docs(tier): correct large-tier galaxy count to 2.5M

CLAUDE.md

@@ -148,14 +148,14 @@ ReadMes for the upstream catalogs live in `data/raw/` (`J_ApJS_199_26_ReadMe`, `
 - **`textureAtlas.ts` + `quadRenderer.ts` + `shaders/quads.wgsl`**: 2048×2048 atlas of 128×128 slots for galaxy thumbnails. LRU eviction.
 - **`galaxyImageQueue.ts`**: priority queue + concurrency limiter (max 4) for thumbnail fetches. Idempotent enqueue (don't re-add in-flight keys — see the long comment for the bug history).
 - **`galaxyImageFetcher.ts`**: SDSS DR18 ImgCutout (CORS-safe) for SDSS galaxies; CDS hips2fits (CORS-safe DSS proxy) for 2MRS/GLADE.
-- **`engine.ts`**: per-frame loop. Per-galaxy `apparentSizePx` gates thumbnail enqueue — but the inner loop hoists `Math.tan` and pre-computes `maxCamDistForVisibility` to avoid 3.5M trig calls per frame.
+- **`engine.ts`**: per-frame loop. Per-galaxy `apparentSizePx` gates thumbnail enqueue — but the inner loop hoists `Math.tan` and pre-computes `maxCamDistForVisibility` to avoid 2.5M trig calls per frame.
 - **`renderScheduler.ts` + `engine.ts` frame tail**: render-on-demand. `requestRender()` from event handlers wakes the loop; the frame body re-schedules only while `autoRotate || currentTween || hasAnyAxis || queue.inFlightCount > 0 || recent-fade` is true.
 
 ## When the user asks you to…
 
 - **"add a feature"** → look in `docs/superpowers/plans/` for an existing plan. If it's substantial, write a new plan via the `writing-plans` skill rather than coding inline.
 - **"fix this bug"** → check tests first; the project favours reproducing bugs as failing tests, then fixing.
-- **"why is this slow?"** → profile mental model first: per-frame work scales with on-screen galaxies (~3.5M total). Inner-loop trig and `Math.sqrt` are real costs. Hoist constants, gate with squared distances, avoid per-galaxy `Math.tan`.
+- **"why is this slow?"** → profile mental model first: per-frame work scales with on-screen galaxies (~2.5M total). Inner-loop trig and `Math.sqrt` are real costs. Hoist constants, gate with squared distances, avoid per-galaxy `Math.tan`.
 - **"refactor X"** → keep the services/ layout. Cross-cutting helpers go in `utils/`; rendering subsystems in `services/gpu/`. Tests mirror the src tree.
 - **"why does the renderer use index Y?"** → check `pointRenderer.ts` SLOTS_PER_POINT and the matching attribute layout in the shader. They must agree byte-for-byte.
 

README.md

@@ -24,7 +24,7 @@ _The full overlay surface — left stack (Navigation cheatsheet, Settings panel,
 
 ![Data-tier selector](docs/screenshots/tier-selector.png)
 
-_Segmented control at the top of the Settings panel hot-swaps the loaded dataset between three sizes: **Small** (~300k galaxies — mobile-friendly), **Medium** (~600k — default for laptops), and **Large** (~3.5M — full catalog). Tier choice on first paint is driven by viewport width; clicking a button re-fetches the relevant `.bin` files and re-uploads the GPU vertex buffers in place — no page reload._
+_Segmented control at the top of the Settings panel hot-swaps the loaded dataset between three sizes: **Small** (~300k galaxies — mobile-friendly), **Medium** (~600k — default for laptops), and **Large** (~2.5M — full catalog). Tier choice on first paint is driven by viewport width; clicking a button re-fetches the relevant `.bin` files and re-uploads the GPU vertex buffers in place — no page reload._
 
 ![Pinned InfoCard for a 2MRS galaxy](docs/screenshots/infocard-detail.png)
 

src/@types/data/Tier.d.ts

@@ -6,7 +6,7 @@
  *              to its brightest ~256k, 2MRS + Famous kept whole (small).
  * - `medium` — default for desktops, ~600k total. Brightest ~156k SDSS +
  *              brightest ~400k GLADE + full 2MRS + full Famous.
- * - `large`  — opt-in full catalog (~3.5M). The pre-tier behaviour.
+ * - `large`  — opt-in full catalog (~2.5M). The pre-tier behaviour.
  *
  * The values are persisted in URL query strings and the runtime API only
  * (never on disk: the binary format is tier-agnostic). String-union — not a

src/data/galaxyCatalogTransfer.ts

@@ -5,7 +5,7 @@
  * ### Why this module exists
  *
  * Sending a GalaxyCatalog across a Worker boundary with structured-clone
- * cost would be prohibitive at ~3.5M galaxies. The cheap alternative is
+ * cost would be prohibitive at ~2.5M galaxies. The cheap alternative is
  * `postMessage(payload, transfer)` with a list of `ArrayBuffer`s to
  * transfer ownership of — but we can't transfer the engine's
  * authoritative buffers (the picker and InfoCard read them after the

src/utils/cluster/clusterMembership.ts

@@ -8,7 +8,7 @@
  * ### Why a pure function (no caching here)?
  *
  * The expensive bit is the cone search itself (one vec3 subtract + one
- * dot product per galaxy, ~3.5M ops for the full loaded catalog).
+ * dot product per galaxy, ~2.5M ops for the full loaded catalog).
  * Memoising against `(poiId, dataRev)` belongs to the subsystem that
  * owns the focus state — see spec §4.3 — because cache invalidation
  * needs to know when a tier swap has bumped `dataRev`, which this pure
@@ -66,7 +66,7 @@ export type ClusterMembershipResult = {
  * `radiusMpc` of `centerMpc` across the supplied catalogs.
  *
  * Time complexity: O(total galaxy count). For the typical loaded
- * footprint (~3.5M galaxies across SDSS + 2MRS + GLADE), one call
+ * footprint (~2.5M galaxies across SDSS + 2MRS + GLADE), one call
  * runs in single-digit milliseconds on the target hardware — see
  * spec §4.2 for the rationale on runtime-vs-build-time computation.
  *
@@ -83,7 +83,7 @@ export function clusterMembership(
   const cx = centerMpc[0];
   const cy = centerMpc[1];
   const cz = centerMpc[2];
-  // Compare against squared distance to avoid 3.5M Math.sqrt calls.
+  // Compare against squared distance to avoid 2.5M Math.sqrt calls.
   const r2 = radiusMpc * radiusMpc;
 
   const packedIds: number[] = [];

src/utils/initialTierFromViewport.ts

@@ -11,7 +11,7 @@
  * enough that users intuitively expect "tablet and up" to behave like a
  * desktop.
  *
- * 'large' is intentionally never auto-selected: the full 3.5M-point catalog
+ * 'large' is intentionally never auto-selected: the full 2.5M-point catalog
  * stresses integrated GPUs and the user should opt-in.
  *
  * ### Defensive edge cases

tools/site/makeOgImage.ts

@@ -55,7 +55,7 @@ const overlay = Buffer.from(
           font-size="34" font-weight="400" fill="#a8d0ff" opacity="0.95">Interactive WebGPU 3D galaxy explorer</text>
     <text x="60" y="${H - 12}"
           font-family="-apple-system, BlinkMacSystemFont, 'SF Pro Display', 'Segoe UI', Helvetica, sans-serif"
-          font-size="22" font-weight="400" fill="#ffffff" opacity="0.65">SDSS · GLADE · 2MRS · ~3.5M galaxies · skymap.rulkens.com</text>
+          font-size="22" font-weight="400" fill="#ffffff" opacity="0.65">SDSS · GLADE · 2MRS · ~2.5M galaxies · skymap.rulkens.com</text>
   </svg>`,
 );