fix(cliffs): sample the fields at the game's lattice - Nauvis is now EXACT (#18) (#70)

The port read `cliff_elevation` and `cliffiness` at `(i*4, j*4 + 0.5)`,
adding the cliff prototype's `grid_offset {0, 0.5}` to the SAMPLE
position. That offset is a CENTRE offset. The game samples the bare
`(i*4, j*4)`.

Two independent sources:

- `CliffGenerator::crossingsForChunk` (arm64 `0x10160c9cc`) reads
  `grid_size` at `[proto+0xb60]`/`[0xb68]` and NEVER `grid_offset` at
  `[0xb70]`/`[0xb78]`; its sample origin is `chunkPos << 5` converted to
  float, step `grid_size`. It computes no other coordinate.
- `base/prototypes/entity/entity-util.lua:305` says it outright:
  "cliffs are auto-placed with centers at (0, 0.5) offset from the grid".

Nauvis goes 0.943 -> **1.0000 recall, 1.0000 precision, ratio 1.000** at
both seeds - exact agreement with `find_entities_filtered`. Vulcanus
recall 0.792/0.870/0.803 -> 0.806/0.938/0.853; its over-placement is
unchanged and stays open on #18.

Why it hid for two months: the error moves NO placed cliff, because cell
centres are derived from their own constants. So `x mod 4 == 2`,
`y mod 4 == 2.5`, the preview agreement, and PR #57's substitution of the
game's own field values (zero cells moved) all passed - that fixture had
itself been captured at the port's assumed site. A substitution test can
falsify a value; it cannot falsify the site it was sampled at.

Fixture handling: the Vulcanus corner-fields oracle is re-captured at the
game's lattice; the previous capture is kept as `-legacy-y0.5`, since it
is valid ground truth for the site it names and is what lets a spec show
the blind spot. Both carry provenance.

Test changes are re-measurements, not accommodations:

- `cliffPlacement.spec.ts` now pins Nauvis at EQUALITY, not >= 0.85.
- `cliffResidual.spec.ts`'s "wrong cells sit on band boundaries" block
  asserted the residual exists; there are no wrong cells left, so it now
  asserts the set is empty.
- `cliffSmoothing.spec.ts`'s plane gains a `.37` offset: at `j*4` the old
  constant put corner elevations exactly ON band edges (measured: all 7
  disagreeing cells at distance 0, bilerp giving 7.1e-15), so it tested
  float rounding rather than smoothing.
- `vulcanusCliffs.spec.ts`'s independent blend now indexes corners
  instead of deriving positions from centres.
- `vulcanusOreCliffSeparation.spec.ts`: full-ore false positives 47 -> 46.
  The game still places zero; that finding is untouched.


Claude-Session: https://claude.ai/code/session_01BkKb3S2cGyRFFHSzuCQBgj

Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>

Author: wormeyman

docs/noise/client-preview-ROADMAP.md

@@ -308,14 +308,20 @@ Done = ore patches overlaid on land, responding to the frequency/size/richness s
       noise fields (`cliff_elevation_nauvis`, `cliffiness_nauvis`) plus a
       disasm-derived geometric placement rule (`crossesCliff` + the 4-tile grid +
       `toMaybeCliffOrientation` none/not-none predicate), levers wired
-      (`nauvis_cliff` frequency/continuity + `cliffSettings`). Validated ~94% against
-      real `find_entities_filtered{type="cliff"}` dumps (frac 0.943 seed 123456,
-      0.942 seed 777771) and cross-checked against a real
+      (`nauvis_cliff` frequency/continuity + `cliffSettings`). Validated **100%**
+      against real `find_entities_filtered{type="cliff"}` dumps (frac 1.000 at
+      both seeds since 2026-07-30; was 0.943/0.942 while the fields were sampled
+      half a tile off in y) and cross-checked against a real
       `factorio --generate-map-preview` render of the same seed/region (95.7%
       spatial agreement at render resolution, using the game's own emitted
       `[144,119,87]` pixels as ground truth). Full writeup: `cliffs-NOTES.md`.
       **All three items previously listed here as "still deferred" are now
-      retired, and none of them explained the ~6% Nauvis residual** (2026-07-28,
+      retired, and none of them explained the ~6% Nauvis residual - which is
+      itself now RESOLVED (2026-07-30): the fields were sampled at `j*4 + 0.5`
+      because the port added the prototype's `grid_offset` (a CENTRE offset) to
+      the sample position. Corrected, Nauvis is **1.0000 recall / 1.0000
+      precision / ratio 1.000** at both seeds, i.e. exact. The five earlier
+      causes were falsified on** (2026-07-28,
       the day five stated causes were falsified - see
       `notes-must-say-how-they-were-measured` and `cliffs-NOTES.md`):
       `fixImpossibleCells` was **ported** in PR #32 (`dea73ac`,

docs/noise/cliffs-NOTES.md

@@ -208,7 +208,34 @@ presets, but all 9 strings in `builtin-presets.json` decode to `u8 = 0` and
 either way; what is unknown is whether the wire carries smoothing at all. Worth an
 issue if the Nauvis render is ever driven by a Lakes/Island preset.
 
-## Validation result and the deferred residual
+## Validation result: EXACT since 2026-07-30 (the residual is resolved)
+
+> **The ~6% Nauvis residual documented throughout this section is GONE, and the
+> cause was none of the six things named for it.** The port sampled the two
+> fields at `(i*4, j*4 + 0.5)`, adding the prototype's `grid_offset {0, 0.5}` to
+> the SAMPLE position. That offset is a **centre** offset - the game data says so
+> at `base/prototypes/entity/entity-util.lua:305` ("cliffs are auto-placed with
+> centers at (0, 0.5) offset from the grid") and `CliffGenerator::crossingsForChunk`
+> reads `grid_size` (`[proto+0xb60]`/`[0xb68]`) and never `grid_offset`
+> (`[0xb70]`/`[0xb78]`), taking its sample origin from `chunkPos << 5`.
+>
+> Correcting it takes Nauvis from **0.943 / 0.943 to 1.0000 recall, 1.0000
+> precision, ratio 1.000 at both seeds** - exact agreement with
+> `find_entities_filtered` - and Vulcanus recall from 0.792/0.870/0.803 to
+> 0.806/0.938/0.853.
+>
+> Why it hid for two months, and why five other causes were "confirmed" instead:
+> the error moves **no placed cliff**. Cell centres are derived from their own
+> constants, so `x mod 4 == 2` / `y mod 4 == 2.5` held, the preview agreement
+> held, and PR #57's substitution of the game's own field values changed zero
+> cells - because that oracle fixture had itself been captured at the port's
+> assumed lattice. A substitution test can falsify a VALUE; it cannot falsify the
+> SITE it was sampled at.
+>
+> Everything below is preserved as the record of the investigation. Read the
+> numbers in it as historical.
+
+## Validation result and the deferred residual (historical - see above)
 
 Reimplementing the rule (sample `cliff_elevation_nauvis` + `cliffiness_nauvis` at the
 corner lattice via the oracle, apply the crossing rule + lattice) reproduced the real

src/noise/cliffs/cliffCatalog.ts

@@ -57,10 +57,31 @@ export const CLIFF_MARK_SIZE_PX = 4;
  */
 export const CLIFF_MARK_BACK_PX = 2;
 
-/** X coordinate of a cliff cell's center, in cell-local tiles. */
+/**
+ * Cliff cell centre, in cell-local tiles: `grid_size/2 + grid_offset`.
+ *
+ * **`grid_offset` belongs HERE, on the centre, and nowhere else.** The
+ * prototype's `grid_offset` is `{0, 0.5}` for `scale == 1` (both `cliff` and
+ * `cliff-vulcanus`), and `base/prototypes/entity/entity-util.lua:305` says what
+ * it is for in as many words: "cliffs are auto-placed with centers at (0, 0.5)
+ * offset from the grid". So `x = 4/2 + 0 = 2` and `y = 4/2 + 0.5 = 2.5`, which
+ * is why every dumped cliff satisfies `x mod 4 == 2`, `y mod 4 == 2.5`.
+ *
+ * The FIELDS are sampled at the bare lattice `(i*4, j*4)` - no offset. This was
+ * wrong here until 2026-07-30: a `CLIFF_CORNER_OFFSET_Y = 0.5` added the centre
+ * offset to the sample position too, displacing every field read half a tile in
+ * y. It was invisible because it does not move a single placed cliff - the
+ * centre constants below are independent - so the mod-4 checks, the preview
+ * agreement and the PR #57 field substitution all passed. Confirmed against the
+ * binary (`CliffGenerator::crossingsForChunk` reads `grid_size` at
+ * `[proto+0xb60]`/`[0xb68]` and never `grid_offset` at `[0xb70]`/`[0xb78]`;
+ * sample origin is `chunkPos << 5` converted to float) and by re-capturing the
+ * oracle at the correct lattice: Vulcanus recall 0.8701 -> 0.9379, FP 301 ->
+ * 235 at `[1500,1500]`.
+ */
 export const CLIFF_CELL_CENTER_X = 2;
 
-/** Y coordinate of a cliff cell's center, in cell-local tiles. */
+/** @see CLIFF_CELL_CENTER_X - carries the prototype's `grid_offset.y` of 0.5. */
 export const CLIFF_CELL_CENTER_Y = 2.5;
 
 /** Default `cliff_elevation_0` map-gen setting (elevation of the first cliff band). */
@@ -72,9 +93,6 @@ export const CLIFF_ELEVATION_INTERVAL_DEFAULT = 40;
 /** seed1 for the low-frequency cliffiness basis noise. */
 export const LOW_FREQ_CLIFFINESS_SEED1 = 86883;
 
-/** Y offset applied to a cliff cell corner. */
-export const CLIFF_CORNER_OFFSET_Y = 0.5;
-
 /** The `nauvis_cliff` autoplace control's frequency/size sliders (size doubles as continuity). */
 export interface CliffControls {
   frequency: number;

src/noise/cliffs/cliffPlacement.ts

@@ -12,7 +12,6 @@ import { makeCliffFields } from "./cliffFields";
 import {
   CLIFF_CELL_CENTER_X,
   CLIFF_CELL_CENTER_Y,
-  CLIFF_CORNER_OFFSET_Y,
   CLIFF_GRID_SIZE,
   getModifiedElevationInterval,
   isCliffPlaced,
@@ -232,11 +231,17 @@ export function makeCliffPlacementFromFields(
       if (bands.disabled === true) return [];
 
       const raw = new Map<string, number>();
+      /**
+       * Sampled at the BARE lattice `(i*4, j*4)`. The prototype's `grid_offset`
+       * is a CENTRE offset, not a sample offset - see `CLIFF_CELL_CENTER_X` -
+       * and `crossingsForChunk` never reads it. Adding it here (as this did
+       * until 2026-07-30) moves no cliff and costs ~7 points of recall.
+       */
       const rawElevation = (i: number, j: number): number => {
         const key = `${i},${j}`;
         let value = raw.get(key);
         if (value === undefined) {
-          value = cliffElevation(i * CLIFF_GRID_SIZE, j * CLIFF_GRID_SIZE + CLIFF_CORNER_OFFSET_Y);
+          value = cliffElevation(i * CLIFF_GRID_SIZE, j * CLIFF_GRID_SIZE);
           raw.set(key, value);
         }
         return value;
@@ -277,7 +282,7 @@ export function makeCliffPlacementFromFields(
         let sample = corners.get(key);
         if (sample === undefined) {
           const wx = i * CLIFF_GRID_SIZE;
-          const wy = j * CLIFF_GRID_SIZE + CLIFF_CORNER_OFFSET_Y;
+          const wy = j * CLIFF_GRID_SIZE;
           sample = { elev: elevationAt(i, j), cliff: cliffiness(wx, wy) };
           corners.set(key, sample);
         }

test/cliffPlacement.spec.ts

@@ -86,15 +86,21 @@ describe("cliff placement vs find_entities (~94% drift guard)", () => {
       const actual = c.cliffs.map(key);
       const matched = actual.filter((k) => predicted.has(k)).length;
       const frac = matched / actual.length;
-      expect(frac).toBeGreaterThanOrEqual(0.85);
-
-      // Over-placement guard. Measured ratio is exactly 1.000 at both seeds, so
-      // this is tight on purpose: cliff placement is deterministic, and anything
-      // that starts inventing cliffs should fail here rather than hide behind a
-      // recall figure.
       const precision = matched / predicted.size;
-      expect(precision).toBeGreaterThanOrEqual(0.85);
-      expect(predicted.size / actual.length).toBeLessThan(1.1);
+
+      // **EXACT since 2026-07-30.** 282/282 at seed 123456 and 52/52 at 777771 -
+      // every real cliff placed, nothing invented. The long-standing ~6%
+      // residual was the port sampling the fields at `j*4 + 0.5`: it added the
+      // prototype's `grid_offset {0, 0.5}`, which is a CENTRE offset, to the
+      // SAMPLE position (see `CLIFF_CELL_CENTER_X`).
+      //
+      // Pinned at equality on purpose. Placement is deterministic given the seed
+      // - there is no roll - so an inequality here would let a real regression
+      // hide inside the slack, which is how the 0.943 sat unexplained for two
+      // months while five different causes were proposed for it.
+      expect(frac).toBe(1);
+      expect(precision).toBe(1);
+      expect(predicted.size).toBe(actual.length);
     });
   }
 });

test/cliffResidual.spec.ts

@@ -127,9 +127,9 @@ describe("Nauvis cliff residual: water rejection cannot be the cause", () => {
   }
 });
 
-describe("Nauvis cliff residual: the wrong cells sit on band boundaries", () => {
+describe("Nauvis cliff residual: RESOLVED 2026-07-30 - there are no wrong cells", () => {
   for (const { seed } of CASES) {
-    it(`seed ${String(seed)}: mismatched cells are closer to a band edge than matched ones`, () => {
+    it(`seed ${String(seed)}: every placed cell is a real cliff, and every real cliff is placed`, () => {
       // Distance from the nearest cliff band boundary (`10 + 40k`), minimised
       // over the cell's four corners. Measured 2026-07-28:
       //
@@ -159,28 +159,33 @@ describe("Nauvis cliff residual: the wrong cells sit on band boundaries", () =>
         }
         return best;
       };
-      const median = (vals: number[]): number => {
-        const s = [...vals].sort((a, b) => a - b);
-        return s[Math.floor(0.5 * (s.length - 1))];
-      };
 
       const matched: number[] = [];
       const mismatched: number[] = [];
       for (const p of placed) (actual.has(key(p)) ? matched : mismatched).push(distance(p.x, p.y));
 
       expect(matched.length).toBeGreaterThan(40);
-      expect(mismatched.length).toBeGreaterThan(0);
-      // Pinned loosely: the measured gap is 3.4x and 4.3x, so 1.8x leaves room
-      // for the port to improve (which would SHRINK the mismatched set and could
-      // move its median either way) without going green on a regression that
-      // erased the effect entirely.
-      expect(median(matched) / median(mismatched)).toBeGreaterThan(1.8);
 
-      // Nauvis's wrong cells are on a KNIFE EDGE: the median sits 0.07 out of a
-      // 40-wide band, 0.18% of an interval. That is the number to compare
-      // Vulcanus against below - not the raw distance, since the two planets run
-      // different intervals (40 vs 120).
-      expect(median(mismatched) / 40).toBeLessThan(0.005);
+      // **The residual is GONE.** Not shrunk - zero. Every cell we place is a
+      // real cliff (no false positives) and the recall/precision spec in
+      // `cliffPlacement.spec.ts` now measures 1.0000 / 1.0000 / ratio 1.000 at
+      // both seeds, up from 0.943 / 0.943.
+      //
+      // The cause was the SAMPLE LATTICE, not the rule and not the field: the
+      // port added the prototype's `grid_offset {0, 0.5}` - a CENTRE offset -
+      // to the field sample position as well, reading every corner half a tile
+      // off in y. It moved no placed cliff, so every positional check passed.
+      // See `CLIFF_CELL_CENTER_X` in cliffCatalog.ts.
+      //
+      // This block used to assert the OPPOSITE - that mismatched cells exist
+      // and sit closer to band edges than matched ones (medians 3.4x and 4.3x
+      // apart). That measurement was real and is preserved in git; it described
+      // a marginal decision, not a cause, exactly as `boundary-proximity-is-not
+      // -a-cause` concluded. `distance` above is kept because the loop still
+      // partitions on it, which is what proves the mismatched set is empty
+      // because there is nothing in it - not because the loop never ran.
+      expect(mismatched.length).toBe(0);
+      expect(actual.size).toBe(placed.length);
     }, 120000);
   }
 });

test/cliffSmoothing.spec.ts

@@ -83,7 +83,20 @@ describe("cliff_smoothing behaviour", () => {
     // A plane. Bilinear interpolation of a plane is the plane itself, whatever
     // the knot spacing - so this is a sharp test that the four weights sum to 1
     // and that lo/hi bracket the corner rather than merely being near it.
-    cliffElevation: (x, y) => 0.7 * x + 0.3 * y + 100,
+    //
+    // The `.37` is load-bearing and was `100` until 2026-07-30. Corners are
+    // sampled at `(i*4, j*4)`, so a constant of 100 makes every corner
+    // elevation `2.8i + 1.2j + 100` - a multiple of 0.4 - and the default bands
+    // sit at `10 + 40n`, which such a value can hit EXACTLY. Measured: all 7
+    // cells that then disagreed between smoothing 0 and 1 had a corner at
+    // distance exactly 0 from a band edge, where the raw path computes 0 and
+    // the bilerp computes 7.1e-15, so a `>=` goes two ways for reasons that have
+    // nothing to do with smoothing. `.37` is not a multiple of 0.4, so no corner
+    // can land on a band edge and the test measures the property it names.
+    //
+    // It passed before only because the port sampled at `j*4 + 0.5`, adding
+    // 0.15 and knocking the plane off the boundaries by accident.
+    cliffElevation: (x, y) => 0.7 * x + 0.3 * y + 100.37,
     cliffiness: alwaysCliffy,
   };
 

test/fixtures/PROVENANCE.json

@@ -184,9 +184,13 @@
       "factorioVersion": "2.1.12",
       "evidence": "stated in the fixture's own _comment"
     },
+    "oracle-vulcanus-cliff-corner-fields-legacy-y0.5.seed123456.json": {
+      "factorioVersion": "2.1.12",
+      "evidence": "captured 2026-07-29 by test/oracle/capture.ts vulcanus-cliff-corner-fields against the installed binary, which pnpm refs:sync --check reported in sync at 2.1.12 at capture time. Superseded 2026-07-30: the sample lattice (j*4+0.5) was the PORT's assumption, not the game's. Retained deliberately - it is valid ground truth for the site it names, and keeping it is what lets a spec show that substituting fields captured at OUR site cannot falsify that site."
+    },
     "oracle-vulcanus-cliff-corner-fields.seed123456.json": {
       "factorioVersion": "2.1.12",
-      "evidence": "captured 2026-07-29 by test/oracle/capture.ts vulcanus-cliff-corner-fields against the installed binary, which pnpm refs:sync --check reported in sync at 2.1.12 at capture time"
+      "evidence": "re-captured 2026-07-30 by test/oracle/capture.ts vulcanus-cliff-corner-fields against the installed binary, which pnpm refs:sync --check reported in sync at 2.1.12 at capture time. Supersedes the 2026-07-29 capture (kept as the -legacy-y0.5 fixture): that one sampled at j*4+0.5, the port's assumed lattice, not the game's."
     },
     "oracle-vulcanus-cliff-entities.seed123456.json": {
       "factorioVersion": "2.1.12",