Rewrite NORTHSTAR.md for an outside reader (v0.14.3)

The manifesto was written in the shorthand of the conversations that produced
it. Dropped the internal "Target / 110%" pillar table (→ four plain commitment
statements), the "purity is an instrument, not an identity" in-joke, the
struck-through "no parsing, ever" amendment (changelog history, not manifesto),
and the "SOTA" bravado. Principles and scope rule kept — they already read for
an outsider. No code changes.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: somejohnbforya

CHANGELOG.md

@@ -4,6 +4,14 @@ All notable changes. The project follows semver; until 1.0, minor versions
 may adjust APIs (taken so far, both pre-npm: `dischargeSPD` → `emissionSPD`
 in 0.12.0, `sodiumSPD` → `lowPressureSodiumSPD` in 0.13.0 — see below).
 
+## 0.14.3 — 2026-06-13
+
+Docs only. Rewrote NORTHSTAR.md for an outside reader: dropped the internal
+"Target / 110%" pillar framing, the "purity is an instrument" in-joke, and the
+struck-through "no parsing, ever" amendment (that history lives here in the
+changelog, not the manifesto). The design principles and scope rule are
+unchanged in substance — only the insider shorthand is gone. No code changes.
+
 ## 0.14.2 — 2026-06-13
 
 Docs correctness. Audited the "Honest comparison" table against the *current*

NORTHSTAR.md

@@ -3,50 +3,57 @@
 > **A research-grade color laboratory for the web.**
 > **Every color space, every illuminant — digit-identical in JS, GLSL, and WGSL.**
 
-Every conversion is generated from one table of cited constants. The matrix that
-runs in your JavaScript is the matrix that runs in your shader — and CI verifies
-it to the last digit. Stop hand-porting color matrices.
-
-## Pillars
-
-| Pillar | Target | 110% |
-|---|---|---|
-| **Coverage** | All 14 CSS Color 4 spaces + OKHSL/OKHSV; full CIE illuminant table + arbitrary white points; gamut checking *and* gamut mapping | Illuminants from correlated color temperature (`illuminantFromCCT`); CAT02/CAT16/von Kries alongside Bradford |
-| **Verified precision** | Every conversion tested against derivational invariants and spec fixtures; round-trip bounds CI-enforced; every constant carries a citation | Differential testing against culori, colorjs.io, and @texel/color as oracles; precision-bounds table in the README, generated by CI |
-| **Everywhere** | Core imports clean in Node / SSR / workers; GLSL + WGSL + JS emitted from a single descriptor pipeline over single-sourced constants | GPU parity tests in CI; shader-side gamut mapping |
-| **Speed** | Zero-allocation out-param paths; precomposed routes; honest published benchmarks | SOTA: nothing maintained is faster, on conversion or gamut mapping — measured with @texel/color's own suite |
+What this library is for, and the rules it holds itself to. Every conversion is
+generated from one table of cited constants — the matrix that runs in your
+JavaScript is the matrix that runs in your shader, verified in CI to the last
+digit. That one idea drives everything below.
+
+## What it commits to
+
+- **Coverage.** Every CSS Color 4 space plus OKHSL/OKHSV; the CIE illuminants and
+  arbitrary white points; chromatic adaptation across multiple transforms
+  (Bradford, CAT02, CAT16, von Kries); illuminants from correlated color
+  temperature; gamut *checking* and gamut *mapping*; and a spectral layer (color
+  matching functions, Planck's law, daylight, emission).
+- **Verified precision.** Conversions are tested against derivational invariants,
+  spec fixtures, and the other major libraries as oracles (culori, colorjs.io,
+  @texel/color). Round-trip error bounds are CI-enforced and published — measured,
+  never implied.
+- **Portability.** The core imports cleanly in Node, SSR, and workers, and one
+  pipeline emits JS, GLSL, and WGSL from the same constants — with CPU/GPU parity
+  checked in CI, so a color is identical on both.
+- **Speed.** Zero-allocation paths and precomposed routes, with benchmarks that
+  are public and reproducible (`npm run bench`).
 
 ## Principles
 
 1. **Derive, don't transcribe.** RGB↔XYZ matrices are computed at module load
-   from cited primaries and white points. A hardcoded matrix is a transcription
-   error waiting to happen; a derived one is checked against published values in CI.
-2. **One source of truth.** Constants live in one table. JS functions, GLSL
-   strings, and WGSL strings are all views of that table.
-3. **Precision is a contract.** No rounding inside conversions. Bytes exist only
+   from cited primaries and white points, then checked against published values
+   in CI. A hardcoded matrix is a transcription error waiting to happen.
+2. **One source of truth.** Constants live in one table; the JS functions, GLSL
+   strings, and WGSL strings are all views of it.
+3. **Precision is a contract.** No rounding inside conversions — bytes exist only
    at the explicit `toBytes`/`fromBytes` boundary. Error bounds are measured,
-   published, and enforced — never implied.
-4. **The math is honest about itself.** Hue is undefined at zero chroma; CSS
-   white points differ from CIE 15 values in the fifth digit; gamut mapping is
-   lossy. We document these instead of papering over them.
-
-## Scope rule
-
-**If a frozen spec defines it on coordinates, implement it exactly; strings
-on the way out only; constants still in motion wait.**
-
-## Anti-goals
-
-- No color wrapper class — plain arrays in, plain arrays out
-- ~~No CSS string parsing, ever~~ — **amended at v0.8**: purity is an
-  instrument, not an identity. Strings are how colors arrive in the world,
-  so the frozen Level 4 grammar is in scope — behind a fenced boundary
-  (src/parse): the conversion core never sees a string, `none` becomes NaN
-  under the numerical policy, and still-moving grammar (Level 5 relative
-  color syntax) waits like every other unfrozen spec
-- Alpha is not a color coordinate — the conversion core stays 3-channel;
-  spec-defined alpha interactions (premultiplied mixing, serialization) are
-  separate, consistently 4-channel entry points
-- No palettes, harmonies, color naming, or gradient objects — those belong
-  to applications (spec-defined interpolation itself is in scope)
-- No unfrozen perceptual models (APCA waits for WCAG 3 to freeze)
+   published, and enforced.
+4. **The math is honest about itself.** Hue is undefined at zero chroma; the CSS
+   white points differ from the CIE 15 values in the fifth digit; gamut mapping
+   is lossy. The library documents these rather than papering over them.
+
+## Scope
+
+One rule decides what belongs here: **if a frozen spec defines it on
+coordinates, implement it exactly; parse strings only at a fenced boundary;
+anything still in motion waits.**
+
+Out of scope, on purpose:
+
+- **No color wrapper class** — plain arrays in, plain arrays out.
+- **Strings only at the edge.** Parsing covers the frozen CSS Color 4 grammar
+  behind a fenced module the conversion core never sees; CSS Level 5
+  relative-color syntax and calc() wait until the spec freezes.
+- **Alpha is not a color coordinate** — the conversion core stays 3-channel;
+  alpha-aware operations (premultiplied mixing, serialization) are separate,
+  4-channel entry points.
+- **No palettes, harmonies, color naming, or gradient objects** — those belong
+  to applications. (Spec-defined interpolation itself is in scope.)
+- **No unfrozen perceptual models** — APCA returns when WCAG 3 freezes.

package.json

@@ -1,6 +1,6 @@
 {
   "name": "whitepoint",
-  "version": "0.14.2",
+  "version": "0.14.3",
   "description": "Research-grade color math: every CSS color space, arbitrary illuminants, CATs, and CCT — digit-identical in JS, GLSL, and WGSL.",
   "type": "module",
   "main": "./src/index.js",