diff --git a/Makefile b/Makefile
index 1ea2cfd5..60f65f02 100644
--- a/Makefile
+++ b/Makefile
@@ -332,6 +332,7 @@ TESTS	= $(BIN)/test_smoke $(BIN)/test_lex $(BIN)/test_parse $(BIN)/test_check \
 	  $(BIN)/test_f64crossmod_run \
 	  $(BIN)/test_tuprecv_run \
 	  $(BIN)/test_f64xmm_run \
+	  $(BIN)/test_f32lit_run \
 	  $(BIN)/test_tuprecv_f64_run \
 	  $(BIN)/test_floats_run \
 	  $(BIN)/test_size_type_run \
@@ -1133,6 +1134,11 @@ $(BIN)/test_f64xmm_run: test/wcc/955_f64xmm_run.c $(BIN)/ww \
 		$(LIB)/libwwrt.a | $(BIN)
 	$(CC) $(CFLAGS) -o $@ $<
 
+$(BIN)/test_f32lit_run: test/wcc/964_f32lit_run.c $(BIN)/ww \
+		$(BIN)/w6c $(BIN)/w6c_ww $(BIN)/w6a $(BIN)/w6l \
+		$(LIB)/libwwrt.a | $(BIN)
+	$(CC) $(CFLAGS) -o $@ $<
+
 $(BIN)/test_tuprecv_f64_run: test/wcc/956_tuprecv_f64_run.c $(BIN)/ww \
 		$(BIN)/w6c $(BIN)/w6c_ww $(BIN)/w6a $(BIN)/w6l \
 		$(LIB)/libwwrt.a | $(BIN)
diff --git a/cmd/w6c/cgen.c b/cmd/w6c/cgen.c
index 903ab0d8..299776a3 100644
--- a/cmd/w6c/cgen.c
+++ b/cmd/w6c/cgen.c
@@ -1887,12 +1887,20 @@ cgexpr(Cg *c, Node *n, Local *locals)
 	case N_RUNELIT:
 		if (node_isfloat(n)) {
 			cgexpr_float(c, (double)(long long)n->uval);
+			/* #104: cgexpr_float materialises a DOUBLE in X0; an
+			 * f32-typed literal must narrow with hardware single-
+			 * rounding so the downstream MOVSS reads a true single. */
+			if (node_isf32(n))
+				ins2(c, A_CVTSD2SS, areg(D_X0), areg(D_X0));
 			break;
 		}
 		cgexpr_int(c, (long long)n->uval);
 		break;
 	case N_FLOATLIT:
 		cgexpr_float(c, n->fval);
+		/* #104: narrow the double in X0 to single for an f32 literal. */
+		if (node_isf32(n))
+			ins2(c, A_CVTSD2SS, areg(D_X0), areg(D_X0));
 		break;
 	case N_STRLIT: {
 		/* str IS []u8: the (ptr, len, cap) triple — ptr in AX, len in
diff --git a/selfhost/cmd/w6c/main.combined.ww b/selfhost/cmd/w6c/main.combined.ww
index 3ab14743..b4a40076 100644
--- a/selfhost/cmd/w6c/main.combined.ww
+++ b/selfhost/cmd/w6c/main.combined.ww
@@ -14209,6 +14209,12 @@ fn cgexpr(c: *cgen, n: *node) void = {
 			let fv: f64 = (n.uval: i64): f64;
 			let pu: *u64 = (&fv): *u64;
 			cgfloatbits(c, *pu);
+			// #104: cgfloatbits materialises a DOUBLE in X0; an
+			// f32-typed literal must narrow with hardware single-
+			// rounding so the downstream MOVSS reads a true single.
+			if (isf32type(c, n)) {
+				emitline("\tCVTSD2SS\tX0, X0\n");
+			};
 			return;
 		};
 		// Print signed (i64), not unsigned (u64). C cgen uses
@@ -14222,10 +14228,12 @@ fn cgexpr(c: *cgen, n: *node) void = {
 	};
 	if (k == nkind.N_FLOATLIT) {
 		// The bits come from n.uval — the parser populates it from
-		// the lexer's bitcast of t.fval. The f32 narrowing is handled
-		// at the consumer site, not here — the literal always carries
-		// the full double precision until typed by context.
+		// the lexer's bitcast of t.fval.
 		cgfloatbits(c, n.uval);
+		// #104: narrow the double in X0 to single for an f32 literal.
+		if (isf32type(c, n)) {
+			emitline("\tCVTSD2SS\tX0, X0\n");
+		};
 		return;
 	};
 	if (k == nkind.N_RUNELIT) {
diff --git a/selfhost/cmd/wcc/cgenexpr.ww b/selfhost/cmd/wcc/cgenexpr.ww
index 00d05171..b78b6e67 100644
--- a/selfhost/cmd/wcc/cgenexpr.ww
+++ b/selfhost/cmd/wcc/cgenexpr.ww
@@ -50,6 +50,12 @@ fn cgexpr(c: *cgen, n: *node) void = {
 			let fv: f64 = (n.uval: i64): f64;
 			let pu: *u64 = (&fv): *u64;
 			cgfloatbits(c, *pu);
+			// #104: cgfloatbits materialises a DOUBLE in X0; an
+			// f32-typed literal must narrow with hardware single-
+			// rounding so the downstream MOVSS reads a true single.
+			if (isf32type(c, n)) {
+				emitline("\tCVTSD2SS\tX0, X0\n");
+			};
 			return;
 		};
 		// Print signed (i64), not unsigned (u64). C cgen uses
@@ -63,10 +69,12 @@ fn cgexpr(c: *cgen, n: *node) void = {
 	};
 	if (k == nkind.N_FLOATLIT) {
 		// The bits come from n.uval — the parser populates it from
-		// the lexer's bitcast of t.fval. The f32 narrowing is handled
-		// at the consumer site, not here — the literal always carries
-		// the full double precision until typed by context.
+		// the lexer's bitcast of t.fval.
 		cgfloatbits(c, n.uval);
+		// #104: narrow the double in X0 to single for an f32 literal.
+		if (isf32type(c, n)) {
+			emitline("\tCVTSD2SS\tX0, X0\n");
+		};
 		return;
 	};
 	if (k == nkind.N_RUNELIT) {
diff --git a/selfhost/cmd/wwdump/main.combined.ww b/selfhost/cmd/wwdump/main.combined.ww
index c0ed90b8..a44c3b54 100644
--- a/selfhost/cmd/wwdump/main.combined.ww
+++ b/selfhost/cmd/wwdump/main.combined.ww
@@ -14209,6 +14209,12 @@ fn cgexpr(c: *cgen, n: *node) void = {
 			let fv: f64 = (n.uval: i64): f64;
 			let pu: *u64 = (&fv): *u64;
 			cgfloatbits(c, *pu);
+			// #104: cgfloatbits materialises a DOUBLE in X0; an
+			// f32-typed literal must narrow with hardware single-
+			// rounding so the downstream MOVSS reads a true single.
+			if (isf32type(c, n)) {
+				emitline("\tCVTSD2SS\tX0, X0\n");
+			};
 			return;
 		};
 		// Print signed (i64), not unsigned (u64). C cgen uses
@@ -14222,10 +14228,12 @@ fn cgexpr(c: *cgen, n: *node) void = {
 	};
 	if (k == nkind.N_FLOATLIT) {
 		// The bits come from n.uval — the parser populates it from
-		// the lexer's bitcast of t.fval. The f32 narrowing is handled
-		// at the consumer site, not here — the literal always carries
-		// the full double precision until typed by context.
+		// the lexer's bitcast of t.fval.
 		cgfloatbits(c, n.uval);
+		// #104: narrow the double in X0 to single for an f32 literal.
+		if (isf32type(c, n)) {
+			emitline("\tCVTSD2SS\tX0, X0\n");
+		};
 		return;
 	};
 	if (k == nkind.N_RUNELIT) {
diff --git a/test/wcc/964_f32lit_run.c b/test/wcc/964_f32lit_run.c
new file mode 100644
index 00000000..14e2e397
--- /dev/null
+++ b/test/wcc/964_f32lit_run.c
@@ -0,0 +1,219 @@
+/*
+ * 964_f32lit_run — runtime + byte-id regression net for #104 fold-1: an
+ * f32-typed float literal must NARROW to single precision in X0 before
+ * the f32 consumer reads it. Both stages (cstage cgen.c cgexpr_float,
+ * wwstage cgenexpr.ww cgfloatbits) materialise a float literal as a
+ * 64-bit DOUBLE in X0 (MOVQ bits -> MOVSD). For an f32-typed literal the
+ * downstream MOVSS store/return then reads the LOW 4 BYTES of that
+ * double — garbage (0x00000000 == 0.0f for most clean values, which is
+ * why 0.0 coincidentally survived the bug). fold-1 appends CVTSD2SS
+ * X0,X0 at both literal sites (N_FLOATLIT + the float-typed N_INTLIT
+ * arm) when the node is f32-typed, so the value reaches X0 as a true
+ * single. Both stages emit byte-identical asm, so the 990-997 byte-id
+ * gates can NEVER catch a reintroduction — only an executed-and-checked
+ * runtime probe can. (951_f64cgen is GATE-BLIND here: its f32 rows only
+ * assert NaN ordering, never a concrete f32 value.)
+ *
+ * SCOPE: fold-1 covers literals that carry an explicit f32 TYPE — the
+ * `f32` suffix (`1.0f32`, `1.5f32`) and the no-decimal N_INTLIT-float
+ * arm (`8f32`). An UN-suffixed literal in an f32 context (`let x: f32 =
+ * 1.0`) stays ty_untyped_float through the checker, so the literal node
+ * is never f32-typed and fold-1's branch can't fire — materialised as a
+ * double, stored low-4-bytes -> 0.0f. Fixing that needs fold-2: the
+ * checker lowering untyped-float literals to their f32 context type
+ * (#104, both checkers). This probe therefore uses suffixed literals
+ * exclusively; the un-suffixed gap is tracked under #104 fold-2.
+ *
+ * Each row carries BOTH dimensions (like 955_f64xmm_run):
+ *   (a) cstage `ww build` + run, asserting the exit code.
+ *   (b) w6c vs w6c_ww `.s` cmp — FAILS if the stages diverge (rule-10).
+ */
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <unistd.h>
+#include <sys/stat.h>
+#include <sys/wait.h>
+
+static int
+runwait(const char *cmd)
+{
+	int rc = system(cmd);
+	if (rc == -1) return -1;
+	if (WIFEXITED(rc)) return WEXITSTATUS(rc);
+	return -1;
+}
+
+struct row { const char *label; const char *src; int want_exit; };
+
+static const struct row rows[] = {
+	/* The hole 951 misses: a CONCRETE non-trivial f32 value. On the
+	 * bug `let x: f32 = 1.0f32` stores the low 4 bytes of double 1.0
+	 * (== 0x00000000 == 0.0f), so x:f64 == 0.0 != 1.0 -> 1. */
+	{ "bare_value",
+	  "package main;\n"
+	  "export fn main() i32 = {\n"
+	  "	let x: f32 = 1.0f32;\n"
+	  "	if (x: f64 != 1.0) { return 1; };\n"
+	  "	return 0;\n"
+	  "};\n", 0 },
+	/* arith on f32 literals: 1.5 + 2.5 == 4.0. On the bug both operands
+	 * land as 0.0f -> sum 0.0 != 4.0. */
+	{ "arith",
+	  "package main;\n"
+	  "export fn main() i32 = {\n"
+	  "	let a: f32 = 1.5f32;\n"
+	  "	let b: f32 = 2.5f32;\n"
+	  "	let s: f32 = a + b;\n"
+	  "	if (s: f64 != 4.0) { return 1; };\n"
+	  "	return 0;\n"
+	  "};\n", 0 },
+	/* value-propagation: f32 returned through an f32 fn + arith on an
+	 * f32 literal, truncated to i32. 2.5 + 1.5 == 4.0 -> 4. */
+	{ "return_arith",
+	  "package main;\n"
+	  "fn g() f32 = { return 2.5f32; };\n"
+	  "export fn main() i32 = {\n"
+	  "	let r: f32 = g() + 1.5f32;\n"
+	  "	return r: i32;\n"
+	  "};\n", 4 },
+	/* the float-typed N_INTLIT arm (`8f32` — no decimal, f32 suffix).
+	 * Same materialiser, same fold-1 branch. -> 8. */
+	{ "intlit_f32_arm",
+	  "package main;\n"
+	  "export fn main() i32 = {\n"
+	  "	let y: f32 = 8f32;\n"
+	  "	return y: i32;\n"
+	  "};\n", 8 },
+	/* genuine single-rounding: 2^24 + 1 is NOT representable in f32 and
+	 * rounds back to 2^24 (round-to-even). If the add ran in double it
+	 * would be 16777217.0 != 16777216.0 -> 1. Proves the value is a
+	 * true single, not the low half of a double. */
+	{ "single_round",
+	  "package main;\n"
+	  "export fn main() i32 = {\n"
+	  "	let big: f32 = 16777216.0f32;\n"
+	  "	let r: f32 = big + 1.0f32;\n"
+	  "	if (r: f64 != 16777216.0) { return 1; };\n"
+	  "	return 0;\n"
+	  "};\n", 0 },
+	{ NULL, NULL, 0 }
+};
+
+static int
+slurp_eq(const char *a, const char *b)
+{
+	FILE *fa = fopen(a, "rb");
+	FILE *fb = fopen(b, "rb");
+	if (!fa || !fb) { if (fa) fclose(fa); if (fb) fclose(fb); return -1; }
+	int rc = 0;
+	for (;;) {
+		int ca = fgetc(fa);
+		int cb = fgetc(fb);
+		if (ca != cb) { rc = -1; break; }
+		if (ca == EOF) break;
+	}
+	fclose(fa); fclose(fb);
+	return rc;
+}
+
+int
+main(void)
+{
+	const char *bin = getenv("BIN");
+	if (!bin) bin = "out/bin";
+	char absbin[1024];
+	if (bin[0] != '/') {
+		char cwd[1024];
+		if (getcwd(cwd, sizeof cwd) == NULL) return 1;
+		snprintf(absbin, sizeof absbin, "%s/%s", cwd, bin);
+		bin = absbin;
+	}
+
+	char w6c[1100], w6c_ww[1100];
+	snprintf(w6c, sizeof w6c, "%s/w6c", bin);
+	snprintf(w6c_ww, sizeof w6c_ww, "%s/w6c_ww", bin);
+	if (access(w6c_ww, X_OK) != 0) {
+		fprintf(stderr, "f32lit: w6c_ww missing — cannot run the "
+		    "cs==ww byte-id gate (the whole point of this test)\n");
+		return 1;
+	}
+
+	int n = 0, fail = 0;
+	for (int i = 0; rows[i].src; i++, n++) {
+		char src[64];
+		snprintf(src, sizeof src, "/tmp/wwf32l_%d_%d.ww", getpid(), i);
+		FILE *f = fopen(src, "wb");
+		if (f == NULL) { fail++; continue; }
+		fputs(rows[i].src, f);
+		fclose(f);
+
+		/* (a) cstage build + run. */
+		char tmpdir[64];
+		snprintf(tmpdir, sizeof tmpdir, "/tmp/wwf32l_%d_d_%d",
+		    getpid(), i);
+		mkdir(tmpdir, 0755);
+
+		char cmd[2048];
+		snprintf(cmd, sizeof cmd, "cd %s && %s/ww build %s",
+		    tmpdir, bin, src);
+		if (runwait(cmd) != 0) {
+			fprintf(stderr, "row[%s]: cstage build failed\n",
+			    rows[i].label);
+			fail++;
+			unlink(src); rmdir(tmpdir);
+			continue;
+		}
+
+		char outbin[128];
+		const char *base = strrchr(src, '/');
+		base = base ? base + 1 : src;
+		snprintf(outbin, sizeof outbin, "%s/%s", tmpdir, base);
+		char *dot = strrchr(outbin, '.');
+		if (dot && strcmp(dot, ".ww") == 0) *dot = '\0';
+
+		int got = runwait(outbin);
+		if (got != rows[i].want_exit) {
+			fprintf(stderr, "row[%s]: cstage exit %d, want %d\n",
+			    rows[i].label, got, rows[i].want_exit);
+			fail++;
+		}
+		unlink(outbin); rmdir(tmpdir);
+
+		/* (b) cs==ww byte-id gate: emit .s from both stages, cmp. */
+		char cs_s[64], ws_s[64];
+		snprintf(cs_s, sizeof cs_s, "/tmp/wwf32l_%d_%d_cs.s",
+		    getpid(), i);
+		snprintf(ws_s, sizeof ws_s, "/tmp/wwf32l_%d_%d_ww.s",
+		    getpid(), i);
+
+		snprintf(cmd, sizeof cmd, "%s -o %s %s 2>/dev/null",
+		    w6c, cs_s, src);
+		if (runwait(cmd) != 0) {
+			fprintf(stderr, "row[%s]: w6c failed\n", rows[i].label);
+			fail++; unlink(src); continue;
+		}
+		snprintf(cmd, sizeof cmd, "%s -o %s %s 2>/dev/null",
+		    w6c_ww, ws_s, src);
+		if (runwait(cmd) != 0) {
+			fprintf(stderr, "row[%s]: w6c_ww failed\n",
+			    rows[i].label);
+			fail++; unlink(src); unlink(cs_s); continue;
+		}
+		if (slurp_eq(cs_s, ws_s) != 0) {
+			fprintf(stderr,
+			    "row[%s]: cstage/wwstage .s DIFFER (rule-10 "
+			    "byte-id violation)\n", rows[i].label);
+			fail++;
+		}
+		unlink(src); unlink(cs_s); unlink(ws_s);
+	}
+
+	if (fail) {
+		fprintf(stderr, "%d/%d f32 literal-materialise tests failed\n",
+		    fail, n);
+		return 1;
+	}
+	printf("f32lit: %d/%d ok (cstage run + cs==ww byte-id)\n", n, n);
+	return 0;
+}