Atlas stabilization: continuous optics and separate motion roles

13October update: the visual status/evidence index covers the later short-trip timing, substrate tint, clipped wide-input repair, long-tap appearance, frame diagnostics and native.NET integrations. The sections below retain the progression and original measurement limits; the earlier baseline failures are not claims about the current implementation.

This is an authored refinement informed by the owner's original iPhone captures. It does not replace historical measured presets or claim recovery of Apple's private rendering implementation.

New labelled motion recordings

The owner's new tap-only and held/throw recordings are being audited at their native timestamps. The first short tap includes a 329×204px visible selector against a 186px bar. A long tap instead has a much flatter 396×159px cross-section. Authored 40–240ms press alternatives in the unchanged controller do not recover these paired shapes at the observed travel phases. These comparisons use inspected pixel cross-sections and model bounds; they are not a complete contour or timing fit.

The live research log preserves hashes, failures, uncertain measurements and next steps. The comparison packet includes original crops, raw probes and the unchanged production baseline. The coupled tap candidate springs projected area and derives radius from the actual spine. This improves the paired short/long shapes without adding application-side tuning. The rates are authored; a remaining early-frame width mismatch and unknown original input timing are recorded. Green regression tests do not prove full parity.

The consecutive short-tap audit then exposes a temporal error hidden by those still comparisons. The local candidate slows adjacent coupled taps to an authored0.75 rate, smoothly returning to the prior rate at two slots. Shape springs and speed demands share the ratio; actual velocities retain their physical time units. Two partial recorded trips improve to about2.5px centre RMS with a single fitted time offset each. The repeated episode still has a nominal width mismatch up to20.6px. It is not a complete matched trajectory or recovery of pointer timing; follow the packet for full/native verification status.

Release attribution correction

The inherited LOLL8185 release calibration is not reliable. Exact native-time crops at CFR1968/1980 still contain a large refracting lens where the previous analysis reported zero raised area and assumed trailing rest. The314ms area decay and367ms fitted equivalent therefore do not identify an isolated release-to-rest or a finger-up time. The historical spring remains unchanged and is labelled an authored baseline. See the original frames and provenance audit. Earlier research is retained as history; this correction supersedes its release attribution.

Evidence changed the model

The previous Atlas magnifier reused Held, a navigation lens whose material and ink intentionally sample different maps. Its central foreground zoom and separate edge band produced the recurring split-face appearance when used as a free lens. That was a role mismatch, not merely an excessive rim colour.

The original Phone T04 sequence also contradicts transferring the free search-button pull model to navigation: the selector travels horizontally, while its own off-axis squeeze was an inference. The new calm preset assigns anchored strain to the bar and leaves travel/pressure in the selector model.

The stabilization record contains the original T04 frame, T01 sequence, native desktop before/after images, line-grid proof and machine-readable test results. The T01 audit's approximate 225px held height over a 186px bar informs a conservative 1.23-height test bound. It does not identify finger distance or timing. The adopted 1.20 settled-height ratio stays within that reference's tolerance. An initial 1.16 candidate measured 215.76px against 225±4px and was rejected by the original-reference test. The slower springs remain authored choices; reducing drag sensitivity does not justify shrinking the settled hold.

A separate whole-bar audit decodes original PTS 20819/600 and 29863/600 from IMG_6756. Four unoccluded columns and three edge thresholds give a 186px resting bar and 204–205px extreme held bar. Growth is mostly above its resting top edge; the bottom changes only about one pixel. The current centre-anchored response is therefore conservative, not a matched reconstruction. Its portable 72-direction extreme-pull regression peaks at 194.2524px against the measured 205px ceiling. This bound prevents excess deformation; it does not identify a finger gain or close the asymmetry gap. Full-density original crops, PTS, thresholds and the reproducible measurement tool are in the stabilization record. No spring or deformation constant was refitted in this audit. An intervening ordinary-held frame at PTS 21019/600 measures 196px with symmetric growth. The extra extreme deformation is only 8–9px beyond that hold, separating it from press expansion. Generic Calm press is 191.58px at this geometry. A fail-first reference regression therefore rejected that amplitude for the navigation bar. Its preset now reuses the earlier rounded 1.05 scale, applied only to material, giving 195.3px against 196±1px. Generic cards and controls retain 3%/2dp growth. The bar's timing and drag coefficients are unchanged. With this press, both Kotlin and portable geometry remain below the 205px extreme ceiling, peaking at 198.017px in the reference-sized fixture. Ordinary-held amplitude now fits; extreme asymmetry and the additional 6–7px deformation remain unmatched. Passing an upper bound alone cannot establish a correct drag model.

A foreground audit also finds that the visible unselected Contacts icon changes from 69×69px at rest to 73×73px on ordinary hold and 69×76px at extreme (threshold 200). Its bounds centre moves 11.5px upward from ordinary to extreme hold at all three tested thresholds. This is visible-pixel evidence, not identification of layout motion versus optical/content warping. The current implementation keeps layout and ordinary labels fixed; that choice does not itself prove matching visible output. The review includes the original crops, threshold results and a Pillow-only reproduction tool.

A follow-up samples 47 native timestamps across the same T04 gesture. Its bar starts and ends at 186px with the same visible centre, holds at 196px, then reaches 205–206px with the centre 10.5–11px higher during the prolonged drag. All 141 threshold/frame combinations have a strict majority of unoccluded columns for both boundaries. The asymmetry persists within one gesture; it is not an offset between recordings. The earlier 205px gate remains a conservative chosen-frame bound, not the complete recording's maximum. The review supplies the plot, raw votes, original crops and an exact decode/measurement command.

Central-row width probes also reject the earlier 872.98×198.02px drag shape. The clear late T04 crop is 833px wide and 206px high. Navigation now opts into an authored area-preserving log-strain model, giving 835.49×204.70px against rounded 834±4 × 205±1.1px bounds. The height gate retains the earlier selected frame; against the late 833×206px frame, residuals are +2.49px width and -1.30px height. This is a conservative multi-frame fit, not exact same-frame parity. Generic cards keep the size-adaptive law. The original's 10.5–11px visible-centre rise is still unmatched: fitting width and height does not establish correct asymmetric geometry. The 0.047 strain bound is a spatial fit, not identification of Apple's physical or input parameters. Full local and hosted verification at 124ae17 passed: 303 library tests plus 27 skipped, both Atlas tests, 13 portable math/controller tests, 12 Skia tests and five browser tests. The review includes fresh native and software gesture captures, raw redraw timings and the remaining asymmetry.

A further candidate checks top and bottom separately against the selected original frame. The centred model failed at181.15/385.85px; a bounded material-only bias gives170.73/375.43px against 171–172/375px. Bias is capped at5.6% of the short side (10.416px here), below the selected10.5px shift. Layout, hit targets and ordinary icon pixels stay fixed. Other directions use an authored extrapolation; timing and the original foreground warp remain unmatched. Twenty focused Kotlin checks,13 portable math/controller checks and13 Skia checks pass; full verification is pending. The review records failure-before evidence and actual software gesture captures separately from prior completed native/CI runs.

The Calm integration also accidentally added generic press brightness on top of the navigation bar's existing held lift. A rendered fail-first test measured 54/255 against the original 44±1 interior plateau. Keeping the bar's held lift and setting its extra pressLift to zero restores 44/255 without changing motion or generic controls. A fresh untouched patch of the decoded ordinary-held reference is uniformly 43/255, within that tolerance. Touch illumination remains separate; this plateau test does not score the fingertip glow or HDR display appearance.

Display colours and the wider dataset (15 October)

The owner's Phone-bar recordings are HDR10 (PQ, BT.2020). Until 15 October the continuous comparison read them as if they were sRGB, which lifts every dark level: the bar read 45 where a display shows 33. Every comparison now decodes them to display sRGB first (the record's decode, checked against the calibration target's grey ramp within one level), and covers five recordings instead of two. Presses are read from the bar itself, which lifts on every touch, rather than from the selected ink alone, which missed holds on the middle tab and split long drags into separate throws.

In display colours the dark Calm bar now rests at iOS's 33 and holds at its 45, its resting selector reads 10-15, its held lens over a black page matches, and a tap's travelling pill reaches 47-51 against iOS's 48-53. The last needed a behaviour, not a level: iOS lifts the whole bar for any touch, a tap included (GlassTabBarStyle.touchLift). Two differences remain open. iOS's held lens brightens page content under it (an avatar reads brighter through it than without it), which neither tint nor raw page share reproduces. iOS's search button presses, stretches toward the finger and lands squashed against its resting edge; since pass 17 ours does too (GlassPullShape.Free, GlassAccessoryLink), within 2-3 px of iOS's centre over seven presses. The calibration made against the sRGB reading is kept as GlassTabBarStyle.CalmRecordedSpace().

Why the magnifier is continuous

For normalized squared radius r², the source offset is -p*m*(1-r²)², with m=1-1/zoom. At the ellipse boundary, the displacement and its first derivative vanish. The radial source derivative remains positive for the supported 1–2.5× centre zoom: source detail is neither reversed nor skipped in an internal annulus. Both production shader passes include the same mapping function.

Rendered ramp tests cover multiple sizes and zooms. A separate endpoint test compares material and ink at every sampled pixel. These tests prove properties of this implementation; they do not prove pixel identity with an iOS screenshot. The porting guide states the same source/alpha contract for other renderers.

Motion is not one scale animation

Press growth, illumination, drag resistance, selector centre and selector spine have separate states. The calm whole-surface preset uses critical geometry springs, limits press growth and progressively resists drag before integration. Large-surface strain stays subtle. Ordinary labels retain layout coordinates, so dragging a long glass bar does not rotate its text.

The calm selector responds to horizontal travel distance through centre speed and a separate spine spring; perpendicular travel contributes neither trace-free strain nor velocity squeeze. Release retains state continuity and travel-driven spine deformation while pressed material fades. The inherited model zeroed shape-driving velocity on release, making throws rigid; the calm preset now separates these roles. Historical expressive dynamics remain an explicit comparison path.

An earlier correction reapplied the original resting-tap recovery gate to Calm and found a missing phase: a long tap stretched but approached rest monotonically. Its shape damping is now 0.45 (from 0.72), while its maximum target half-spine gain falls from 0.55 to 0.42 slots. The centre and press springs are unchanged. In the 1150×186px five-slot fixture, four-slot travel peaks at 450.63px width (previously 457.64), briefly compresses to 238.08px and returns to 252px. The original 6690–6701 stills support the ordering of stretch, compressed arrival and recovery; they do not establish those numerical gains or elapsed time.

A subsequent audit of original IMG_6698 found that this correction still changed height too late. Six unoccluded column scans give 171–173px height near 49% of the trip; the model gave 162.96px. Optical formation and pressure had two consecutive slow springs. Calm now drives transit pressure directly from the existing movement target through one authored critical spring (45/s). Its spine gain is 0.36 slots because restoring cap height also adds width. The same spatial-phase test now gives 329.74×171.36px, against rounded gates of 330±12 × 172±4px. The height-only candidate failed the width gate at 344.81px; both dimensions must pass together. Optical formation, held/released pressure and historical presets are unchanged. Untimed stills constrain spatial shape, not frequency. At the separate 1150×186px fixture, long-travel peak/rebound are now 429.27/232.87px before settling at 252px. These supersede the earlier model outcomes above, not the reference measurements. The review record includes the unscaled original, pixel probes, reproduction command and failed candidates.

Two original held-travel states then exposed excessive width with insufficient flattening. Measured selectors are 326×174px and 313.5×142px inside a 186px bar. Calm now separates held contraction from ordinary touch growth and uses a concave held-spine response with a smaller maximum gain. Authored speed sweeps reach 326.03×173.93px and 314.03×144.19px; these are shape-envelope fits, not timed iOS trajectories. The earlier one-state fit failed the intermediate-width check. An independent tap allocation also prevents held tuning from clipping tap spring overshoot: 481 tap samples match the preceding curve within 0.001px. Whole-bar squeeze remains separate. The held-travel research includes original crops, probe uncertainty, failed candidates, traces and verification status.

Input targets are integrated on their timestamps in the calm path. Previously, multiple pointer samples between display frames overwrote the preceding target, making the same gesture depend on presentation cadence. The new reversal/throw regression compares 30, 60, 90 and 120Hz against the same 120Hz input stream, within 1.5px geometry and 0.015 formation. This is simulation consistency, not a measurement of display latency. After a stationary hold, the frame loop resumes from the latest input time rather than replaying the idle interval as frozen frames.

An additional end-anchor regression found that raw horizontal finger speed still excited the selector after its centre had exhausted its available travel. Alternating outward and diagonal pulls beyond the same end changed its width by 72.47px and height by 7.30px in a 360×64px fixture, while its centre was stationary. Calm now derives shape-driving speed from the same bounded target used by grasp tracking. Further motion outside a saturated anchor adds no fictitious travel; returning to the tab range resumes the ordinary deformation. Both end anchors report zero width/height change in the corrected test. Whole-bar feedback and raw navigation intent remain separate. This is a causal interaction correction, not a new fitted Apple speed or timing parameter.

The portable Skia adapter executes these same production shaders outside Compose. On the current software ramp fixtures, all four material profiles differ from JVM Skia by at most one channel level out of 255. This is a backend agreement measurement, not an iOS equivalence score or a browser GPU benchmark. Its high-level drawSurface now builds rounded in-app material and the wide tone kernel. Eighteen parameter fixtures generated by the real Kotlin factory guard appearance/tint drift. An independent sinusoidal blur check measures 0.6137 amplitude against the Gaussian prediction 0.6176; coordinate tests include negative local samples. The tone cache recognizes native-canvas aliases, so repeated JavaScript wrappers do not force a blur each frame. These adapter tests do not establish native framework support, identical host compositing or live-backdrop performance.

An edge audit then found transparent-black contamination with small opaque sources: a 1×1 (64,128,192,255) image produced (4,8,11,15) in the wide buffer. Extending the image with clamped shader sampling before blur corrects it. Five source sizes, three sigmas and five locations each now preserve the original constant within one code value; 12 portable Skia tests pass. The same independent sinusoid check measures 0.6119 after correction, within 0.6176 ± 0.04. The earlier surface revision also passed four hosted browser tests; its labelled captures are in the review.

Endpoint precision and native clipping

The legacy endpoint gate remains 1.4256 output levels against its strict one-level target. Its material and ink are first combined in an 8-bit layer, then rounded again after coverage. The new Desktop Calm path keeps those inputs separate until the final float composition and coverage pass. Across the same 18 page/ink/body fixtures, its maximum error is 0.9569 levels (199,788 visible pixels, including 11,124 fractional-coverage pixels). The independent oracle starts from the material and ink outputs and evaluates source-over in double precision. It does not measure optical agreement with Apple or the subsequent display framebuffer's rounding.

CanvasKit's separate-input test also passes: maximum error 0.9668 levels across rounded, asymmetric and pose outlines, including 1,376 fractional-coverage pixels. Native validation caught a failed first integration: sampling the stored ink through a shader offset let the GPU cull it outside the window. Declaring that offset in the image-filter graph restores the selected label. The regression now exercises the actual graph with viewport clipping, and storage rows must stay transparent. Android retains the historical compositor because its current Java API does not expose the required multiple dynamic inputs. This precision improvement is enabled only for Desktop Calm navigation.

See the host compositor contract and the native before/candidate/corrected captures in the stabilization record. A direct source shader and an image filter have different input-bounds requirements; compilation alone cannot establish equivalence.

Public sources and limits

Apple describes interactive glass, illumination and material roles in Meet Liquid Glass and distinguishes Clear/Regular usage in its materials guidance. Those sources provide design guidance, not the numerical spring constants used here. Pixel comparisons must also account for source content, density, colour handling and time alignment.

The native desktop evidence uses Direct3D, not an assumed Vulkan pipeline. Its static redraw timings are not presented-frame or Android performance. No new physical device capture was available. Full 1:1 parity and native ports for every framework remain open work. The strict endpoint gate now passes for separate-input desktop/CanvasKit composition; the historical and Android path remains above it.