the plate lies across "Dolphin" and "Firewall", whose glyphs show above it and through its .92 grey: two layers of text. HA §4: the neighbours' labels under the plate fade to .22 (the mockup hides them). Fix: fade the other cells' labels while a plate shows (:has() on the grid, keyed on the input mode: .gpfocus under lgs-input-pad, .lgs-dwell:hover under lgs-input-laser, as the plate itself), with P5's fade token. The plate also appears in one frame (m2).

M3 (major) PLAN-2b-2 (cmp against both mockups) was never run

There is no docs/phase2/wp/C2b-cmp.json, and neither home-apps-plus.html nor -plus-t1.html (both C2b's) has a data-id, so glass.py cmp can only place them side by side. The log's look rows say "Matches … by eye", but the live popup differs visibly: cut labels (M1), the plate over its neighbours (M2), focus lighting the disc only where the mockup lights the cell (M4), one generic window glyph where the mockup draws terminal, gear, remote and screens glyphs (m4), the All Apps circle (T3; HA hides it while LQ5 fails, which must then be said), the toggle platter's radius (m5). G-MOCK needs named rects within ±8 px or each difference explained, and an agent verdict. Fix: data-ids in both mockups (panel, header, grid, cell-n, disc-n, label-n, plate, toggle), C2b-cmp.json, glass.py cmp for both.

M4 (major) G-FOCUS (P-14, must) fails: the focused cell is +21 to +42 L over rest; on bright icons even the disc misses +40

glass.py focus barpopup --mode pad (pairs: the cell focused by a D-pad move vs the same cell at rest; luma per VP SHOT), native off:

Tier	Whole cell (the focusable, 72 × 96)	Disc only (52 px)	Toggle row
T3 + toggle (13:37)	+21.8 to +41.7; 22 of 23 FAIL	+30.6 to +77.0; FAIL: Chromium +38.7, RenderDoc +30.6	+78.3 PASS
T1 (13:41)	+21.0 to +42.5; 22 of 24 FAIL	FAIL: Liquid Glass +39.8, qBittorrent +38.9, RenderDoc +30.6	—

The focus light is the disc's fill only (+.32 and a spot), and a bright 34 px icon covers most of the disc. The mockup lights the whole cell (.hp-cell.is-focus white .14) plus the disc's arc and glow. Fix: light the cell as well (a row-style focus fill on the cell, not a scale), then re-measure every cell, including the Liquid Glass cell in T1.

M5 (major) P-18 (must): the first frame of gamepad focus shows about a quarter of the look

theme/32-launcher.css lines 173–174: the disc's focus fill rides a plain transition: background-color 294ms b0 (hover-in), with no first-frame step (PLAN §1.4: "the first frame shows 60 %"; P4/P5 provide --lgs-focus-first, lgs-focus-in and P3's .lgs-focus-in). First session (12:13, native off): first frame +20.8 L over rest against +77.6 L settled on the same disc = 27 % (needs ≥ 60 %). Re-measured (mkfirst.py pairs, 13:53, native off; the disc's transitions caught by a MutationObserver as .gpfocus lands after a D-pad move, held at 16.7 ms): CMake +14.0 L first frame against +55.1 L settled = 25 %; Hide/Show Screens +19.4 against +77.0 = 25 %. The held transitions are background-color and box-shadow, 294 ms; the fill reads white .14 at 16.7 ms (rest .11, final .43). Fix: P4's first-frame recipe for focus (60 % at once, then the rest on hover-in), as the tab bar and rows use.

M6 (major) G-MOTION: the popup's content fades in with its glass instead of after it

DOM probe of the popup's entry (entry.js, 13:43, T3+toggle and T1, native off): the only animation is lgs-mat-large-in (250 ms, linear, backwards) on the glass (the panel in T3, the list in T1); paused at 15 %, 35 % and 60 %, the glass, the header and a cell all read opacity .20 / .48 / .82: content and glass are one channel. HA §4: "glass 0–92 %, content 35–100 %"; contracts/motion.md §2 gives the content channel lgs-mat-content-in on the glass's children; G-MOTION's filmstrip criterion "glass before content on entry" is not met. The swell (s0 1.04 at the bottom centre) and 250 ms are right; Reduce Motion is right (opacity only, 180 ms, entry_reduce, PASS P-56). Fix: lgs-mat-content-in on the card's children. (glass.py motion barpopup could not film this in the first session: its warm-up capture of the hidden barpopup times out; REQ below.) Fourth session, filmed (P10's fixed motion; pre openfast.js: open "+" and return as soon as the first cell exists; --flags wp.c2b,c2bToggle --mode pad, 16:24, native off; shots/p2_motion_c2b_r4_plus_open_strip.png, viewed): frozen: 1, the only animation is lgs-mat-large-in on .lgs-c2b-panel (props scale, opacity; keyframe easings linear(0, 1 92 %) and linear(0, 1 73.6 %)), 250 ms, token PASS, 0 geometry issues. No animation targets the header, the grid or the toggle row, so they ride the panel's opacity: there is no "content 35–100 %" channel. (The first try with the 1.1 s OPEN pre caught 0 animations: the entry had finished before the pause; such a strip is vacuous.)

M7 (major) READY is not supported by the card's tests

• PLAN-2b-1 fails (M1): by hand, and on the current gate in T1 and T3, both modes; T3 + toggle AUD FAILs with 7 issues (the 5 cut labels and Steam's missing "Liquid Glass" row, fourth session).
• PLAN-2b-2 never ran (M3); PLAN-2b-4 never ran (M2; it arrived with R2-15, but the open REQ asks for it).
• G-FOCUS was never run with glass.py focus (M4); P-18 was never measured (M5); the entry filmstrip was never recorded (M6).
• The log's Status still lists the word-aware "…" as a documented deviation, which R2-15 rejects; REQ Coordinator->C2b is unanswered.

Minors

• m1 P-01 (must) on one rule. 32-launcher.css line 294: %{DashboardBarPopupListItem}:is(.gpfocus, :hover) %{Marquee>Content} { display: block; white-space: nowrap; … } has no input-mode scope. Under the laser a stale .gpfocus (IM §3.3) turns that cell's label into one truncated line (a look change on a cell the laser is not on). Scope the .gpfocus half with &:not(.lgs-input-laser) and the :hover half with &:not(.lgs-input-pad), as the plate rules above it are. (Line 314's z-index is paint-neutral.)
• m2 Plate timing and motion. The plate shows at once on gamepad focus and after P3's 80 ms dwell under the laser, and leaves in one frame. HA §4 and §11: after 0.4 s of attention, materialize-in 250 ms, materialize-out 350 ms. Hover-out of the disc uses hover-in 294 ms instead of fade 441 ms (HA §11, tokens.md §3).
• m3 Grid edges close the popup. D-pad traversal (pad.js, T1 13:25 and T3+toggle 13:28; popup nav tree active): every cell reached, every move reversible (P-20 PASS), B closes (P-24 PASS). But Left from column 1, Right from column 4, Down from the last row (T1) and Left/Right/Down from the toggle row all close the popup (focus goes back to the bar, Steam's exit). The 4-column grid has four times the side edges of Steam's list, so a user stepping sideways loses the popup often. visionOS keeps focus inside the container at an edge (P-79's bumper). Consider keeping Left/Right inside the grid in T3 (the panel's Focusable can take onGamepadDirection at the edge) and say so in the log.
• m4 Program glyphs. Six programs without an icon (Camera Switch, Desktop, Frametop Remote Access, Hide/Show Screens, KDE System Settings, Konsole) all get the same window glyph (CSS line 194). HA §0.3 and the mockup give them distinct glyphs (terminal, gear, remote, screens). With labels cut (M1) the identical glyphs leave six cells hard to tell apart.
• m5 Small spec drifts. Toggle platter radius 32 (CSS line 391) against HA §4's 28; label line-height 17 against the mockup's 18; the All Apps circle not built (T3; HA hides it while LQ5 fails: say so in the log and the cmp).
• m6 A second non-Steam scan per press. Toggle calls R.data.useNonSteamApps() (line 174), which runs Steam's ScanForInstalledNonSteamApps again on every open, beside the scan Steam's own popup runs. The switch's command line could be read from Steam's own row element (props.lg) or from the scan Steam already did.
• m7 No exit motion. Steam hides the popup at once on close; HA §4's materialize-out 350 ms is neither built nor recorded as a deviation.
• m8 High Contrast. gates --media contrast (13:30–13:31, T1 and T3+toggle): OUTLINE edge ratio 1.0 on all four sides, which is P4's one allowed 2 px stroke (tokens.md, High Contrast), so a gate question for P10, not a C2b defect. But the cells' focus look does not follow P4's High Contrast focus (white .30 + the stroke): the disc rule is a literal fill.
• m9 P-03. Laser hover on the disc: +9.5 to +26.3 L (RenderDoc +9.5, under 10; Frametop Remote Access, Hide/Show Screens, KDE System Settings +26.3, over 25), lfocus_* 13:39/13:42. On the whole cell +5.3 to +15.8.
• m10 Toggle row semantics. aria-checked is always "true" and the row draws the switch "on" even when the press is only logged (actionsLive off) or refused; with B1 fixed, a refused action should leave Steam's own row in place rather than a switch that claims a state it cannot change.

Function retention (HA §12.7) and the card's other tests

Steam's own launch call SteamClient.Apps.LaunchNonSteamApp is called by Steam's row as Ae=H.useCallback(et=>{SteamClient.Apps?.LaunchNonSteamApp(et.strCmdline)},[]) (read from steamui/sp.js on the Frame), so a spy on that property catches every press. Each probe replaced it with a recorder, pressed A only after checking that Steam's active nav tree was the popup's and the target's nav node had focus, and restored it in finally (restored: true in every run). Nothing was launched and the theme was never toggled.

#	Function	Laser	Gamepad	Evidence
P1	Open the list	PASS: L.click on %{AddWindowButton} (every step)	Steam's own bar path, unchanged; not reproducible in the lab (no controller: activating the bar's nav tree parks focus in vr-null-tree, padopen.js)	Steam's button (C3a)
P2	Launch a program	PASS T1 (spy_t1: a click on CMake → cmake-gui); in T3 the cells are the same Steam rows and handler (not clicked again)	PASS T1 and T3+toggle: A on CMake → cmake-gui (pad_t1, pad_t3tog, misc)	spy
P3	Toggle Liquid Glass	PASS T1 (Steam's row → lgs "toggle"); T3 toggle row → P2 launchNonSteam logged (b1 part 2)	PASS T1 (A on Steam's row → lgs "toggle"); T3 toggle row: A → logged (pad_t3tog, misc)	but B1
P4	Add a desktop window	Not live: SteamVR reports no desktop windows today. Static: the T1 grid also styles the "Windows" list (both lists have a header); T3 rewrites only the last list; Steam's rows and DashboardDesktopWindowClicked handler are kept	as laser	source read
P5	Close the list	Steam's outside click and 2 s auto-close: unchanged (our panel sits inside Steam's popup contents)	PASS: B closes (T1 and T3+toggle)	pad_*

• AT-5: (a) PASS: T1's first cell is exactly "Liquid Glass" (Steam's row, green pip); the T3 toggle row's title is "Liquid Glass". (d) PASS as far as the lab can open it: with the popup's nav tree activated, the entry focus is "Camera Switch", the first program (misc, 13:46); the toggle row is reached only by Down from the last row, and Up from it returns to the cell it came from. (e) = PLAN-2b-3 below.
• PLAN-2b-3 (lq10.js, --flags wp.c2b,c2bToggle,wp.c2a, 13:55, native off): PASS. Main on /library/lgs/steamhome, a click on the toggle row logs navigate /library/home replace=true and then launchNonSteam … lgs "toggle", both logged; the route stays /library/lgs/steamhome; the spy saw no call; the route was put back afterwards. (B1 still applies: the row is dead without the finder.)
• AT-12: PASS: 4 columns of 72 px, display: grid, no flow-children, scrollHeight 576 = clientHeight 576, every visible label ≤ 70 wide and ≤ 2 lines, every plate inside x 6–294; every cell reachable, reversible, B closes.
• G-REMOVE (remove.js, 13:48; native mode was on during the step, which does not touch this DOM and runtime check): PASS. Flag off while the T3 popup is open: the popup closes, 0 lgs-c2b nodes and 0 data-lgs-row / data-lgs-c2b / data-lgs-wait attributes in any popup, c2b.plus patch gone, P2 patchedLeft: 0; reopened: Steam's scan order ("Liquid Glass" first), no tag.
• G-REMOVE, lgs off (third session): PASS.
  • Static: lgs off runs rt_teardown(s, "off") (device/lgs.py line 761), which calls each module's remove(). That is the launcher's own remove(), which closes a "+" popup showing our panel, drops P2's patch and untags Steam's row. P1's leftovers sweep (device/lgs_core.js line 296) runs after it.
  • Real teardowns: the device log shows the theme turned off at 13:32:37, 13:34:25 and 13:36:19, inside the window when the stale session flags kept the launcher module live. leftnow.js (15:17, theme on, no flags) finds 0 lgs-c2b nodes and 0 data-lgs-row / data-lgs-c2b / data-lgs-wait / data-lgs-fit attributes in all 11 popup documents. It also finds no C2b patch in P2's list, patchedLeft: 0, and Steam's own LaunchNonSteamApp.
  • Off state: stockoff.js under --stock (15:20; the theme off inside the lock and given back at its exit; native off) finds no runtime (__LGS_RT and __LGS_MOTION undefined) and no Glass Shell stylesheet. The "+" popup opens stock: flex list, 24 rows of 260 × 40, Steam's scan order (Liquid Glass, Visual Studio Code, qBittorrent), no panel. The only marks are lgs-input-laser and data-lgs-vr-mode on each document root. The lab's own --mode stub sets those two (lab/lab.py lines 568–571) and removes them at the step's exit, so they are not left by lgs off.
• Robustness: 8 open/close cycles with T3+toggle: the module's subscriptions 4 → 4, runtime listeners 9 → 9, popupCreated/popupDestroyed 6/3 → 6/3; at rest getAnimations() is empty (G-MOTION at rest, every gates run); the module runs no interval (two rt.setTimeouts per show, 120 and 500 ms). Fail-closed install: installT3 throws on a missing required finder or Focusable and the T1 grid stays (static); the gap is the optional finder (B1).
• T4: PASS, see Native.

Gates (CSS-only, native=off)

glass.py gates barpopup --pre <OPEN> --flags … --mode … --json --shot p2_c2b_r2b_<name>:

Run	Time	AUD	SIZE	TYPE	OUTLINE	MOTION
T1 laser (wp.c2b=false,c2bToggle=false)	13:20	PASS*, 24 E-GRID + 10 label waivers	PASS (24)	PASS (25)	PASS, top 0.063	PASS
T1 pad	13:21	PASS* (same)	PASS	PASS	PASS	PASS
T3 laser (wp.c2b,c2bToggle=false)	13:21	PASS* (same)	PASS	PASS	PASS	PASS
T3 pad	13:21	PASS* (same)	PASS	PASS	PASS	PASS
T3 + toggle laser	13:25	PASS*, 0 controls, 0 texts compared	PASS (24)	PASS (26)	PASS, top 0.09	PASS
T3 + toggle pad	13:26	as above	PASS	PASS	PASS	PASS
T3 + toggle, --media reduce	13:29	as above	PASS	PASS	PASS	PASS
T3 + toggle, --media contrast	13:30	as above	PASS	PASS	FAIL: ratio 1.0 (P4's HC stroke, m8)	PASS
T1, --media contrast	13:31	PASS*	PASS	PASS	FAIL: ratio 1.0 (m8)	PASS
Third session, P10's 14:51 gate, --only aud:
T1 laser	15:21	FAIL: 5 SHRUNK (cut labels), 24 E-GRID + 5 label waivers	—	—	—	—
T3 laser	15:24	FAIL: the same 5	—	—	—	—
T1 pad	15:27	FAIL: the same 5	—	—	—	—
T3 + toggle laser	15:21	FAIL: 49 GONE (47 path-key mismatches; 2 real, Steam's Liquid Glass row and label: B1)	—	—	—	—
T3 + toggle pad	15:28	FAIL: the same 49	—	—	—	—
Fourth session, P10's 16:09 gate, --only aud,motion:
T3 + toggle pad	16:22	FAIL: 7 (5 cut labels; GONE Steam's "Liquid Glass" row and label, B1); 47 re-matched, 28 waivers	—	—	—	PASS

* AUD PASS was false for the label runs (M1); the third-session re-runs with P10's corrected gate FAIL.

Native (one native-session, 13:46–13:53, P7 complete since 12:45)

run_native.py: two probes, two sgcheck, native gates, two hv --look; the session returned to CSS only ("back to CSS only: yes").

• Cover (nprobe.js): T3+toggle, the reporter's barpopup cover is our panel, [0, 447, 450, 1089] texture px, r 60 (= the CSS card 300 × 726 at (0, 298), radius 40, × 1.5); T1, the list card [0, 558, 450, 978], r 60 (= 300 × 652 at (0, 372)). Both acked (data-lgs-cover), and the CSS glass is dropped there (--lgs-mat-panel-bg: transparent, --lgs-edge: none), so glassd's slab and the CSS card have the same shape. No plates, no layers.
• T4: popups.status(): the live "+" barpopup has entry launcher, z 0.06775 units (25 mm / 369 at S, bar space). In T1 (wp.c2b off) the entry is not used (z 0.01, Steam's 3.7 mm) unless C3a's wp.c3a entry applies.
• G-DEPTH (sgcheck, T3+toggle pad and T1 laser, 13:49): PASS, 0 pops on every surface, depths {0} (C2b adds no crop; P-46, P-47, P-48 not engaged).
• Native gates (T3+toggle, pad, 13:49): all five PASS; AUD's "0 findings" is the T3+toggle case, whose AUD compared nothing in CSS-only mode (M1).
• G-HV: T3+toggle look (frame viewed, then deleted on both machines): one copy of the popup in front of the window, header, cells and toggle row legible at the frame's 960 × 540, no closed rim. Metrics: doubling 0.315 (not doubled), but the auto rect [591, 8, 953, 476] is not the popup, so glassL 72.9 does not describe it; verdict withheld (exit 3, no --rect). T1: the look folder was reaped at 120 s before it could be viewed (this agent was blocked in a locked step); metrics only (doubling 0.28, not doubled; auto rect again not the popup). A G-HV verdict needs --rect on the popup's quad; V1's native gate should take it.

Conformance (VP §6) for this area

Item	Verdict	Evidence
P-01	FAIL on one rule (m1)	CSS line 294
P-02	PASS	laser looks key on :hover (lines 243, 472)
P-03	Mostly PASS (m9)	lfocus_*
P-04, P-05	PASS	no scale or translate on cells, rows or the toggle
P-07	PASS	cells abut, gaps 0 (E-GRID why)
P-08	PASS by E-GRID	cells 72 × 96 popup px (87 × 116 main), whole-cell hit 100 % own
P-13	PASS	one .gpfocus after a D-pad move
P-14	FAIL (M4)	focus_*
P-17	PASS	FocusRing hidden; the disc's glow is blurred, not a ring
P-18	FAIL (M5)	25 % (CMake, Hide/Show Screens), first session 27 %
P-20	PASS	traversal reversible, all reached
P-22	PASS (lab-reachable path)	entry "Camera Switch"
P-24	PASS	B closes the popup
P-38, P-84	PASS	G-TYPE (15 popup px = 18 main)
P-42, P-43	PASS (CSS-only, both modes)	G-OUTLINE
P-46, P-47, P-48	PASS (no pops)	sgcheck
P-52, P-58	PASS	G-MOTION at rest, tokens
P-56	PASS	entry_reduce
P-89	PASS	the plate's :hover reveal has its .gpfocus twin
G-MOTION filmstrip (glass before content)	FAIL (M6)	entry.js; fourth session motion barpopup strip (one animation, on the panel)
G-MOCK	not run (M3)	—

Requests

• REQ C2b-R2->P10: lab/lab_gates.js labelFacts (E-GRID (labels) clause 3) passes a single unbreakable word that overflows a display: -webkit-box; -webkit-line-clamp: 2 box because the box computes text-overflow: ellipsis; Chrome draws no ellipsis across there (live: LXTerminal, qBittorrent, RenderDoc, Chromium, Hide/Show Screens cut mid-glyph while the gate says "no glyph cut"). Suggest: on a clamped box (clampOf(c) > 0), horizontal overflow is a cut whatever text-overflow says. Also: with C2b's T3 panel the AUD diff matches nothing (controls 0, texts 0, --flags wp.c2b,c2bToggle), so G-AUD is vacuous there; and glass.py motion barpopup times out in its warm-up capture (first session, 11:22–11:27). For the fixture: a clamped label with one long word (SHRUNK kept).
  • P10 (2026-10-07 session 5): done, all four, with the step-flag hygiene you raised to the coordinator. (1) Clamp: on a line-clamped box (clampOf(c) > 0) an overflow across is a cut whatever text-overflow computes. Fixture case 12a2 (one long word in a clamped box: SHRUNK kept) and 12a3 (the same word on one line with nowrap + text-overflow: ellipsis: waived). Live T1 (16:07) and T3 (16:11): qBittorrent, Chromium, LXTerminal, RenderDoc keep their SHRUNK with "(3) cut across at the run without an ellipsis (a line-clamped box draws none across)", and Hide/Show Screens too. (2) T3 AUD: three causes, all fixed. The theme switch closed the popup the pre had opened, so the stock or themed snapshot was empty: the pre now runs again (preRerun), and a snapshot that is still empty makes AUD fail as VACUOUS instead of 0 issues. barpopup could name the hidden one of Steam's two popup windows: the shown window wins now, in the JS helpers and for CDP. T3 moves Steam's rows into its panel and sorts them, so their DOM-path keys changed: AUD re-matches a stock record by a unique identity (kind, first class, text; then a class unique in both snapshots), and does not compare records at the same path that are other elements (T3's toggle row sat at the old list box's path). Live, gates barpopup --pre <open "+"> --only aud --flags wp.c2b,c2bToggle --mode laser (16:11; another agent's native session was on, which changes no geometry): 51 stock and 53 themed records, 49 matched across the move, 1 path collision not compared, 23 cells and 5 labels waived, and 7 issues left that are real: the 5 cut labels above, and "GONE … Liquid Glass" (Steam's row and its label: your toggle row replaces it, a different control, so AUD cannot count it as the same function; PLAN-2b-3 and the coordinator judge that). Before: 49 GONE. (3) motion barpopup: the warm-up capture waited for a frame from a popup window that was not shown yet. It is taken after the pre when the pre opens the surface, and CDP captures the shown window. Live (16:04): 34 s, six frames of the "+" opening (strip viewed: shots/p2_motion_p10_s5_plus_open_strip.png), automatic part PASS; its P-53 "window scale 1.0..1.04" false positive is gone too (the 600 px of P-53 are now 600 × m, 498 bar px, not 141). (4) Flags hygiene: each step that writes /tmp/lgs/flags.json also records its keys in /tmp/lgs/flag-steps/; the next lab lock entry puts back the keys of a step whose process is gone (pid and start time), only where the file still holds that step's values. The two keys from 13:13 predate the records, so lgs flags (your REQ to the coordinator) is still the way to drop them. contracts/lab.md §1.1, §4, §10.
  • Reviewer, third session (15:31): the label part is in lab_gates.js (14:51) and confirmed live: the five cut labels now keep their SHRUNK, and the five two-line "…" names stay waived (M1). Still open: (a) with C2b's T3 panel, AUD now reports all 49 stock records GONE, because pathOf keys by index path and the panel's wrapper div shifts every path (15:21, 15:28). Matching an unmatched record by kind and text (for example the stock record's text against the themed records' text, before reporting GONE) would let AUD see that the header and 23 cells are present and Steam's "Liquid Glass" row is not. (b) The motion barpopup warm-up timeout (not re-tried).
  • Reviewer, fourth session (16:30): both confirmed. (a) T3 + toggle AUD re-matches 47 records and reports the 7 real issues (16:22). (b) motion barpopup films the entry (16:24); note for users of motion on a popup: the pre must return as soon as the popup's content exists, or the 250 ms entry is over before the pause and the strip is empty.
• REQ C2b-R2->Coordinator: device hygiene, not C2b's code: /tmp/lgs/flags.json has carried c2bToggle: true and wp.c2b: true as a session override since at least 13:13 (a step killed before its lock exit), so the live "+" shows an inert switch. Someone with the authority should drop the two keys (lgs flags), and P10 might make the step's flags-file entries expire with its runtime overlay (TTL) rather than live until lgs off.
  • Reviewer, third session (15:16): the keys are gone. /tmp/lgs/flags.json does not exist, the runtime's flagsSet is {} and the launcher module is off. Only the TTL suggestion remains, as a guard against a recurrence.

Evidence

Scratch probes and JSON in <scratchpad>/c2b_r2/ (spy_t1.js, pad.js, padopen.js, misc.js, mkpairs.py, mklaser.py, mkfirst.py, entry.js, b1.js, remove.js, nprobe.js, run_native.py, and their .out files; third session: leftnow.js, stockoff.js and the AUD re-runs ga_{t1,t3_laser,t1_pad,t3tog,t3tog_pad}.out); fourth session: r4_ga_t3tog_pad.out, openfast.js, r4_motion.out (vacuous, slow pre) and r4_motion2.out; the label probe p1_t1.js and the gates outputs g_*.out in <scratchpad>/r2/. Kept shots: shots/p2_c2b_r2b_{t1,t3,t3tog}_{laser,pad}.png, p2_c2b_r2b_t3tog_{reduce,contrast}.png, p2_c2b_r2b_t1_contrast.png, p2_c2b_r2b_pf3_*.png and pf1_* (gamepad focus per cell), p2_c2b_r2b_lf3_*.png and lf1_* (laser hover), p2_c2b_r2b_native_t3.png (DOM only).