spec(conv2d_layer): forward slides its kernel; scalar conv_at for the nn-conv widget (Closes #10612)

.trinity/seals/Conv2d.json

@@ -1,19 +1,19 @@
 {
-  "gen_hash_c": "sha256:64cc98a2d8a6874d5648322d454539821632ae913fab3dbebeb8265a4712d14f",
-  "gen_hash_rust": "sha256:1260822c3590b2ec7a7dc5d33d16e0db36f6b94e13bf459874fafa7339b681b1",
-  "gen_hash_verilog": "sha256:2e469acfecbbb53a7d0187c89ef6e8133bd185977fed689d5edd11b45a76f26c",
-  "gen_hash_zig": "sha256:1a8ca800f2b2cb4496acfb41062a9ef5cda7634f84034f7f6363017f924d593f",
+  "gen_hash_c": "sha256:60f9b428ab8ceb7d2f4654a727012c82d12eafbf92d8dde167418d5a43fe0e10",
+  "gen_hash_rust": "sha256:1b00b1fe570ae33bde1434d0125dc478df1d00f0470df7efb378e5dff574bb30",
+  "gen_hash_verilog": "sha256:83f9eae101994f85bb5be9fb7f154f01d88ff0330c3286322f584689c910b420",
+  "gen_hash_zig": "sha256:c1ec25dbf1daf046dfa6600b183fe3b9501525038b576582e66947e39379074f",
   "module": "Conv2d",
   "ring": 12,
-  "sealed_at": "2026-10-11T10:43:18Z",
-  "spec_hash": "sha256:90bf260b6fe12b6d64a323eec8c887a7ad4882d73f0307c20a240317ed64f543",
+  "sealed_at": "2026-10-19T01:01:23Z",
+  "spec_hash": "sha256:9c3e20d01e59ec8df1ae330ac9c3c27f8e67db34d524475733abaf0f3382b7f0",
   "spec_path": "specs/ml/layers/conv2d_layer.t27",
   "tests": {
     "failed": 0,
     "failing": [],
     "forced": false,
-    "passed": 2,
-    "total": 2,
+    "passed": 14,
+    "total": 14,
     "vacuous": 0
   }
 }

.trinity/seals/layers_Conv2d.json

@@ -1,5 +1,5 @@
 {
-  "built_by": "t27c-bootstrap@0.5.2+cc19d1fcd",
+  "built_by": "t27c-bootstrap@0.5.2+239b4b059",
   "closure": [
     {
       "spec": "specs/base/types.t27",
@@ -11,27 +11,27 @@
     }
   ],
   "config": "gen=c,rust,verilog,zig;tests=zig",
-  "gen_hash_c": "sha256:64cc98a2d8a6874d5648322d454539821632ae913fab3dbebeb8265a4712d14f",
-  "gen_hash_rust": "sha256:1260822c3590b2ec7a7dc5d33d16e0db36f6b94e13bf459874fafa7339b681b1",
-  "gen_hash_verilog": "sha256:2e469acfecbbb53a7d0187c89ef6e8133bd185977fed689d5edd11b45a76f26c",
-  "gen_hash_zig": "sha256:1a8ca800f2b2cb4496acfb41062a9ef5cda7634f84034f7f6363017f924d593f",
+  "gen_hash_c": "sha256:60f9b428ab8ceb7d2f4654a727012c82d12eafbf92d8dde167418d5a43fe0e10",
+  "gen_hash_rust": "sha256:1b00b1fe570ae33bde1434d0125dc478df1d00f0470df7efb378e5dff574bb30",
+  "gen_hash_verilog": "sha256:83f9eae101994f85bb5be9fb7f154f01d88ff0330c3286322f584689c910b420",
+  "gen_hash_zig": "sha256:c1ec25dbf1daf046dfa6600b183fe3b9501525038b576582e66947e39379074f",
   "module": "Conv2d",
   "ring": 12,
   "seal_schema": 2,
-  "sealed_at": "2026-10-11T10:43:18Z",
+  "sealed_at": "2026-10-19T01:01:23Z",
   "sealed_by": "t27c-bootstrap@0.5.2",
-  "spec_hash": "sha256:90bf260b6fe12b6d64a323eec8c887a7ad4882d73f0307c20a240317ed64f543",
+  "spec_hash": "sha256:9c3e20d01e59ec8df1ae330ac9c3c27f8e67db34d524475733abaf0f3382b7f0",
   "spec_path": "specs/ml/layers/conv2d_layer.t27",
   "tests": {
     "failed": 0,
     "failing": [],
     "forced": false,
-    "passed": 2,
-    "total": 2,
+    "passed": 14,
+    "total": 14,
     "vacuous": 0
   },
   "toolchain": {
-    "test_runner": "sha256:1982ae948e24d00c5d221ccec12dfce1eb01204b53e37b0f857f93b16d64629d",
+    "test_runner": "sha256:8512b80da533b79289110fca4a65f3336abd7a059ca4b72ea595d5e33ff8ad10",
     "zig": "0.16.0"
   }
 }

docs/now/2026-10-19-conv2d-slides-its-kernel.md

@@ -0,0 +1,14 @@
+# NOW -- conv2d_layer.t27 slides its kernel, plus scalar fns for the nn-conv widget (2026-10-19)
+
+## specs/ml/layers/conv2d_layer.t27 (Closes #10612)
+
+- Why: `forward` multiplied one kernel-sized patch elementwise by the weights and added the bias, so the 2x2 input with the 2x2 kernel gave four numbers where a convolution gives one (3.25), and its test checked the four. No fn lowered to WASM, so course 10 lesson 10 (trinity `widgets/nn-conv`) falls back to average pooling.
+- `forward` is now the deep-learning convolution (2-D cross-correlation, kernel not flipped) over one H x W channel with stride and zero padding; output side `(n + 2p - k) / s + 1`; bias once per output cell. `backward` is the sliding gradient (grad_input, grad_weights; the bias gradient is its return value, which keeps it at 8 parameters and the site compiler warning-free). Both now take `height, width`.
+- Scalar entry points: `out_size(n, kernel, stride, padding)`, `conv_at(kernel_id, stride, padding, index)` on the fixed 5x5 square image with 4 fixed 3x3 kernels (identity, Sobel x, Sobel y, box blur), `kernel_at`, `image_at`, and `default_at(index)` for the old 2x2 case. `scripts/t27-logic.mjs` lowers 5 fns (`outSize`, `convAt`, `kernelAt`, `imageAt`, `defaultAt`); in node `defaultAt(0)` is 3.25 and every row of the issue's table comes back exact.
+- VERSION 2 with a dated header note and a test that VERSION is 2.
+- t27c (bootstrap unchanged since 239b4b059), zig 0.16.0: test-report 14 of 14 pass, 0 vacuous, 5 invariants proved. Each row of the issue's table has its own test; a 3x3 input with the 2x2 kernel pins the slide (3.6, 5.45, 9.15, 11.0).
+- Negative controls: 9 mutants (elementwise fallback, stride ignored, padding ignored, kernel flipped, bias k*k times, and in backward padding ignored, kernel flipped, has_bias ignored, stride ignored), 9 killed.
+- Site compiler (trinity `public/t27/t27_compiler.wasm`): typecheck ok, 0 errors, 0 discarded, all seven backends print.
+- `t27c dupes --bodies` names no group in this spec (master had one: three unused `default_*` helpers with one body; removed).
+- Seal saved (`Conv2d.json`, `layers_Conv2d.json`). No spec `use`s Conv2d.
+- Not here: re-vendoring into trinity and switching `nn-conv/tool.js` from `pool4At` to `convAt`.