Gravitational lensing: SIS + NFW thin-lens with precomputed NFW image-finding LUT

[rulkens]

What this is

Vertex-stage gravitational lensing for the galaxy point cloud: each background
galaxy is deflected by foreground cluster lenses, rendered as a primary image
plus a counter image (12 verts/point when lensing is on). Two lens profiles —
SIS (analytic) and NFW (driven by a precomputed image-finding LUT).

The branch is large because it stacks several prior lensing commits; the
NFW image-finding LUT is the work from this session and the part most worth
reviewing closely.

Where to focus the review (this session's LUT work)

These commits, newest first:

• e6d1194a fix(lensing): drive NFW primary image from the LUT — the
  consistency fix. Before this, the NFW primary used the source-plane summed
  deflection while the counter used the exact LUT inverse, so the two images
  were solved by different methods and visibly crossed (inner/outer rings
  overlapping, unlike SIS). Now the LUT sample is hoisted above the image-kind
  branch so the dominant NFW lens's primary and counter read one inverse
  solution. The primary's source-plane push is swapped for the LUT-exact image
  position only inside the LUT's y-domain (y <= Y_MAX), continuous at the
  boundary (xPrimary -> y as y -> Y_MAX). SIS + NFW-counter math unchanged.
• 9622ece4 drive the NFW counter image from the LUT — option (b) texture
  params on lensedPosition, the (y, s) -> UV inverse map, the 12-vert gate
  for both modes, and the TS↔WESL constants parity test.
• 88e4d21b wire the NFW LUT into the @group(3) scene group — BGL bindings
  2/3 (rgba16float texture + filtering sampler, both VERTEX), bind-group
  entries, LUT built at startup, lifecycle/teardown.
• 41b6317e createNfwLensLutTexture GPU resource — rgba16float + linear
  sampler (dodges the non-universal float32-filterable feature), f16 pack.
• bd8f7bfe extract lensedPosition into lib/lensing.wesl — faithful 1:1
  lift, pre-wires the shared deflection model.
• 4372bdb0 NFW image-finding LUT generator (CPU) — buildNfwLensLut:
  dense-scan + bisection root finder, log-scaled s-axis, channels
  [xPrimary, muPrimary, xCounter, muCounter].

Architecture

• The LUT inverts y = x − s·nfwShape(x) on the CPU into a 256×64 rgba16float
  texture, sampled in the points vertex stage — no per-vertex root-finding.
• s-axis is log-scaled (LOG_K = 4.0); the shader UV map is the algebraic
  inverse of the generator's forward map. TS↔WESL constant parity is test-guarded
  (nfwLutConstants.parity.test.ts) since ?static WESL linking injects no values.
• LUT texture + sampler co-host in @group(3) (scene group) alongside focus and
  the lensing uniform buffer — the only group secondary pick renderers don't declare.

Verification

• npm run build — green (WESL links; the iOS-fatal-shader class is the reason
  the build gate matters here).
• npm test — 3070 passing.
• npm run typecheck — clean.

Outstanding before merge

• Visual confirmation of NFW lensing is the one manual DoD gate not yet
  satisfied — I have not eyeballed the NFW rings rendering. Worth a look on the
  dev server before this lands.
• Plan: docs/superpowers/plans/2026-06-24-nfw-lensing-lut.md (+ -part2.md).
  Disk/quad lensing coverage was deferred (galaxies large enough to be disks are
  near-camera and rarely lensed); the lensedPosition extraction pre-wires it.

🤖 Generated with Claude Code

[cloudflare-workers-and-pages]

Deploying with    Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status	Name	Latest Commit	Preview URL	Updated (UTC)
✅ Deployment successful! View logs	skymap	6a90a46	Commit Preview URL Branch Preview URL	Jun 25 2026, 12:17 AM