MichaelScottsman--LiquidAss-Frame/docs/phase2/wp/P6-review-R1.md

P6 Compositor, Steam side: review R1

Reviewer: independent adversarial review of P6 round R1 (builder report: M3 / complete). Date 2026-10-07, 07:00-07:45. Steam build 11094443 (/tmp/lgs/steam-build.json). Every device step ran through python glass.py ... or the steam-frame-ssh helper under the lab locks; native work only inside glass.py native-session.

Verdict: fix (2 major, 10 minor, no blocker).

The reporter, fragments and popup wrapper work as the card asks. I re-ran every acceptance test on the Frame and RP-1 to RP-8 and RP-X pass, with one RP-6 harness flake. Removal is clean.

Two things block acceptance:

1. In native mode, 05-native.css does not drop the CSS glass on every acknowledged element. The popped bottom ornament keeps its CSS E3 edge and its liquid tint above glassd's slab. The window cover keeps a CSS top sheen.
2. The popup wrapper ignores flag changes while a popup is shown. A flagged host entry does not apply when its flag turns on. A flagged change would stay after the flag turns off, so a lab --flags step could leave its effect behind.

Re-run results

Harness: the builder's p6test.py copied to /tmp/lgs/p6r/. The files under test were taken fresh from the synced tree; their md5 hashes match the working tree. p1_layers.js is checked as b8fea9a:device/lgs_layers.js with __LGS_LAYERS renamed. My own probes are rv6*.py, rvflag.py and rvperf.py in the same folder. That folder is in RAM: the files hold no camera data and no names.

Test	Command	Result	Output
RP-1	p6test.py rp1	PASS. The surfaces are byte-identical with Phase 1 on /library/home, AllGames, AllGames after L.pad('down'), and /library/app/735580. Both the fragment form and the Phase 1 object form match	rp1.json
RP-2	p6test.py rp2	PASS. 20 plates with the right materials and rects. Window mode is 1920 × 984 r 81, window-full is 1920 × 1080, and windowless has shapes: [] and 4 mosaic bands	rp2.json
RP-3/4/5	p6test.py rp345	PASS on all three	rp345.json
RP-6	p6test.py rp6, run twice	First run FAIL: the pad state reported card: null 400 ms after the D-pad move, while laser, sweep and rest all passed. Second run PASS. An independent probe (rv6.py) reported the pad lift on the focused poster at 0.3, 0.8 and 1.6 s after each of 4 moves (down, down, right, up). See minor m1	rp6.json, rp6b.json, rv6.json
RP-7	p6test.py rp7	PASS, 13 of 13 checks. Systemui z changed 0.0919 → 0.0500 and meters-per-pixel ×1.25; both were restored. Both TTL cases restored Steam	rp7.json
RP-8	glass.py native-session --step "eval SharedJSContext @/tmp/lgs/p6r/rp8.js" --step 'js "..."' (real glassd)	PASS, 8 of 8. Plates were acked in 600 ms; fill transparent, shadow none; 4 tagged pops = 4 acked. Back to CSS only. The harness checks too little, though: see major M1	scratchpad/rv/native1.txt
RP-X	p6test.py rpx	PASS, 18 of 18	rpx.json
Cost	rvperf.py: rpperf with the injection order swapped	Not convincing. The builder ran Phase 1 first and got v3 2.0 ms against Phase 1 3.55 ms. With v3 first, v3 takes 3.4 ms and Phase 1 1.87 ms. Whichever runs first pays for the layout flush, so the two cost about the same. See m2	rvperf.json
G-DEPTH (info)	native-session --step "sgcheck --route R" for AllGames, Home and the game page	FAIL on all three. Every failure is a legacy pop: R2 depths 3-7.5 mm are outside the set; R5 tab-arrow is 32 × 32 and play is 240 × 54; R7 finds 5 depths on the game page. Expected until the areas supersede the legacy rules (PLAN §6); see m3	native1.txt
G-REMOVE (P6 part)	glass.py js after the native sessions	PASS. No __LGS_LAYERS*, __LGS_POPUPS* or __LGS_P6* global; no data-lgs-cover, -pop, -plate-ack, -noslab or data-lgst-* attribute and no fixture in any window; no own method on vrPooledPopupStore; lgs-native is on 0 windows	stdout
Flag following (new)	rvflag.py (wp.p6 on, test popup, apply([00-base, entry with "flag": "p6RevFlag"]), then the flag pushed on and popped)	FAIL. The request stayed z_pixels 60 for 2 s with the flag on, although the entry asks for z 0.05. See M2	rvflag.json
Native visuals	native-session --step "shot main p2_p6_rev_native_{allgames,home}", plus CSS-only shots p2_p6_rev_css_*; native PNGs composited over a glass tone rgb(62 66 78)	Clean except the footer edge and tint and the window sheen (M1). The overlapping header on AllGames (the search field over the tab row) appears in CSS-only mode too: not P6	shots/p2_p6_rev_*.png

Findings

Major

M1. Native mode keeps CSS glass on acknowledged elements that are not painted through tokens.

• Files: theme/05-native.css §4, contract reporter.md §8.
• What I measured (native session, 07:24 and 07:31, /library/tab/AllGames and /library/app/735580):
  • #Footer is popped (data-lgs-pop="after", acked as footer).
  • Its computed --lgs-edge is still liquid, so P4's E3 arc on #Footer::before stays: a conic gradient with a white .682 arc at the top.
  • #Footer::after keeps C1a's literal liquid tint (linear-gradient(rgb(255 255 255 / .12), rgb(255 255 255 / .03) 55%) over a dark tint, with backdrop blur).
  • In p2_p6_rev_native_allgames.png, the footer's top edge at (960, 942-943) is white rgb ≈ 200 at alpha 196, the same as in CSS-only mode. glassd's liquid slab under that crop draws its own crescent, so the headset shows a doubled edge and a doubled tint.
  • The acked main cover keeps %{BasicUiRoot}::after with a white .08 → 0 top sheen (C1a, 20-shell.css around line 91) above glassd's window glass.
• Why it happens:
  • §4 sets --lgs-edge: none only on [data-lgs-pop~="self"]. When the popped part is a pseudo-element, the host's E3 edge belongs to that pseudo capsule, and it is never turned off.
  • §4 zeroes only --lgs-mat-* tokens, and C1a paints these capsules with literal colours.
  • Contract §8 promises "tint, shading and E3 edge go" for every popped part. RP-8 passed because it checks only --lgs-glass-bg on the footer and one token on Back.
• Fix:
  • In 05-native.css, turn off --lgs-edge on hosts of popped before/after parts that are not glass themselves (today only #Footer qualifies; a generic [data-lgs-pop~="after"]:not([data-lgs-pop~="self"]):not([data-lgs-noslab~="after"]) works).
  • On the acked main cover, set background: none on ::after.
  • File a REQ asking C1a to paint the window sheen and the footer capsule through --lgs-mat-liquid-* / --lgs-mat-window-*, or to key native rules on [data-lgs-pop~="after"].
  • Extend RP-8 to read the computed paint (background-image, box-shadow, --lgs-edge and the host's ::before) of every tagged element.

M2. The popup wrapper does not follow flag (or geometry) changes on popups that are already shown.

• File: device/rt/08-popups.js.
• The gap: entryFor() checks an entry's flag only inside transform(), which runs only when Steam sends a request. The module never subscribes to rt.flags.on / onAny (P1 provides both, runtime.md §3.4) or to rt.bridge geom.
• Live evidence (rvflag.json): with wp.p6 on and a shown test popup matched by an entry {"z": 0.05, "flag": "p6RevFlag"}, pushing p6RevFlag changed nothing in 2 s (z_pixels 60, no z_meters).
• The reverse case: a flag that goes off leaves the changed request in place until Steam happens to re-send. The frame menu, the tab bar of PLAN §1.7 and the wrapper's main client, stays shown and is rarely re-sent.
• Consequences:
  • A lab step with --flags tabBarDepth (or tabBarAlways) tests nothing on the live tab bar.
  • If the step does apply something, the change can outlast the step (team rule 9: no flag effect left behind).
  • A change of r leaves zMm entries stale.
• Why it is latent: no fragment in the tree uses a flagged popup entry yet, and wp.p6 is off.
• Fix: in install, subscribe with rt.flags.onAny and the rt.bridge geom listener, and call resendLive(false) (a re-send is idempotent); add the case to RP-7.

Minor

• m1. The RP-6 harness depends on where focus is at route entry.
  • Sometimes focus is not on a poster when the route opens; the failing run's first poster was unscaled, 172 × 258.
  • The D-pad move then lands on a poster that is still running its focus-scale animation. The reporter skips it (animating (new)) past the 400 ms sample.
  • The mechanism itself is fine (rv6.json). Put focus on the first poster before the move, or poll for up to 1.2 s.
• m2. The cost claim is an artifact of run order (see the table). Say "same cost as Phase 1" in the log, or alternate the injection order between runs.
• m3. Legacy pops fail G-DEPTH on every route, and the R2/R5 failures are P-46/P-47 "must" items.
  • This is expected under PLAN §6.
  • Nothing yet records which area supersedes which legacy id, so 99-legacy.json has no path to empty.
  • Proposed mapping, by PLAN §1.1 ownership: hdr-back, hdr-search and footer to C1a; tabs, tab-arrow, sort-filter and card to C2c/C2a; play and app-button to C5a; menu, sheet, sheet-panel and filters to C1c; button to whoever claims it.
  • Suggest adding this table to reporter.md.
• m4. Rule 7 counts only admission pops (lgs_layers.js around line 1407). On a mixed route (area plus legacy) the reporter stays silent at 6-7 depths. The card asks for a warning "when a route shows more than 4 distinct dz". Count every kept layer, and mark the legacy ones in the warning.
• m5. Duplicate-id handling ignores flags (compile, ids set).
  • An area rule that reuses a legacy id (say card behind wp.c2a) always knocks the legacy rule out as a "duplicate", even while the area flag is off.
  • The flag-gated supersedes path is meant to bring the legacy rule back in that case.
  • Make the duplicate check flag-aware, or document "never reuse a legacy id; use supersedes".
• m6. A rule-level "admission": false silently bypasses every admission rule in any fragment. The contract allows false only at fragment level, for 99-legacy.json. Ignore it outside 99-legacy.json, or report it in errors.
• m7. Rule 2 counts only focusables inside the popped element. PLAN §1.7 says "the smallest visible focusable that intersects the crop". A neighbour that an outset covers is not counted. Low risk; align the code or record the deviation.
• m8. RP-7's frame-menu check reads back what the wrapper itself wrote.
  • Systemui reports only-visible-with-laser: false for the frame menu in every state, so the check proves only the request.
  • The effect is unverified: record it as PLAUSIBLE (needs a wearer) in the log.
  • zMm/space conversion is not exercised live either: RP-7 uses z.
• m9. Runtime-contract hygiene in 08-popups.js:
  • It uses a raw setInterval and a global __LGS_POPUPS_API. runtime.md rules 3-4 prefer rt.setInterval and an API returned from install.
  • It defines __LGS_POPUPS_MODULE at load time when no runtime is present.
  • restored() leaves an empty offset: {} on a request that had none.
• m10. Evidence log precision:
  • Logged times do not match the result files: rpx.json mtime is 06:32 against "06:40 run", and rpint.json is 06:30 against 06:36.
  • Contract §10 says the P7 request is filed in P6.md § Requests, but it is not there (P7 consumes mosaic anyway: lgs_sg.js around line 828).
  • The REQ answers in C1c.md are paragraphs, not one line.
  • RP-5's media case is a <video> fixture, not a media route (disclosed).

The six checks

• (a) Ownership.
  • git diff f5190bb touches only P6 files: lgs_layers.js, 08-popups.js, 05-native.css, layers/00-base.json and contracts/reporter.md.
  • theme/layers.json is unchanged since b8fea9a, as the retired-but-kept rule requires; 99-legacy.json and popups/00-base.json are unchanged this session.
  • The REQ answers in C3a.md, P1.md and C1c.md follow hard rule 1.
  • shots/p2_p6_rp8_*.png follow the naming rule (and shots/ is ignored).
  • OK.
• (b) Safety.
  • Nothing autostarts. Runtime and test state live only in /tmp/lgs/p6*.
  • The tests stay off the never-list: Steam's CONFIRM is opened with no-op handlers and closed without confirming, and the test popup is non-interactive and closed. Locks are respected: lab.Lock, both=True for systemui, native-session.
  • After my runs the device is in CSS-only mode, native off, with no P6 flag or global. A wp.p3 test token that appeared afterwards (fy10893.2) is not mine or P6's.
  • Exception: M2 (a flagged popup change can outlast its flag).
• (c) Robustness.
  • Fail-closed paths are in place: an unresolved token or a bad selector skips the rule or entry and is listed in errors; transform errors fall back to Steam's params; install throws before it patches anything.
  • Removal works by flag-off, remove() and both TTL paths (RP-7 re-run). The reporter detaches its MutationObservers and listeners and clears its attributes.
  • At idle: the reporter polls every 500 ms, but only in native mode, as in Phase 1. The wrapper polls every 1 s, only while wp.p6 is on.
  • Exception: M2.
• (d) Conformance.
  • The machinery is in place: rule 2 (P-46 set and click-safe cap), rule 5 (P-47), interactive false in the default profile (P-11, R8 clean), and occluder plates.
  • Live native routes fail P-46/P-47, all from legacy rules (m3).
  • In native mode the footer keeps a CSS edge next to glassd's (M1), against the "edges only from the material and shader" bar.
• (e) Function retention.
  • P6 changes nothing in CSS-only mode: 05-native.css is scoped to .lgs-native, the wrapper is off, and the reporter only runs in native.
  • In native, the default-profile crops are non-interactive (R8 clean), so the laser reaches Steam's panel.
  • Lifts follow both paths: gamepad .gpfocus (rv6.json, RP-6) and laser .lgs-dwell:hover (RP-6, with no lift during the 45 ms sweep).
• (f) Visual quality.
  • P6 has no mockup of its own. I compared the native shots, composited over a glass tone, with the CSS-only shots of the same routes.
  • Covers, plates and pops drop their CSS glass cleanly, except M1.
  • Outside P6, for the coordinator: on AllGames the search field overlaps the tab row and the "Back" label overlaps its arrow, in CSS-only mode as well. That comes from in-progress area CSS (C1a/C1b), not from P6.

What the builder must do to pass

1. Fix M1: §4 of 05-native.css, plus a REQ to C1a. Re-run RP-8 with checks on the computed paint of every tagged element, and take a native shot of the footer.
2. Fix M2: subscribe to flags and geometry and re-send. Add an RP-7 case where an entry flag toggles while the popup is shown.
3. Correct the evidence log: the cost wording (m2), the times (m10), and RP-6's flakiness (m1).