Snapshot: fog implementation + fidelity tooling baseline (pre bilinear-clamp fix)

Per-vertex GS fog end-to-end (gs_stub emit incl. persp_emit5, gs_prim_list_feeder
XYZ2->XYZF2 on PRIM.FGE, gs_make_sh3_scheduler_fixture.py F/FGE packing), new fog
TBs, fidelity attribution tooling. Functional baseline before removing the dead
bilinear lerp8 clamps (Codex: 161-node comb loop -> -0.042ns setup fail).

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

tools/gs_localmem.py

@@ -13,6 +13,9 @@ Address convention matches the codebase: VRAM byte base = PTR*256 (TBP0/DBP/CBP
 and PTRs are page-aligned (multiple of 32) so page_index*8192 composes correctly off that base.
 """
 
+import os
+import re
+
 # block grid is shared by PSMCT32 and PSMT8 (4 rows x 8 cols), value = block index within page
 BLOCK = [
     [ 0, 1, 4, 5,16,17,20,21],
@@ -52,6 +55,40 @@ def psmt8_addr(tbp, fbw, x, y):
     block_idx = BLOCK[(y >> 4) & 3][(x >> 4) & 7]
     return tbp*256 + page_index*8192 + block_idx*256 + COL8[y & 15][x & 15]
 
+_COL4 = None
+_BLOCK4 = (
+    (0,2,8,10), (1,3,9,11), (4,6,12,14), (5,7,13,15),
+    (16,18,24,26), (17,19,25,27), (20,22,28,30), (21,23,29,31),
+)
+
+def _col4_table():
+    """Load the canonical PSMT4 permutation from the RTL contract."""
+    global _COL4
+    if _COL4 is not None:
+        return _COL4
+    rtl = os.path.join(os.path.dirname(os.path.dirname(os.path.abspath(__file__))),
+                       "rtl", "gif_gs", "gs_swizzle_psmt4_stub.sv")
+    text = open(rtl, encoding="utf-8").read()
+    lo = text.index("function automatic logic [8:0] col_idx_psmt4")
+    hi = text.index("endfunction", lo)
+    tab = [None] * 512
+    for key, value in re.findall(r"9'd(\d+)\s*:\s*return\s+9'd(\d+)", text[lo:hi]):
+        tab[int(key)] = int(value)
+    tab[511] = 511  # explicit default arm in the RTL function
+    if any(v is None for v in tab) or sorted(tab) != list(range(512)):
+        raise RuntimeError("PSMT4 RTL column table is incomplete or non-bijective")
+    _COL4 = tuple(tab)
+    return _COL4
+
+def psmt4_addr(tbp, fbw, x, y):
+    """Return (byte address, high-nibble select) for a PSMT4 texel."""
+    if fbw & 1:
+        raise ValueError(f"PSMT4 TBW must be even, got {fbw}")
+    page_index = (y >> 7) * (fbw >> 1) + (x >> 7)
+    block_idx = _BLOCK4[(y >> 4) & 7][(x >> 5) & 3]
+    nib = _col4_table()[((y & 15) << 5) | (x & 31)]
+    return tbp*256 + page_index*8192 + block_idx*256 + (nib >> 1), bool(nib & 1)
+
 class LocalMem:
     """4 MiB GS VRAM. Seed from the dump's initial snapshot, then replay host->local uploads in order."""
     SIZE = 0x400000
@@ -82,6 +119,17 @@ class LocalMem:
                 out[r+x] = self.m[a] if 0 <= a < self.SIZE else 0
         return out
 
+    def read_psmt4(self, tbp, fbw, tw, th):
+        """Return one unpacked 0..15 index byte per raster-order texel."""
+        out = bytearray(tw*th)
+        for y in range(th):
+            r = y*tw
+            for x in range(tw):
+                a, hi = psmt4_addr(tbp, fbw, x, y)
+                b = self.m[a] if 0 <= a < self.SIZE else 0
+                out[r+x] = ((b >> 4) & 0xF) if hi else (b & 0xF)
+        return out
+
     def read_ct32_word(self, dbp, dbw, x, y):
         a = ct32_addr(dbp, dbw, x, y)
         return int.from_bytes(self.m[a:a+4], "little") if a+4 <= self.SIZE else 0