MichaelScottsman--LiquidAss-Frame/docs/phase2/wp/P5-review-R1.md

P5 Motion library: independent review, round R1

Reviewer: independent adversarial review agent. Date 2026-10-07, 06:40-07:12 (Frame clock). Steam build 11094443. Scope: everything the builder reported for P5 at M3 (MO-1..MO-5, live load, filmstrips, G-MOTION smoke, contract contracts/motion.md, decisions D-P5-1..9, the three generated outputs and springs.py).

Verdict: accept. No blocker or major finding. Every acceptance test re-ran and passed on the live device, in both input-mode stubs and under emulated Reduce Motion. The numbers match the builder's to the last digit. The generator is the real single source: --emit reproduces each output byte for byte, and the C++ header compiles cleanly. Live on real Steam menus and alerts, the keyframes behave as the contract says. Nine minor findings follow, mostly about Reduce Motion polish and contract wording. One incident (a crash of Steam's SharedJSContext during the review) is recorded in §5. It could not be reproduced and is not attributed to P5.

1. What was re-run

Test	Command	Result	Agrees with builder?
MO-1	python docs/phase2/research/springs.py --check	PASS: b0, b15, b20 and b25 are byte-identical to D2 §11.3 and present in 02-motion.nowrap.css; all 3 outputs are up to date; version ed4e044810 in all three. --emit css|js|h also equals each written file byte for byte (after normalising CRLF on stdout); the files are LF	yes
MO-2 (node)	springs.py --test-js	PASS, worst 1.01e-3 (sheet-out, at the settle clamp); create() is done exactly at 441 / 488 / 735 / 210 / 250 ms; retarget jumps ≤ 3.6e-5 in value and ≤ 4.7e-4 /s in velocity	yes
MO-2 in CEF	springs.py --live cef (06:44)	PASS, worst 1.01e-3; 5e-10 before the clamp for every token; create('depth') done at 441 ms; runtime global restored	yes
MO-3	springs.py --live mo3 (06:45)	PASS 20/20, nothing left, stage removed, timeline advanced 1866 ms. Samples identical to the builder's (e.g. lgs-mat-glass-in coverage .565 at f .15 and 1 at .35, content 0 until 35 %; lgs-sheet-out .747 / .209 / .070 at .15 / .35 / .5; lgs-mat-glass-out scale 1 until 55 %, 1.0167 at 75 %)	yes
MO-3 pad / laser	--live mo3 --mode pad, --mode laser (06:56)	PASS 20/20 each. The stub reports mode=… via classes (no rt.input): P3's adapter is off, so the class fallback is used. P5's keyframes do not depend on the input mode	yes
MO-4	--live mo4 (06:49, 06:55), --live mo4 --mode pad (06:57), --mode laser (06:59)	PASS 20/20 in all four: matchMedia(reduce) true; tokens 180 / 200 / 150 / 180 ms; --lgs-ease-snappy = b0; every keyframe opacity-only or empty; every duration 150–200 ms; nothing left	yes
Reduce Motion exit probe (new)	glass.py js <one .lgs-mat-out node, removed> --media reduce	180 ms, but opacity holds at 1 until 55 % and then fades over 81 ms (finding m1)	builder did not sample it
MO-5	frame_ssh run "~/.local/share/glass-shell/native/glassd/glassd --selftest phase"	PASS, worst 0.0010 (sheet-in 735, sheet-out 515, linear 250 / 350, reduce 180). The installed binary reports header 2428c76f74. Since then the header changed only in kVersion: git diff f5190bb is one line, and D-P5-8 changed CSS composites only. test_phase.sh (GL-4) was not re-run here because it opens SteamVR overlays; P9's reviewer re-ran it (PASS, 40 values)	yes
Header compile	g++ -std=c++17 -Wall -Wextra -Wpedantic -O2 on the Frame (GCC 15.1.1) of a test that includes motion_tokens.h; /tmp/lgs/p5rev deleted after	PASS, no warnings. |y| ≤ 0.00101 at every token's settle time and > 0.001 two ms earlier, so the settle times are the true crossings; sample(t90) within .005 of .9; coverage(.15) .5000; kPhaseUp / kPhaseDown = sheet-in / sheet-out	yes
check-theme	python glass.py check-theme	PASS (41 files)	yes
Live load	glass.py js; glass.py js … --in vr:systemui	rt.shared = [motion], rt.shared.motion === window.__LGS_MOTION, version ed4e044810, frozen. P1's runtime lists __LGS_MOTION in status().globals, so teardown deletes it. In vr:systemui (CSS-only mode) the tokens and keyframes are bundled (--lgs-d-snappy 488ms, lgs-mat-glass-in rule present), and __LGS_MOTION is absent, as expected without the scene graph. 0 animations at rest	yes
G-MOTION smoke	glass.py gates main --route /zoo/buttons --only motion --pre <lgs-mat-glass-in on mat-in, lgs-sheet-in on sheet-in, lgs-fade-in 300ms ease; removed after 6 s>	Only finding: lgs-fade-in on DIV: duration 300 ms. P5's keyframes, with their composite per-keyframe curves, pass the token checks	yes
R8 perf (new)	glass.py shot main p2_p5rev_perf --route /zoo/buttons --back --pre <rAF timing: baseline 700 ms, then 6 visible 320 × 76 glass nodes on .lgs-mat, .lgs-mat-out, lgs-toast-in; removed>	Baseline median 11.1 ms, max 11.2. .lgs-mat ×6: max 22.3 ms (one dropped frame at the class flip). .lgs-mat-out ×6: max 12.3. lgs-toast-in ×6: max 11.7. No frame > 34 ms, so R8's precondition holds for P5's blur ramps	builder had no perf evidence (m3)
Real Steam surfaces (new)	glass.py shot main p2_p5rev_{menu,alert}_f<f> --route /zoo/modals --back --pre <open the zoo's no-op "Simple Menu" / "GenericConfirmDialog Dialog", pause our lgs-* animations at f>; plus menu_f0.999 and menu_rest	C1c's real menu: lgs-morph 607 ms clips from the top-centre capsule (f .15: inset 30 % / 60 %; f .35: 8 % / 17 %, rounded corners), items 0 at f .15 and .57 at .35. Real alert: lgs-mat-large-in opacity .20 and scale 1.0157 at f .15, content 0	builder tested its own scenes only
Filmstrips	looked at shots/p2_cmp_p5_motion.png, p2_p5_film_b.png, p2_p5_film_a_reduce.png	See §3	yes

2. Findings

None blocker, none major.

Minor

• m1 Reduce Motion: the small-glass exit is a 99 ms hold, then an 81 ms fade. springs.py's Reduce Motion block (lines 584–598) redefines the durations and the bounce curves, but not --lgs-ease-mat-out (linear(0, 0 55%, 1)). JS has the same gap: timing('mat-out', {reduce: true}) returns {duration: 180, easing: 'linear(0, 0 55%, 1)'}.
  • Live under --media reduce, a .lgs-mat-out node samples opacity 1 / 1 / 1 / .889 / .556 / .222 at f 0 / .25 / .55 / .6 / .75 / .9.
  • MO C8, MO §9 and D-P5-1 call for a cross-dissolve of 150–200 ms. P3's tooltips (IN-13) and every other .lgs-mat-out or --lgs-motion-mat-out caller get this abrupt exit.
  • MO-4 passes because it only checks duration ≤ 200 ms.
  • Fix: in the reduce block, emit --lgs-ease-mat-out: linear, and in easingOf() return linear for mat-out when reduce.
• m2 Contract §1.6 overclaims. "So a call site that uses the tokens and keyframes is Reduce-Motion-correct without its own media query."
  • That is true for P5's keyframes, but not for the contract's own §1.1 idiom, transition: scale var(--lgs-motion-snappy). Under Reduce Motion that is still a 180 ms scale (or translate) animation, which P-56 forbids. Callers in the tree use the idiom: theme/41-home.css 770, 990, 1008, 1029, 1037, 1252–1260 and theme/10-primitives.css 1087–1090, 1147.
  • Reword the sentence to say keyframes only. Add the rule: a transitioned scale/translate needs its state value at rest under Reduce Motion, as P4 does with --lgs-press-grow: 0. Owners' P-56 runs (gates/motion --media reduce) will catch the rest.
• m3 R8 precondition was not measured by P5. MO R8 amends the filter-animation ban only "if perf shows no frames > 34 ms". P5 introduced the backdrop-filter/filter ramps, but its log has no perf row.
  • This review measured it (§1, R8 perf: no long frames; one 22 ms frame when six nodes start together).
  • Add that row, or the equivalent, to wp/P5.md so V2 does not have to rediscover it.
• m4 lgs-toast-in keeps a backdrop-filter animation alive for 488 ms. The ramp ends at 47.1 % (230 ms), but the animation object animates backdrop-filter, background-color and box-shadow for the full snappy time. That exceeds R8's "≤ 350 ms" in letter, though not in visible effect.
  • Measured cost is nil (toast ×6: max 11.7 ms).
  • Either note it in the contract as an accepted reading of R8, or split the toast into a 250 ms glass keyframe plus a snappy translate (lgs-shift-in) on the same node.
• m5 lgs-knob-lift is a linear trapezoid, not a spring. MO §4.15 specifies a lift on interactive (b15, 210 ms) and a settle on snappy. The keyframe instead ramps linearly in 98 ms, holds to 65 %, then returns linearly in 171 ms (var(--lgs-d-snappy) linear).
  • Visible in p2_p5_film_b.png: the knob is already fully translucent at f .15. The return has no b15 ease-out.
  • Have springs.py generate per-keyframe literal linear() curves for the 0 % and 65 % blocks (the b15 rise and the b15 settle, rescaled to their segments), as it already does for the coverage curves of the other keyframes.
• m6 create() accepts non-finite targets, and audit() throws without a document. __LGS_MOTION.create('depth').to(NaN, 0) gives value() = NaN and done() false for 4 × settle, then NaN forever. P7's depth channel would push NaN to the scene graph. audit() with no document throws TypeError. Ignore non-finite to() targets (keep the old one), and return [] from audit() without a document.
• m7 Documentation drift.
  • wp/P5.md Status says REQ P5->P10 and REQ P5->P7 "are answered by their owners", but P5->P7 is still - [ ]. It is optional, so say so.
  • contracts/motion.md §2 says lgs-mat-glass-out drops coverage opacity "over the last ≈ 30 %". Measured, it is 1 until 86.5 % of the time and falls over the last 13.5 % (30 % of m, not of time).
• m8 Evidence naming and overwrite. springs.py --live mo3|mo4 always writes shots/p2_p5_mo3.png / p2_p5_mo4.png, which show only /zoo/buttons after the step, so any re-run overwrites the evidence (this review did, 06:50–06:58, with equivalent content). The filmstrips are named p2_p5_film_*, while PLAN §2.1 asks for p2_motion_<id>_<interaction>_<f>.png. Suggest p2_motion_p5_<set>[_reduce].png, and either drop the uninformative mo3 and mo4 shot or name it per run.
• m9 The morph clip also clips hit testing for about 200 ms. In Chromium, clip-path limits hit testing, so while lgs-morph grows, a laser click on an item outside the clip misses it. MO §4.6 says "Items accept input from the first frame". In practice the window is too short to aim at an item (the clip covers about 83 % by 212 ms in the real menu). Gamepad focus is unaffected. Informational: mention it in the contract next to R9.

Accepted deviations (checked, no finding)

• D-P5-6: no @property in 02-motion, although the card lists it. P4 registers --lgs-hover, --lgs-press and --lgs-focus (theme/00-tokens.nowrap.css), and a second registration would override P4's. The decision is sound and documented.
• D-P5-8: large glass fades over min(1, m/.8). In the storyboard comparison it removes the pop-in, and it matches the mockup's sample points (alert .20 at 66 ms, sheet .50 at 110 ms).
• D-P5-9: no overshoot on the clip. MO §6.4 itself prescribes b0 for the clip.

3. Visual review (against "feels native visionOS, no outlines")

• Storyboard comparison (shots/p2_cmp_p5_motion.png, live rows under window-nav-motion.html). All five rows agree frame by frame:

  • Menu morph: clip about 40 % at 91 ms; content in from 15 % to 50 %.
  • Alert: faint glass at 66 ms; content from 35 %; swell 1.019.
  • Sheet: .97 → 1; glass half at 110 ms; content 25–70 %.
  • Dismiss: content out first, then the glass.
  • Route: old content out in 150 ms; new content in with 16 px parallax.

  Known differences: the clip has no overshoot (D-P5-9), and the storyboard has a room photo behind the window. No frame shows a closed outline or a hard rim. The keyframes animate box-shadow none → rest, not a stroke.

• Set B (p2_p5_film_b.png): small glass leads with frost, and content sharpens from 35 % to 100 %. On exit the content goes first. The toast materializes with an 8 px drop, and focus starts at .6. The knob lift reads as "into clear glass", but linearly (m5).

• Reduce Motion (p2_p5_film_a_reduce.png): every row is a plain dissolve with no scale, clip morph or parallax. That is correct, apart from the exit timing in m1.

• Real Steam surfaces (shots/p2_p5rev_menu_f0.15.png, _f0.35, _f0.999, _rest, p2_p5rev_alert_f0.15.png, _f0.35):

  • On C1c's real zoo menu, the clip grows from the top-centre capsule with rounded corners and no edge line, and the items fade in late as specified.
  • The alert is faint at f .15, half there at .35, with the small swell.
  • P5's part reads as native Liquid Glass motion.
  • Not P5's, recorded for C1c and P4: at rest (_rest, identical to _f0.999) the zoo menu's thick glass shows the buttons behind it sharp, with no visible backdrop blur. The morph is not the cause, since the rest frame and the paused end frame are the same.

4. Checklist (a)–(f)

• (a) Ownership.
  • filesChanged covers only P5's five owned files (PLAN §2.6), its evidence log and its own shots/p2_p5_* / p2_cmp_p5_*.
  • git status shows no P5 file outside its patterns: lab/lab_motion.js is P10's, and no theme/_wip motion file exists.
  • The other agents' edits to wp/P5.md are the sanctioned REQ answers from P4, P8, P9 and P10.
• (b) Safety.
  • P5 writes nothing at runtime. Its live steps are locked glass.py steps whose stage nodes are removed; the CEF test restores the runtime's global.
  • Nothing persists, and nothing is installed or restarted. /tmp/lgs has no p5* folder.
  • lgs off: P1 tracks __LGS_MOTION as a runtime global and deletes it (status().globals). P8 removes it in vr:systemui (P8's selftest plates).
  • Never-list respected: zoo demo menus and dialogs were only opened, then closed by the step.
• (c) Robustness. motion.js has no timers, listeners or side effects beyond the frozen global. A throw would be caught by P1's fail-closed loader. Nothing runs at idle: 0 lgs-* animations in main and in vr:systemui after every step. Weak spots are listed in m6.
• (d) Conformance.
  • P-52 (iterations 1, nothing left), P-53/P-54 (≤ 16 px translate defaults, swell ≤ 1.019 on large glass), P-55 (bounce ≤ .25), P-56 (opacity-only ≤ 200 ms under reduce) and P-58 (tokens; allowed / isToken* used by G-MOTION): pass.
  • P-06: --lgs-delay-dwell 80 ms provided.
  • MO C1–C3 and C8: pass, except m1. MO §4.15: m5.
• (e) Function retention. P5 adds no function of its own. No keyframe touches pointer-events, visibility or layout, and every keyframe ends at the element's own rest (MO-3 rest diff empty in pad and laser). The one input-side effect is the transient clip in m9.
• (f) Visual quality. See §3. The motion reads as Liquid Glass: glass first, content second, springs settling without bounce on large surfaces, and no outlines.

5. Incidents during the review

• 06:49:56: Steam's SharedJSContext crashed.
  • lgs.log shows "an evaluation timed out" at 06:49:46, then Runtime.evaluate: Target crashed.
  • The gamepadui windows were recreated (ids 7664xxxx → 8510xxxx). The theme was re-injected before Steam's webpack modules loaded, and showed 735 unresolved tokens and classModules 3 until about 06:53:53.
  • The crash came in the same second that this review's first springs.py --live mo4 step ended (Reduce Motion had been emulated since 06:49:31). Other agents' steps were interleaved: theme reloads at 06:49:24 and 06:49:31.
  • The same MO-4 step then ran three more times (06:55 default, 06:57 pad, 06:59 laser), plus one Reduce Motion probe, with no crash, no timeout and no theme loss. MO-3 also ran three times without a problem.
  • The cause is not established. It is recorded so the coordinator can correlate it with other agents' logs. It is not counted as a P5 finding.
  • The three --live runs that fell into the broken window (06:50–06:51) failed with lgs-on False / NO ANIMATION and are not counted.
  • The theme recovered by itself. Turning it on before Steam's modules exist, and never re-resolving the tokens afterwards, may deserve a look by P1/P8.
• Re-running springs.py --live mo3|mo4 overwrote shots/p2_p5_mo3.png and shots/p2_p5_mo4.png (m8).
• Device left as found: theme on, CSS-only mode; no flags (status().flagsSet {}); no Reduce Motion emulation; no test node (#lgs-p5-test, #lgs-p5rev-perf, #lgs-p5rev-gates absent); 0 lgs-* animations; menus closed by each step; /tmp/lgs/p5rev deleted. Review shots kept: shots/p2_p5rev_* (our own nodes and the zoo's demo page; no room imagery, no personal names).