Owner, 2026-08-04, on yesterday's attempt: «с лучами солнца ты сделал фигню —
не надо размывать их боковые грани».
They are right. 22b588e answered "the shafts run into something invisible" with
a Gaussian blur of the WHOLE wedge (`raySoftness`, filter `hp-sunsoft`), which
feathered the sides as well as the tip. A shaft of light through a window has
crisp sides; only its reach fades. The blur turned every wedge into a smudge.
GONE. `raySoftness()`, the `<filter>`/`feGaussianBlur` in <defs>, the `<g
filter clip-path>` wrapper, and with it the `hp-sunclip` clipPath — that clip
existed only so the blur could not bleed through a wall. The polygons come out
of `computeSunRays()` already intersected with the room, so a wall still stops
the light by geometry (demo/smoke_sun.mjs, wedgeClippedToRoom). The sun layer
is plain `<polygon fill="url(#hp-sun-i)">` again.
THE KERB DID NOT COME BACK, and not by luck. The old bright edge floating in
mid-floor was never about softness: the gradient's iso-alpha lines are square
to the SUN, while a parallelogram's far edge is parallel to the WALL. Head-on
they coincide; at any other angle one far corner sits at offset `1 − 0.5/k` —
0.71 of the way at a low sun, 0.11 at a high one — i.e. still lit when the
polygon ends. So `rayQuad()` no longer builds a parallelogram: each side is
extruded until it reaches the same distance `len` ALONG `dir`, which puts the
far edge on one iso-alpha line of the gradient. Combined with the untouched
`RAY_FADE_END` = 85 %, the last 15 % of every wedge is empty and its outline
has nothing left to draw. The sides stay razor-sharp on purpose.
Length (×0.7) and the live sky catch-up are untouched.
Tests: unit — `rayQuad` at six sun angles (sides exactly parallel to the ray,
both far corners at offset 1, far edge ⊥ ray, nothing past the gradient) plus
the head-on parallelogram pinned; the `raySoftness` test is gone with the
function. Smoke — demo/smoke_sun_soft.mjs keeps the reach and the "dead at
85 %" checks and flips the feather assert into its opposite: no filter on any
wedge, no `feGaussianBlur` in the tree, and at an OBLIQUE sun (230°/8° and
225°/55°) no vertex is drawn past the end of the gradient. Verified to fail on
the previous bundle on exactly those four. All 247 unit tests and all 97 smokes
green. Stills: sun_sharp_low / sun_sharp_high (demo/shot_sun_short.mjs now
takes a file prefix).
Owner, 2026-08-04: «лучи от солнца сделать короче на 30%, проверить, чтобы они
всегда плавно рассеивались (сейчас есть ощущение, что они упираются во что-то
невидимое)».
SHORTER. `rayLength` is now the v1.56 curve times RAY_LENGTH_K = 0.7 — 1.75
window lengths at sunrise, 0.56 at the zenith. Scaling the whole curve instead
of re-picking the constants keeps the shape the owner approved: a low sun still
reaches three times further than a high one.
WHAT THEY WERE BUMPING INTO. Nothing invisible — the wedge's own outline, in
three places at once.
1. The gradient runs ALONG the sun, so its iso-alpha lines are perpendicular to
the sun, while the wedge's far edge is parallel to the WALL. The two
coincide only for a sun hitting the glass dead-on; at any other angle one
half of that far edge was cut while it still carried colour — a straight
bright kerb hanging in the middle of the floor. The single `100% → alpha 0`
stop hid this from the reader of the code and from nobody else.
2. The two SIDES of the wedge had no falloff at all: two razor lines from the
window into the room, brightest exactly where they are most visible.
3. Where the room outline clips the wedge — the opposite wall, the inner corner
of an L, and above all an OPEN (virtual) boundary, which has no wall drawn
at all — the shaft was chopped at whatever alpha it still had.
WHAT IT IS NOW. The gradient still spans the FULL wedge (geometry and gradient
must describe the same shaft), but `rayStops()` eases it to a hard zero at
RAY_FADE_END = 85% of the length, so the last 15% of every wedge is guaranteed
empty and a shaft ending in mid-air has nothing left to draw an edge with. Each
wedge is then drawn inside `<g filter clip-path>`: SVG applies the filter FIRST
and the clip SECOND, so a Gaussian blur of `raySoftness(len)` (7% of the shaft,
clamped 3…18 render units) feathers the sides and the tip and the room outline
cuts that feather off. Light still never crosses a wall — but where it reaches
one, the kerb is a soft ramp that reads as light landing ON the wall.
Clipping by the room is untouched; only its visible edge changed.
Tests: unit — rayLength pinned at exactly 70% of the old curve at ten
elevations, rayStops (monotone, dead at/after 85%, bright at the glass),
raySoftness clamps. Smoke — demo/smoke_sun_soft.mjs, which fails on the
previous tip (lowSunIs70Percent, highSunIs70Percent, gradientSpansWholeWedge,
deadWellBeforeTheEnd, everyWedgeFeathered). Stills: demo/shot_sun_short.mjs.
Matysh