feat: add real-sky-poster skill

.claude/skills/real-sky-poster/SKILL.md

@@ -0,0 +1,125 @@
+---
+name: real-sky-poster
+description: Replace invented constellations in artwork with the real sky for a given date and place, then export a full sized asset kit. Auto-triggers on "the constellations are wrong/fake/bullshit", "make the stars accurate", "real constellations", "what did the sky actually look like on <date>", "fix the sky in this poster", "build an asset kit from this artwork".
+---
+
+# Real Sky Poster
+
+Take artwork with a made-up night sky, give it the sky that was actually there, and ship every size the campaign needs.
+
+## Success Criteria
+
+Done when all of these hold:
+
+- [ ] **The sky is real.** Every star is a catalogue star at its true magnitude; every figure has its true shape and its true rotation against the horizon for the stated date, time and location.
+- [ ] **The astronomy gate passes.** `python3 tests/skills/real-sky-poster/test_sky.py` is green. Polaris altitude equals the observer's latitude, or the reduction is wrong.
+- [ ] **Nothing below the horizon is drawn.** If a constellation had not risen, it is not in the picture. Say so rather than faking it.
+- [ ] **The artwork is untouched.** `plate.disturbance()` returns a number in the low hundreds. Thousands means the wipe ate somebody's painting — stop and fix the geometry.
+- [ ] **Stars are native at every size.** No output contains an upscaled star.
+- [ ] **No crop discards constellations.** Tall formats extend the sky; they do not centre-crop.
+- [ ] **Copy is verbatim.** Every word traces to a source file. Nothing drafted, nothing paraphrased.
+- [ ] Every exported file hits its exact pixel dimensions.
+
+## Why not just regenerate the image
+
+Because it will not work. An image model asked for "accurate constellations" produces a *fresh set of hallucinated star patterns* — the same failure in a new arrangement. Star positions are exact data. Port the data; do not ask a model to imagine it.
+
+This is the same lesson as optical illusions: when the output must be *correct*, compute it, don't generate it.
+
+## Instructions
+
+### Inputs
+
+A **profile** (`profiles/<name>.json`) carrying:
+
+| Field | What it is |
+|---|---|
+| `source` | path to the artwork (never committed — see [Boundary](#boundary)) |
+| `observer` | `utc` (**UTC, not local** — the easiest way to draw the wrong sky), `lat`, `lon` |
+| `protect` | boxes the wipe must never touch: type, figures, anything painted |
+| `horizon` | the top edge of the painted mass; above it is open sky |
+| `layout` | where each figure sits, and at what scale |
+| `zenith` | figures for the extended sky in tall formats |
+| `copy` | approved strings + palette |
+
+`profiles/futureproof-salmon-starfield.json` is the worked example.
+
+### Stages
+
+Run in order. Each stage is independently inspectable — look at the output before moving on.
+
+```
+1. plate.py    wipe the invented sky, keep the aurora and the brushwork
+               -> check disturbance(); should be low hundreds
+2. upscale.py  4x the CLEAN PLATE with a local ESRGAN (free, on-device)
+3. render.py   draw the real stars at target scale — natively, never upscaled
+4. kit.py      extend the sky for tall formats, set type, export every size
+```
+
+**Stage order is the whole trick.** Upscale the *plate*, then redraw the stars. Upscaling a finished poster gives you 4x-blurry stars; this gives you 4x sharp ones. It also means the upscaler never sees a fabricated dot it might sharpen into an artifact.
+
+### Placing figures
+
+Shape and rotation are computed. **Placement is composed** — and you should say so out loud rather than overclaim. Honest framing: *"every figure has its true shape and orientation; where each sits on the canvas is arranged to fit the art, following the real bearings."*
+
+Read `Sky.report()` first, then place west-bearing figures left, north low and central, east right. That arrangement is usually a gift: it gave the Futureproof poster "summer setting on one side, winter rising on the other."
+
+### Growing the canvas, not cropping it
+
+A 16:9 painting cropped to 1:1 loses its left and right edges — exactly where the constellations live. So tall formats **grow upward**. Upward is toward the zenith, which is where the figures that could not fit the wide frame actually were. The square gains real sky instead of losing it.
+
+## Quality Check
+
+- [ ] Did I check `Sky.report()` before placing anything, or did I assume?
+- [ ] Is anything drawn that was below the horizon?
+- [ ] Did `disturbance()` stay low, and did I *look* at a 100% crop of the artwork's fine detail?
+- [ ] Did I A/B the upscale against plain Lanczos before trusting it? (Photo-trained models plasticize brushwork — use an illustration model, never `RealESRGAN_x4plus`.)
+- [ ] Does any asset stack a second wordmark on artwork that already has one?
+- [ ] Is every string traceable to a source file?
+
+## HITL Checkpoint
+
+**Type:** Two checkpoints, because the two halves fail differently.
+
+**Checkpoint 1 — after the sky (stage 3).** Show the corrected artwork and an annotated key naming each constellation with its real bearing and altitude. The key is what lets a human *check the claim* instead of taking it on faith.
+
+Human reviews:
+- [ ] Does the sky look right, and does the key survive a fact-check?
+- [ ] Is the artwork genuinely untouched at 100%?
+
+**Checkpoint 2 — before publish.** The kit is a marketing artifact.
+
+Human reviews:
+- [ ] Is the copy the approved copy?
+- [ ] Do the story safe areas hold?
+- [ ] Any brand conflict surfaced rather than silently resolved?
+
+**Surface conflicts, never resolve them silently.** The Futureproof run turned up two (an OG size that disagreed between code and spec; two competing wordmarks). Both went in the PR body for a human to rule on.
+
+## Boundary
+
+Rafiki's public repo is **tool-only**. Artwork, generated assets and campaign copy are gitignored and live in the private knowledge base. This skill ships **code and geometry**; profiles reference artwork by path and never embed it. The smoke test runs with no image, no GPU and no network — which is also why the astronomy is standard-library only.
+
+## Setup
+
+Two assets are fetched, not committed (both are binaries, and the repo is tool-only):
+
+```bash
+# Upscale weights — an ILLUSTRATION model. Never RealESRGAN_x4plus on painted art.
+curl -L -o ~/Code/ComfyUI/models/upscale_models/4x-UltraSharp.pth \
+  https://huggingface.co/Kim2091/UltraSharp/resolve/main/4x-UltraSharp.pth
+
+# Type faces for stage 4 (swap for whatever the brand actually uses)
+mkdir -p .claude/skills/real-sky-poster/scripts/fonts
+curl -L -o .claude/skills/real-sky-poster/scripts/fonts/Inter.ttf \
+  'https://raw.githubusercontent.com/google/fonts/main/ofl/inter/Inter%5Bopsz,wght%5D.ttf'
+curl -L -o .claude/skills/real-sky-poster/scripts/fonts/Cormorant.ttf \
+  'https://raw.githubusercontent.com/google/fonts/main/ofl/cormorantgaramond/CormorantGaramond%5Bwght%5D.ttf'
+```
+
+Stages 1–3 need only `opencv-python`, `pillow`, `numpy`. Stage 2 runs on ComfyUI's venv (torch + MPS + spandrel already there). **The smoke test needs none of this** — that is the point of keeping the astronomy in the standard library.
+
+## Reference
+
+- Worked example: the Futureproof salmon-starfield poster — [WalksWithASwagger/futureproof-festival#644](https://github.com/WalksWithASwagger/futureproof-festival/pull/644)
+- Star data: Bright Star Catalogue, J2000.

.claude/skills/real-sky-poster/profiles/futureproof-salmon-starfield.json

@@ -0,0 +1,72 @@
+{
+  "$comment": "Geometry only. The artwork itself is never committed to this repo (see .gitignore: assets/, outputs/). Point `source` at a local file.",
+
+  "name": "futureproof-salmon-starfield",
+  "source": "~/Code/futureproof-festival/public/media/launch/futureproof-salmon-starfield.jpg",
+  "width": 1376,
+  "height": 768,
+
+  "observer": {
+    "$comment": "UTC, not local. 21:00 PDT on Wed 28 Oct 2026 is 04:00 UTC on the 29th.",
+    "utc": [2026, 10, 29, 4, 0, 0],
+    "lat": 49.2827,
+    "lon": -123.1207,
+    "place": "Vancouver, BC"
+  },
+
+  "protect": {
+    "$comment": "Painted elements the wipe must never touch. Generous on purpose: over-protecting costs a few square pixels of star field, under-protecting destroys the art.",
+    "title": [412, 0, 962, 188],
+    "figures": [
+      [968, 52, 1112, 222],
+      [676, 192, 816, 306],
+      [830, 156, 1000, 280],
+      [1036, 172, 1220, 334]
+    ]
+  },
+
+  "horizon": {
+    "$comment": "Top edge of the painted mass, left to right. Everything above is open sky. Where paint reaches high (x 350-560: kelp, eyes, a tail fin) the line must stay above it -- the wipe is destructive.",
+    "points": [
+      [0, 384], [100, 386], [170, 372], [240, 350], [300, 322], [350, 306],
+      [400, 302], [460, 300], [520, 294], [560, 288], [620, 250], [700, 264],
+      [780, 244], [860, 234], [920, 218], [990, 228], [1060, 266], [1120, 294],
+      [1180, 322], [1240, 340], [1300, 346], [1376, 348]
+    ],
+    "residual_zones": [[0, 340], [1120, 1376]]
+  },
+
+  "layout": {
+    "$comment": "Where each figure sits: [centre x, centre y, pixels per degree]. Shape and rotation are computed from the real sky; only placement is composed, and it follows the true bearings -- west left, north low and central, east right.",
+    "Cygnus":     [116, 96, 6.6],
+    "Lyra":       [306, 182, 7.4],
+    "Ursa Major": [176, 292, 7.8],
+    "Cassiopeia": [1232, 84, 8.2],
+    "Perseus":    [1298, 236, 5.4],
+    "Pleiades":   [1318, 308, 24.0]
+  },
+
+  "zenith": {
+    "$comment": "For tall formats, the canvas grows upward -- toward the zenith. These are the figures that actually live up there and cannot fit the 16:9 crop. [centre x, height as fraction of the extension, pixels per degree].",
+    "Ursa Minor": [430, 0.19, 11.0],
+    "Andromeda":  [1140, 0.34, 8.4],
+    "Pegasus":    [1180, 0.72, 11.5],
+    "Cepheus":    [742, 0.82, 10.0],
+    "Draco":      [300, 0.56, 5.2]
+  },
+
+  "copy": {
+    "$comment": "Verbatim from the festival repo. Never drafted, never paraphrased. The painting already carries the wordmark, the date and the presenter -- the type layer adds only what the painting does not say.",
+    "line": "The most honest AI conversation happening this year.",
+    "line_source": "lib/brand-guide.ts:37",
+    "date": "October 28-30, 2026",
+    "venue": "H.R. MacMillan Space Centre  ·  Vancouver",
+    "url": "futureproof.website",
+    "palette": {
+      "ink": [10, 10, 11],
+      "parchment": [245, 241, 232],
+      "aurora": [94, 234, 212],
+      "ember": [252, 165, 165]
+    }
+  }
+}