From 2e9e4fe8731b82297509b09e7439725f6db4194d Mon Sep 17 00:00:00 2001
From: agra <alex@swipelab.co>
Date: Thu, 4 Jun 2026 16:48:34 +0300
Subject: [PATCH 1/5] =?UTF-8?q?WIP:=20float=20numeric-limit=20accessors=20?=
 =?UTF-8?q?(NL.2)=20=E2=80=94=20blocked=20on=200091=20(nan=20!=3D=20nan);?=
 =?UTF-8?q?=20examples/docs=20pending?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 src/ir/expr_typer.zig    | 15 +++++++++-----
 src/ir/lower.zig         | 44 +++++++++++++++++++++++++---------------
 src/ir/type_resolver.zig | 43 +++++++++++++++++++++++++++++++++++++++
 3 files changed, 81 insertions(+), 21 deletions(-)

diff --git a/src/ir/expr_typer.zig b/src/ir/expr_typer.zig
index 51a156d..1b8d8a9 100644
--- a/src/ir/expr_typer.zig
+++ b/src/ir/expr_typer.zig
@@ -126,10 +126,13 @@ pub const ExprTyper = struct {
                         if (info.ty) |t| return t;
                     }
                 }
-                // Numeric-limit accessor: `<IntType>.min` / `.max` is a comptime
-                // const of the queried integer type — mirrors the lowerFieldAccess
-                // intercept so inference reports the same type (without it the
-                // const would be mistyped, e.g. boxed into an Any slot).
+                // Numeric-limit accessor: `<Type>.min`/`.max` (int or float) or a
+                // float-only `.epsilon`/`.min_positive`/`.true_min`/`.inf`/`.nan`
+                // is a comptime const of the queried type — mirrors the
+                // lowerFieldAccess intercept so inference reports the same type
+                // (without it the const would be mistyped, e.g. boxed into an Any
+                // slot). Only valid folds carry a type here; the cross-type error
+                // cases fall through (lowerNumericLimit emits the diagnostic).
                 {
                     const type_name: ?[]const u8 = switch (fa.object.data) {
                         .identifier => |id| id.name,
@@ -137,7 +140,9 @@ pub const ExprTyper = struct {
                         else => null,
                     };
                     if (type_name) |tn| {
-                        if (TypeResolver.integerLimitFor(tn, fa.field) != null) {
+                        if (TypeResolver.integerLimitFor(tn, fa.field) != null or
+                            TypeResolver.floatLimitFor(tn, fa.field) != null)
+                        {
                             if (TypeResolver.resolveBuiltinName(tn, &self.l.module.types)) |t| return t;
                         }
                     }
diff --git a/src/ir/lower.zig b/src/ir/lower.zig
index a83feb0..e4951f9 100644
--- a/src/ir/lower.zig
+++ b/src/ir/lower.zig
@@ -4904,33 +4904,45 @@ pub const Lowering = struct {
         return self.lowerFieldAccessOnType(obj, obj_ty, fa.field, span);
     }
 
-    /// Numeric-limit accessor intercept (`<IntType>.min` / `.max`), a sibling of
-    /// the `error.X` / `Struct.CONST` / pack-arity identifier-receiver intercepts
-    /// in `lowerFieldAccess`. Folds an integer type's `.min`/`.max` to a comptime
-    /// const of that type via the shared `TypeResolver` width logic (no second
-    /// width parser) + the existing `constInt` const path. Returns null when this
-    /// is not an integer-limit access, so the caller continues normal field
-    /// lowering. A `.min`/`.max` on a builtin NON-numeric receiver
-    /// (`bool`/`string`/`void`/`Any`/`noreturn`) is a clean diagnostic here (then
-    /// a placeholder, so lowering finishes and `hasErrors()` aborts the build); a
-    /// float receiver falls through (float limits are NL.2).
+    /// Numeric-limit accessor intercept (`<Type>.min`/`.max`/`.epsilon`/
+    /// `.min_positive`/`.true_min`/`.inf`/`.nan`), a sibling of the `error.X` /
+    /// `Struct.CONST` / pack-arity identifier-receiver intercepts in
+    /// `lowerFieldAccess`. Folds the limit to a comptime const of the queried
+    /// type via the shared `TypeResolver` logic (no second computor) + the
+    /// existing `constInt` / `constFloat` const paths:
+    ///   - integer `.min`/`.max` → `constInt` (NL.1, via `integerLimitFor`);
+    ///   - float `.min`/`.max`/`.epsilon`/`.min_positive`/`.true_min`/`.inf`/
+    ///     `.nan` → `constFloat` (via `floatLimitFor`).
+    /// Returns null when the field is not a limit accessor, or the receiver is not
+    /// a builtin type (a user struct → ordinary field lowering reports
+    /// field-not-found). Two clean diagnostics (then a placeholder, so lowering
+    /// finishes and `hasErrors()` aborts the build):
+    ///   - a FLOAT-only accessor on an integer type (`s32.epsilon`, `u8.inf`);
+    ///   - any accessor on a builtin NON-numeric receiver
+    ///     (`bool`/`string`/`void`/`Any`/`noreturn`).
     fn lowerNumericLimit(self: *Lowering, fa: *const ast.FieldAccess, span: ast.Span) ?Ref {
         const name = switch (fa.object.data) {
             .identifier => |id| id.name,
             .type_expr => |te| te.name,
             else => return null,
         };
-        if (!std.mem.eql(u8, fa.field, "min") and !std.mem.eql(u8, fa.field, "max")) return null;
+        if (!TypeResolver.isLimitField(fa.field)) return null;
         const ty = TypeResolver.resolveBuiltinName(name, &self.module.types) orelse return null;
         if (TypeResolver.integerLimitFor(name, fa.field)) |value| {
             return self.builder.constInt(value, ty);
         }
-        // A builtin receiver that is not an integer: floats are NL.2 (fall
-        // through), every other builtin (bool/string/void/Any/noreturn) has no
-        // numeric limit.
-        if (ty == .f32 or ty == .f64) return null;
+        if (TypeResolver.floatLimitFor(name, fa.field)) |value| {
+            return self.builder.constFloat(value, ty);
+        }
+        // The field is a limit accessor, but it does not apply to this type.
         if (self.diagnostics) |d| {
-            d.addFmt(.err, span, "type '{s}' has no '.{s}' — numeric limits apply only to integer and float types", .{ name, fa.field });
+            if (TypeResolver.integerWidthSign(name) != null) {
+                // Integer receiver + a float-only accessor.
+                d.addFmt(.err, span, "type '{s}' has no '.{s}' — '.{s}' applies only to float types (f32/f64); integer types expose only '.min'/'.max'", .{ name, fa.field, fa.field });
+            } else {
+                // Non-numeric builtin receiver (bool/string/void/Any/noreturn).
+                d.addFmt(.err, span, "type '{s}' has no '.{s}' — numeric limits apply only to integer and float types", .{ name, fa.field });
+            }
         }
         return self.emitPlaceholder(fa.field);
     }
diff --git a/src/ir/type_resolver.zig b/src/ir/type_resolver.zig
index 3b673a1..aed74e5 100644
--- a/src/ir/type_resolver.zig
+++ b/src/ir/type_resolver.zig
@@ -137,6 +137,49 @@ pub const TypeResolver = struct {
         return integerLimitBits(wi, want_max);
     }
 
+    /// The full numeric-limit accessor field set: `.min`/`.max` (valid on int AND
+    /// float) plus the float-only `.epsilon`/`.min_positive`/`.true_min`/`.inf`/
+    /// `.nan`. THE single trigger for the `lowerNumericLimit` intercept — only a
+    /// field in this set is treated as a limit access; anything else falls through
+    /// to ordinary field lowering. Keeps the accessor name set in one place so the
+    /// intercept and `expr_typer` can't recognize different surfaces.
+    pub fn isLimitField(field: []const u8) bool {
+        const names = [_][]const u8{ "min", "max", "epsilon", "min_positive", "true_min", "inf", "nan" };
+        for (names) |n| if (std.mem.eql(u8, field, n)) return true;
+        return false;
+    }
+
+    /// `<FloatType>.<field>` → the limit as an `f64` value (the queried type is
+    /// `f32`/`f64`; every f32 limit is exactly representable in f64, so widening
+    /// is lossless and the caller pairs the value with the queried `TypeId` —
+    /// `builder.constFloat` narrows it back at emit), or null when `name` is not a
+    /// builtin float type or `field` is not a limit accessor. Values come straight
+    /// from `std.math` (`floatMax`/`floatEps`/`floatMin`/`floatTrueMin`/`inf`/`nan`):
+    ///   - `.min` = most-NEGATIVE finite (`-max`, NOT C's DBL_MIN)
+    ///   - `.max` = largest finite
+    ///   - `.epsilon` = ULP of 1.0 (`floatEps`; f64 = 2^-52, f32 = 2^-23)
+    ///   - `.min_positive` = smallest positive NORMAL (`floatMin`; = C DBL_MIN)
+    ///   - `.true_min` = smallest positive SUBNORMAL (`floatTrueMin`)
+    ///   - `.inf` = +infinity, `.nan` = a quiet NaN
+    /// THE single name+field → float fold, shared by the value path (lower.zig)
+    /// and `expr_typer` so they can't disagree.
+    pub fn floatLimitFor(name: []const u8, field: []const u8) ?f64 {
+        if (std.mem.eql(u8, name, "f64")) return floatLimitValue(f64, field);
+        if (std.mem.eql(u8, name, "f32")) return floatLimitValue(f32, field);
+        return null;
+    }
+
+    fn floatLimitValue(comptime T: type, field: []const u8) ?f64 {
+        if (std.mem.eql(u8, field, "min")) return -@as(f64, std.math.floatMax(T));
+        if (std.mem.eql(u8, field, "max")) return @as(f64, std.math.floatMax(T));
+        if (std.mem.eql(u8, field, "epsilon")) return @as(f64, std.math.floatEps(T));
+        if (std.mem.eql(u8, field, "min_positive")) return @as(f64, std.math.floatMin(T));
+        if (std.mem.eql(u8, field, "true_min")) return @as(f64, std.math.floatTrueMin(T));
+        if (std.mem.eql(u8, field, "inf")) return @as(f64, std.math.inf(T));
+        if (std.mem.eql(u8, field, "nan")) return @as(f64, std.math.nan(T));
+        return null;
+    }
+
     /// Single owner of structural AST-type-shape construction. Builds the
     /// shapes whose `TypeId` is fully determined by their node kind plus their
     /// element types resolved through `inner.resolveInner`: `*T`, `[*]T`, `[]T`,

From 463557990f62a3c694ac09e64da33315c135e593 Mon Sep 17 00:00:00 2001
From: agra <alex@swipelab.co>
Date: Thu, 4 Jun 2026 23:30:41 +0300
Subject: [PATCH 2/5] =?UTF-8?q?feat(lang):=20float=20numeric-limit=20acces?=
 =?UTF-8?q?sors=20=E2=80=94=20examples,=20unit=20tests,=20docs=20[NL.2]?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Finish NL.2 on top of the WIP compiler impl (2e9e4fe): f32/f64 expose
.min/.max plus the float-only .epsilon/.min_positive/.true_min/.inf/.nan,
folded via the shared lowerNumericLimit intercept + builder.constFloat.

- examples/0159: pins every f32/f64 accessor by untagged-union bit
  reinterpret against exact IEEE-754 hex (true_min read before any
  arithmetic — FTZ/DAZ), plus the defining-property checks
  ((1+eps)!=1 / (1+eps/2)==1, inf>max, min==-max, true_min<min_positive,
  true_min>0, nan!=nan).
- examples/0160: float-only accessor on an int (s32.epsilon/u8.inf/
  s64.true_min) and any accessor on a non-numeric type compile-error
  cleanly (exit 1, pinned stderr).
- type_resolver.test.zig: floatLimitFor bit-pattern + property tests for
  f32/f64, isLimitField coverage, null for non-float/non-limit fields.
- specs.md Numeric Limits: float accessors + the min=-max / min_positive=
  smallest-normal / epsilon=ULP-of-1.0 / true_min=smallest-subnormal
  clarifications, with the mandatory FTZ/DAZ flush-to-zero caveat.
  readme.md overview updated.
---
 examples/0159-types-float-numeric-limits.sx   | 89 +++++++++++++++++++
 .../0160-types-float-numeric-limits-errors.sx | 27 ++++++
 .../0159-types-float-numeric-limits.exit      |  1 +
 .../0159-types-float-numeric-limits.stderr    |  1 +
 .../0159-types-float-numeric-limits.stdout    | 19 ++++
 ...160-types-float-numeric-limits-errors.exit |  1 +
 ...0-types-float-numeric-limits-errors.stderr | 35 ++++++++
 ...0-types-float-numeric-limits-errors.stdout |  1 +
 readme.md                                     |  8 +-
 specs.md                                      | 58 ++++++++++++
 src/ir/type_resolver.test.zig                 | 71 +++++++++++++++
 11 files changed, 310 insertions(+), 1 deletion(-)
 create mode 100644 examples/0159-types-float-numeric-limits.sx
 create mode 100644 examples/0160-types-float-numeric-limits-errors.sx
 create mode 100644 examples/expected/0159-types-float-numeric-limits.exit
 create mode 100644 examples/expected/0159-types-float-numeric-limits.stderr
 create mode 100644 examples/expected/0159-types-float-numeric-limits.stdout
 create mode 100644 examples/expected/0160-types-float-numeric-limits-errors.exit
 create mode 100644 examples/expected/0160-types-float-numeric-limits-errors.stderr
 create mode 100644 examples/expected/0160-types-float-numeric-limits-errors.stdout

diff --git a/examples/0159-types-float-numeric-limits.sx b/examples/0159-types-float-numeric-limits.sx
new file mode 100644
index 0000000..f2753ab
--- /dev/null
+++ b/examples/0159-types-float-numeric-limits.sx
@@ -0,0 +1,89 @@
+// Float numeric-limit accessors: `f32`/`f64` expose `.min` / `.max` (sibling of
+// the integer `.min`/`.max`, NL.1) plus the float-only `.epsilon`,
+// `.min_positive`, `.true_min`, `.inf`, and `.nan`. Each folds, at compile time,
+// to a constant of the QUERIED float type via the same `lowerNumericLimit`
+// intercept as the integer case (`builder.constFloat` + the `std.math`
+// constants), driven by `TypeResolver.floatLimitFor`.
+//
+// The lexer has no exponent notation and the default float formatter is crude
+// (issue 0090), so these limits can be pinned NEITHER by literal comparison NOR
+// by printing. Every accessor is asserted instead by reinterpreting its bits
+// through an untagged union and comparing against the exact IEEE-754 hex
+// pattern — plus the defining-property checks that no other value could satisfy.
+//
+// Semantics (Agra-ruled, consistent with the integer accessors):
+//   .min          = most-NEGATIVE finite (= -max), NOT C's DBL_MIN
+//   .max          = largest finite
+//   .epsilon      = ULP of 1.0 (next f after 1.0 minus 1.0), NOT C#'s denormal Epsilon
+//   .min_positive = smallest positive NORMAL (= C DBL_MIN / Rust MIN_POSITIVE)
+//   .true_min     = smallest positive SUBNORMAL (next value above 0.0)
+//   .inf          = +infinity
+//   .nan          = a quiet NaN
+//
+// Regression (issue 0091): `f64.nan != f64.nan` is true — native float `!=`
+// lowers UNORDERED, so a NaN compares unequal to everything including itself.
+#import "modules/std.sx";
+
+// `bits` mirrors each float's raw IEEE-754 storage. f64 needs 64 bits, f32 32.
+// The f64 union uses a `u64` view so the all-ones-ish positive patterns read as
+// their true magnitude; the negative `f64.min` pattern (0xFFEF…) overflows the
+// i64 literal parser, so it is pinned by the `min == -max` property instead.
+Uf64 :: union { f: f64; bits: u64; }
+Uf32 :: union { f: f32; bits: u32; }
+
+main :: () -> s32 {
+    o : Uf64 = ---;
+
+    // Read `.true_min` (a subnormal) FIRST and through the union only — never via
+    // arithmetic. Under flush-to-zero / denormals-are-zero CPU modes a subnormal
+    // can flush to 0.0 on the first arithmetic op, so the bit reinterpret is the
+    // only reliable channel for it.
+    o.f = f64.true_min;
+    print("f64.true_min     {}\n", o.bits == 0x0000000000000001);  // true
+
+    o.f = f64.max;
+    print("f64.max          {}\n", o.bits == 0x7FEFFFFFFFFFFFFF);  // true
+    o.f = f64.epsilon;
+    print("f64.epsilon      {}\n", o.bits == 0x3CB0000000000000);  // true
+    o.f = f64.min_positive;
+    print("f64.min_positive {}\n", o.bits == 0x0010000000000000);  // true
+    o.f = f64.inf;
+    print("f64.inf          {}\n", o.bits == 0x7FF0000000000000);  // true
+
+    p : Uf32 = ---;
+    p.f = f32.true_min;
+    print("f32.true_min     {}\n", p.bits == 0x00000001);          // true
+    p.f = f32.max;
+    print("f32.max          {}\n", p.bits == 0x7F7FFFFF);          // true
+    p.f = f32.min;
+    print("f32.min          {}\n", p.bits == 0xFF7FFFFF);          // true
+    p.f = f32.epsilon;
+    print("f32.epsilon      {}\n", p.bits == 0x34000000);          // true
+    p.f = f32.min_positive;
+    print("f32.min_positive {}\n", p.bits == 0x00800000);          // true
+    p.f = f32.inf;
+    print("f32.inf          {}\n", p.bits == 0x7F800000);          // true
+
+    // Defining-property checks — true epsilon is the ULP of 1.0: adding it to 1.0
+    // changes the value, adding half of it does not (round-to-nearest-even).
+    print("(1+eps)!=1       {}\n", (1.0 + f64.epsilon) != 1.0);        // true
+    print("(1+eps/2)==1     {}\n", (1.0 + f64.epsilon/2.0) == 1.0);    // true
+    print("inf>max          {}\n", f64.inf > f64.max);                 // true
+    // f64.min = -max (the 0xFFEF… bit pattern overflows the i64 literal parser).
+    print("min==-max        {}\n", f64.min == -f64.max);              // true
+    print("true_min<min_pos {}\n", f64.true_min < f64.min_positive);  // true
+    print("true_min>0       {}\n", f64.true_min > 0.0);               // true
+    // Quiet NaN: unequal to everything, itself included (mantissa bits not pinned).
+    print("nan!=nan         {}\n", f64.nan != f64.nan);               // true
+
+    // Result carries the QUERIED type: each binding is declared with the float
+    // type and round-trips, so a mistyped fold (boxed as Any / wrong width) would
+    // not type-check here.
+    e64 : f64 = f64.epsilon;
+    e32 : f32 = f32.epsilon;
+    q : Uf64 = ---;  q.f = e64;
+    r : Uf32 = ---;  r.f = e32;
+    print("typed eps bits   {}\n", q.bits == 0x3CB0000000000000 and r.bits == 0x34000000); // true
+
+    return 0;
+}
diff --git a/examples/0160-types-float-numeric-limits-errors.sx b/examples/0160-types-float-numeric-limits-errors.sx
new file mode 100644
index 0000000..f008213
--- /dev/null
+++ b/examples/0160-types-float-numeric-limits-errors.sx
@@ -0,0 +1,27 @@
+// Cross-type rules for the numeric-limit accessors. `.min` / `.max` are valid on
+// BOTH integer and float types, but `.epsilon` / `.min_positive` / `.true_min` /
+// `.inf` / `.nan` are FLOAT-ONLY. Applying a float-only accessor to an INTEGER
+// type, or ANY accessor to a non-numeric type, is a clean compile error — never
+// a silent value, never the `.unresolved` sentinel reaching codegen.
+//
+//   - float-only accessor on an integer (`s32.epsilon`, `u8.inf`,
+//     `s64.true_min`) → a dedicated "applies only to float types" diagnostic
+//     from the accessor intercept, located at the access;
+//   - any accessor on a non-numeric builtin (`bool.nan`, `string.max`) → the
+//     "numeric limits apply only to integer and float types" diagnostic;
+//   - a user struct (`MyStruct.epsilon`) → the type name is not a builtin, so the
+//     intercept stays out and the existing field-not-found path reports it.
+// Each case is accurate and located at the access; the program exits non-zero.
+#import "modules/std.sx";
+
+MyStruct :: struct { a: s64; }
+
+main :: () -> s32 {
+    a := s32.epsilon;
+    b := u8.inf;
+    c := s64.true_min;
+    d := bool.nan;
+    e := string.max;
+    f := MyStruct.epsilon;
+    return 0;
+}
diff --git a/examples/expected/0159-types-float-numeric-limits.exit b/examples/expected/0159-types-float-numeric-limits.exit
new file mode 100644
index 0000000..573541a
--- /dev/null
+++ b/examples/expected/0159-types-float-numeric-limits.exit
@@ -0,0 +1 @@
+0
diff --git a/examples/expected/0159-types-float-numeric-limits.stderr b/examples/expected/0159-types-float-numeric-limits.stderr
new file mode 100644
index 0000000..8b13789
--- /dev/null
+++ b/examples/expected/0159-types-float-numeric-limits.stderr
@@ -0,0 +1 @@
+
diff --git a/examples/expected/0159-types-float-numeric-limits.stdout b/examples/expected/0159-types-float-numeric-limits.stdout
new file mode 100644
index 0000000..f562d64
--- /dev/null
+++ b/examples/expected/0159-types-float-numeric-limits.stdout
@@ -0,0 +1,19 @@
+f64.true_min     true
+f64.max          true
+f64.epsilon      true
+f64.min_positive true
+f64.inf          true
+f32.true_min     true
+f32.max          true
+f32.min          true
+f32.epsilon      true
+f32.min_positive true
+f32.inf          true
+(1+eps)!=1       true
+(1+eps/2)==1     true
+inf>max          true
+min==-max        true
+true_min<min_pos true
+true_min>0       true
+nan!=nan         true
+typed eps bits   true
diff --git a/examples/expected/0160-types-float-numeric-limits-errors.exit b/examples/expected/0160-types-float-numeric-limits-errors.exit
new file mode 100644
index 0000000..d00491f
--- /dev/null
+++ b/examples/expected/0160-types-float-numeric-limits-errors.exit
@@ -0,0 +1 @@
+1
diff --git a/examples/expected/0160-types-float-numeric-limits-errors.stderr b/examples/expected/0160-types-float-numeric-limits-errors.stderr
new file mode 100644
index 0000000..7f9b2df
--- /dev/null
+++ b/examples/expected/0160-types-float-numeric-limits-errors.stderr
@@ -0,0 +1,35 @@
+error: type 's32' has no '.epsilon' — '.epsilon' applies only to float types (f32/f64); integer types expose only '.min'/'.max'
+  --> examples/0160-types-float-numeric-limits-errors.sx:20:10
+   |
+20 |     a := s32.epsilon;
+   |          ^^^^^^^^^^^
+
+error: type 'u8' has no '.inf' — '.inf' applies only to float types (f32/f64); integer types expose only '.min'/'.max'
+  --> examples/0160-types-float-numeric-limits-errors.sx:21:10
+   |
+21 |     b := u8.inf;
+   |          ^^^^^^
+
+error: type 's64' has no '.true_min' — '.true_min' applies only to float types (f32/f64); integer types expose only '.min'/'.max'
+  --> examples/0160-types-float-numeric-limits-errors.sx:22:10
+   |
+22 |     c := s64.true_min;
+   |          ^^^^^^^^^^^^
+
+error: type 'bool' has no '.nan' — numeric limits apply only to integer and float types
+  --> examples/0160-types-float-numeric-limits-errors.sx:23:10
+   |
+23 |     d := bool.nan;
+   |          ^^^^^^^^
+
+error: type 'string' has no '.max' — numeric limits apply only to integer and float types
+  --> examples/0160-types-float-numeric-limits-errors.sx:24:10
+   |
+24 |     e := string.max;
+   |          ^^^^^^^^^^
+
+error: field 'epsilon' not found on type 'Any'
+  --> examples/0160-types-float-numeric-limits-errors.sx:25:10
+   |
+25 |     f := MyStruct.epsilon;
+   |          ^^^^^^^^^^^^^^^^
diff --git a/examples/expected/0160-types-float-numeric-limits-errors.stdout b/examples/expected/0160-types-float-numeric-limits-errors.stdout
new file mode 100644
index 0000000..8b13789
--- /dev/null
+++ b/examples/expected/0160-types-float-numeric-limits-errors.stdout
@@ -0,0 +1 @@
+
diff --git a/readme.md b/readme.md
index 0c71b8e..a6b70d9 100644
--- a/readme.md
+++ b/readme.md
@@ -90,7 +90,13 @@ Options:
 a compile-time constant of that type: `s64.max` → `9223372036854775807`,
 `u8.min` → `0`, `s3.max` → `3`. It works for every width `s1`..`s64` / `u1`..`u64`
 plus `usize`/`isize`, and is usable anywhere a constant of that type is — including
-array dimensions (`[u8.max]T` is a 255-element array).
+array dimensions (`[u8.max]T` is a 255-element array). The float types `f32`/`f64`
+expose `.min` / `.max` too (with `.min` = most-negative finite = `-max`, **not**
+C's `DBL_MIN`) plus the float-only `.epsilon` (ULP of 1.0, not C#'s denormal
+`Epsilon`), `.min_positive` (smallest normal = C `DBL_MIN`), `.true_min` (smallest
+subnormal — beware flush-to-zero CPU modes), `.inf`, and `.nan`. A float-only
+accessor on an integer (`s32.epsilon`), or any accessor on a non-numeric type, is
+a clean compile error. See `specs.md` → Numeric Limits.
 
 ### Declarations
 
diff --git a/specs.md b/specs.md
index b30853e..506c6ea 100644
--- a/specs.md
+++ b/specs.md
@@ -283,6 +283,64 @@ n      := u64.max;   // 18446744073709551615 (all-ones)
   `string`, a pointer, a `struct`, `void`, an `enum`) is a compile error, never
   a silent value.
 
+The **float** types `f32` and `f64` expose the same `.min` / `.max` plus a set of
+float-only accessors. Each folds, at compile time, to a constant of the queried
+float type (the same `lowerNumericLimit` intercept, via `builder.constFloat`):
+
+```sx
+hi  := f64.max;           // largest finite double
+lo  := f64.min;           // most-NEGATIVE finite = -max  (NOT C's DBL_MIN)
+eps := f64.epsilon;       // ULP of 1.0  (f64 = 2^-52, f32 = 2^-23)
+mp  := f64.min_positive;  // smallest positive NORMAL  (= C DBL_MIN / Rust MIN_POSITIVE)
+tm  := f64.true_min;      // smallest positive SUBNORMAL (next value above 0.0)
+pin := f64.inf;           // +infinity
+qn  := f64.nan;           // a quiet NaN
+```
+
+- **Receiver.** `f32` or `f64`.
+- **Shared with integers.** `.min` / `.max` are valid on BOTH integer and float
+  types. `.min` is the most-NEGATIVE finite value, i.e. `-max` — consistent with
+  the integer `.min`, and deliberately **NOT** C's `DBL_MIN`/`FLT_MIN` (which is
+  the smallest positive normal; that is `.min_positive` here).
+- **Float-only accessors.**
+  - `.epsilon` — the ULP of `1.0`: the gap between `1.0` and the next
+    representable value (`f64 = 2^-52 ≈ 2.22e-16`, `f32 = 2^-23`). This is the
+    **machine epsilon** used for relative-tolerance comparisons, **NOT** C#'s
+    `Double.Epsilon` (which is the smallest denormal — that is `.true_min` here).
+    Defining property: `1.0 + epsilon != 1.0` while `1.0 + epsilon/2.0 == 1.0`.
+  - `.min_positive` — the smallest positive **NORMAL** value (`f64 = 2^-1022`,
+    `f32 = 2^-126`). Equals C's `DBL_MIN` / Rust's `MIN_POSITIVE`.
+  - `.true_min` — the smallest positive **SUBNORMAL**: the next value above `0.0`
+    (`f64` bits `0x0000000000000001 = 2^-1074`, `f32` bits `0x00000001 = 2^-149`).
+    Named `true_min` (after Zig's `floatTrueMin`) to avoid the Java/Go/JS
+    `MIN_VALUE` footgun, where a bare `MIN_VALUE` names the smallest *subnormal*
+    yet reads like the most-negative value.
+    - **FTZ/DAZ caveat.** Subnormals are exactly the values that vanish under
+      flush-to-zero (FTZ) / denormals-are-zero (DAZ) CPU modes. If such a mode is
+      active, a loaded `.true_min` can flush to `0.0` on the **first arithmetic
+      operation** that touches it. The folded constant always carries the exact
+      subnormal bit pattern; read or store it through a bit reinterpret *before*
+      any arithmetic if you need the true value to survive. Numerical-library
+      authors who toggle FTZ/DAZ should not be surprised when `true_min * 1.0`
+      reads back as `0.0`.
+  - `.inf` — positive infinity (`inf > max`).
+  - `.nan` — a quiet NaN. The exact mantissa bits are not pinned; the only
+    guaranteed property is that it is unequal to everything, itself included
+    (`nan != nan` is `true` — native float `!=` lowers unordered, issue 0091).
+- **Float-only on an integer is an error.** `.epsilon` / `.min_positive` /
+  `.true_min` / `.inf` / `.nan` applied to an integer type (`s32.epsilon`,
+  `u8.inf`, `s64.true_min`) is a clean compile error — integer types expose only
+  `.min` / `.max`.
+- **Pinning the values.** The lexer has no exponent notation and the default
+  float formatter is crude (issue 0090), so float limits can be asserted neither
+  by literal comparison nor by printing. Reinterpret the bits through an untagged
+  union (`union { f: f64; bits: u64 }`) and compare against the exact IEEE-754
+  pattern — `f64.max = 0x7FEFFFFFFFFFFFFF`, `min = 0xFFEFFFFFFFFFFFFF`,
+  `epsilon = 0x3CB0000000000000`, `min_positive = 0x0010000000000000`,
+  `true_min = 0x0000000000000001`, `inf = 0x7FF0000000000000`; the `f32` set is
+  `0x7F7FFFFF` / `0xFF7FFFFF` / `0x34000000` / `0x00800000` / `0x00000001` /
+  `0x7F800000`.
+
 ### Enum Types
 User-defined sum types with named variants. Variants may optionally carry typed data (tagged unions). Internally, payload-less enums are represented as `i64` (variant index). Enums with payloads are represented as `{ i64, [max_payload_size x i8] }` (tag + data).
 
diff --git a/src/ir/type_resolver.test.zig b/src/ir/type_resolver.test.zig
index 70e895d..b81945f 100644
--- a/src/ir/type_resolver.test.zig
+++ b/src/ir/type_resolver.test.zig
@@ -246,3 +246,74 @@ test "TypeResolver.integerLimitFor: null for non-integer receivers and non-limit
     try std.testing.expect(TypeResolver.integerLimitFor("s64", "len") == null);
     try std.testing.expect(TypeResolver.integerLimitFor("u8", "epsilon") == null);
 }
+
+test "TypeResolver.isLimitField: the accessor set, nothing else" {
+    // The full numeric-limit surface — int .min/.max plus the float-only ones.
+    for ([_][]const u8{ "min", "max", "epsilon", "min_positive", "true_min", "inf", "nan" }) |f| {
+        try std.testing.expect(TypeResolver.isLimitField(f));
+    }
+    // Ordinary fields / near-misses are not limit accessors.
+    try std.testing.expect(!TypeResolver.isLimitField("len"));
+    try std.testing.expect(!TypeResolver.isLimitField("ptr"));
+    try std.testing.expect(!TypeResolver.isLimitField("maximum"));
+    try std.testing.expect(!TypeResolver.isLimitField("Min"));
+    try std.testing.expect(!TypeResolver.isLimitField(""));
+}
+
+test "TypeResolver.floatLimitFor: pinned f64 bit patterns" {
+    const L = struct {
+        fn bits(name: []const u8, field: []const u8) u64 {
+            return @bitCast(TypeResolver.floatLimitFor(name, field).?);
+        }
+    };
+    // Exact IEEE-754 double bit patterns (the same values the example pins via
+    // a runtime bit reinterpret).
+    try std.testing.expectEqual(@as(u64, 0x7FEFFFFFFFFFFFFF), L.bits("f64", "max"));
+    try std.testing.expectEqual(@as(u64, 0xFFEFFFFFFFFFFFFF), L.bits("f64", "min")); // -max
+    try std.testing.expectEqual(@as(u64, 0x3CB0000000000000), L.bits("f64", "epsilon"));
+    try std.testing.expectEqual(@as(u64, 0x0010000000000000), L.bits("f64", "min_positive"));
+    try std.testing.expectEqual(@as(u64, 0x0000000000000001), L.bits("f64", "true_min"));
+    try std.testing.expectEqual(@as(u64, 0x7FF0000000000000), L.bits("f64", "inf"));
+    // .min = -max (NOT C's DBL_MIN, which is min_positive); the ordering holds.
+    const v = struct {
+        fn f(name: []const u8, field: []const u8) f64 {
+            return TypeResolver.floatLimitFor(name, field).?;
+        }
+    };
+    try std.testing.expectEqual(v.f("f64", "min"), -v.f("f64", "max"));
+    try std.testing.expect(v.f("f64", "true_min") < v.f("f64", "min_positive"));
+    try std.testing.expect(v.f("f64", "true_min") > 0.0);
+    try std.testing.expect(std.math.isInf(v.f("f64", "inf")));
+    // Quiet NaN: unequal to itself; exact mantissa bits intentionally not pinned.
+    try std.testing.expect(std.math.isNan(v.f("f64", "nan")));
+}
+
+test "TypeResolver.floatLimitFor: pinned f32 bit patterns (widened value narrows losslessly)" {
+    const L = struct {
+        // floatLimitFor widens every f32 limit to f64; narrowing back is lossless
+        // (the codegen path does the same via constFloat at the queried width).
+        fn bits(name: []const u8, field: []const u8) u32 {
+            return @bitCast(@as(f32, @floatCast(TypeResolver.floatLimitFor(name, field).?)));
+        }
+    };
+    try std.testing.expectEqual(@as(u32, 0x7F7FFFFF), L.bits("f32", "max"));
+    try std.testing.expectEqual(@as(u32, 0xFF7FFFFF), L.bits("f32", "min")); // -max
+    try std.testing.expectEqual(@as(u32, 0x34000000), L.bits("f32", "epsilon"));
+    try std.testing.expectEqual(@as(u32, 0x00800000), L.bits("f32", "min_positive"));
+    try std.testing.expectEqual(@as(u32, 0x00000001), L.bits("f32", "true_min"));
+    try std.testing.expectEqual(@as(u32, 0x7F800000), L.bits("f32", "inf"));
+    const nan_v: f32 = @floatCast(TypeResolver.floatLimitFor("f32", "nan").?);
+    try std.testing.expect(std.math.isNan(nan_v));
+}
+
+test "TypeResolver.floatLimitFor: null for non-float receivers and non-limit fields" {
+    // Integer / non-numeric / user names are not float-limit folds.
+    try std.testing.expect(TypeResolver.floatLimitFor("s32", "epsilon") == null);
+    try std.testing.expect(TypeResolver.floatLimitFor("u64", "max") == null);
+    try std.testing.expect(TypeResolver.floatLimitFor("usize", "min") == null);
+    try std.testing.expect(TypeResolver.floatLimitFor("bool", "nan") == null);
+    try std.testing.expect(TypeResolver.floatLimitFor("MyStruct", "inf") == null);
+    // A builtin float with a non-limit field is not a fold here.
+    try std.testing.expect(TypeResolver.floatLimitFor("f64", "len") == null);
+    try std.testing.expect(TypeResolver.floatLimitFor("f32", "ptr") == null);
+}

From b0cc22a8c00006a66ef6ddfb258a4fec4db1be97 Mon Sep 17 00:00:00 2001
From: agra <alex@swipelab.co>
Date: Thu, 4 Jun 2026 23:59:11 +0300
Subject: [PATCH 3/5] fix(lang): numeric-limit intercept no longer shadows raw
 value bindings [NL.2]
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

The numeric-limit accessor intercept (NL.1 integer `.min`/`.max`, NL.2 float
`.epsilon`/`.min_positive`/`.true_min`/`.inf`/`.nan`) treated ANY receiver
whose text matched a builtin numeric type name as a TYPE receiver, without
first checking for an in-scope VALUE binding. An F0.6 backtick raw identifier
(`` `f64 := … ``) binds a local under the stripped name `f64`; field access on
it (`` `f64.epsilon ``) parses as an `.identifier` receiver, which the intercept
silently folded to the type's numeric limit — a silent-wrong-value bug
(issue 0092).

Fix: for `.identifier` receivers, prefer an in-scope value binding
(`Scope.lookup`) over the fold — defer to ordinary field lowering when the
identifier resolves to a value. `.type_expr` receivers are unambiguous types
and are never shadowed, so a bare `f64.epsilon`/`s32.max` still folds even in a
scope where `` `f64 `` is bound (the parser classifies a bare builtin name as a
`.type_expr`). Mirrored in expr_typer.zig so inference matches lowering
(avoids the issue-0083 two-resolver desync). Float-only-on-int and
non-numeric-receiver errors are unchanged.

- src/ir/lower.zig: value-binding guard in lowerNumericLimit.
- src/ir/expr_typer.zig: same guard in the numeric-limit inference arm.
- src/ir/expr_typer.test.zig: unit test pinning the two-resolver agreement.
- examples/0161-types-numeric-limit-value-shadow.sx: regression — raw
  `` `f64 ``/`` `s32 ``/`` `u8 `` value reads coexisting with bare folds.
- issues/0092: RESOLVED banner.
- specs.md / readme.md: receiver-vs-shadowing-value-binding note.
---
 .../0161-types-numeric-limit-value-shadow.sx  | 50 ++++++++++++
 ...0161-types-numeric-limit-value-shadow.exit |  1 +
 ...61-types-numeric-limit-value-shadow.stderr |  1 +
 ...61-types-numeric-limit-value-shadow.stdout |  9 +++
 ...raw-reserved-value-numeric-limit-shadow.md | 77 +++++++++++++++++++
 readme.md                                     |  5 +-
 specs.md                                      |  8 ++
 src/ir/expr_typer.test.zig                    | 45 +++++++++++
 src/ir/expr_typer.zig                         | 11 ++-
 src/ir/lower.zig                              | 13 ++++
 10 files changed, 217 insertions(+), 3 deletions(-)
 create mode 100644 examples/0161-types-numeric-limit-value-shadow.sx
 create mode 100644 examples/expected/0161-types-numeric-limit-value-shadow.exit
 create mode 100644 examples/expected/0161-types-numeric-limit-value-shadow.stderr
 create mode 100644 examples/expected/0161-types-numeric-limit-value-shadow.stdout
 create mode 100644 issues/0092-raw-reserved-value-numeric-limit-shadow.md

diff --git a/examples/0161-types-numeric-limit-value-shadow.sx b/examples/0161-types-numeric-limit-value-shadow.sx
new file mode 100644
index 0000000..8e7dc20
--- /dev/null
+++ b/examples/0161-types-numeric-limit-value-shadow.sx
@@ -0,0 +1,50 @@
+// Numeric-limit accessor vs. a raw value binding that shadows a builtin type
+// name. A backtick raw identifier (F0.6) can legitimately bind a value whose
+// spelling is a reserved numeric type name (`` `f64 ``, `` `s32 ``, `` `u8 ``).
+// Field access on such a value is an ORDINARY field read — the numeric-limit
+// intercept (NL.1 integer `.min`/`.max`, NL.2 float `.epsilon`/… ) must NOT
+// hijack it. An adjacent BARE `f64.epsilon` / `s32.max` / `u8.max` — which the
+// parser classifies as a type receiver, not the raw value — STILL folds to the
+// numeric limit. Both behaviors coexist in one scope: the raw receiver reads
+// the value, the bare receiver folds the limit.
+//
+// Regression (issue 0092): the intercept previously treated any identifier
+// whose text matched a builtin numeric type name as a TYPE receiver, silently
+// shadowing the in-scope value binding (`` `f64.epsilon `` folded to 2^-52,
+// `` `s32.max `` folded to 2147483647 — a silent wrong value).
+#import "modules/std.sx";
+
+FBox :: struct { epsilon: s64; max: s64; min_positive: s64; }
+IBox :: struct { max: s64; min: s64; }
+UBox :: struct { max: s64; }
+
+main :: () -> s32 {
+    // Raw value bindings whose spelling shadows a builtin numeric type name.
+    `f64 := FBox.{ epsilon = 11, max = 22, min_positive = 33 };
+    `s32 := IBox.{ max = 78, min = -78 };
+    `u8  := UBox.{ max = 7 };
+
+    // Raw receiver → ordinary field READ (the value), never the numeric limit.
+    print("val f64: epsilon={} max={} min_positive={}\n",
+          `f64.epsilon, `f64.max, `f64.min_positive);   // 11 22 33
+    print("val s32: max={} min={}\n", `s32.max, `s32.min); // 78 -78
+    print("val u8: max={}\n", `u8.max);                    // 7
+
+    // The value-field read carries the field type (s64 here): round-trips
+    // through a typed binding, so a mistyped/boxed read would not type-check.
+    e : s64 = `f64.epsilon;
+    print("typed val e={}\n", e);                          // 11
+
+    // Bare receiver (a type receiver, NOT the raw value) → STILL folds to the
+    // numeric limit, even though `s32`/`u8`/`f64` are bound in this same scope.
+    print("lim s32.max={} s32.min={}\n", s32.max, s32.min); // 2147483647 -2147483648
+    print("lim u8.max={}\n", u8.max);                       // 255
+
+    // Bare float accessors still fold; the formatter is crude (issue 0090), so
+    // pin the values by their defining properties rather than by printing.
+    print("lim (1.0+f64.epsilon)!=1.0: {}\n", (1.0 + f64.epsilon) != 1.0); // true
+    print("lim f64.inf > f64.max: {}\n", f64.inf > f64.max);               // true
+    print("lim f64.min == -f64.max: {}\n", f64.min == -f64.max);           // true
+
+    return 0;
+}
diff --git a/examples/expected/0161-types-numeric-limit-value-shadow.exit b/examples/expected/0161-types-numeric-limit-value-shadow.exit
new file mode 100644
index 0000000..573541a
--- /dev/null
+++ b/examples/expected/0161-types-numeric-limit-value-shadow.exit
@@ -0,0 +1 @@
+0
diff --git a/examples/expected/0161-types-numeric-limit-value-shadow.stderr b/examples/expected/0161-types-numeric-limit-value-shadow.stderr
new file mode 100644
index 0000000..8b13789
--- /dev/null
+++ b/examples/expected/0161-types-numeric-limit-value-shadow.stderr
@@ -0,0 +1 @@
+
diff --git a/examples/expected/0161-types-numeric-limit-value-shadow.stdout b/examples/expected/0161-types-numeric-limit-value-shadow.stdout
new file mode 100644
index 0000000..e54c05a
--- /dev/null
+++ b/examples/expected/0161-types-numeric-limit-value-shadow.stdout
@@ -0,0 +1,9 @@
+val f64: epsilon=11 max=22 min_positive=33
+val s32: max=78 min=-78
+val u8: max=7
+typed val e=11
+lim s32.max=2147483647 s32.min=-2147483648
+lim u8.max=255
+lim (1.0+f64.epsilon)!=1.0: true
+lim f64.inf > f64.max: true
+lim f64.min == -f64.max: true
diff --git a/issues/0092-raw-reserved-value-numeric-limit-shadow.md b/issues/0092-raw-reserved-value-numeric-limit-shadow.md
new file mode 100644
index 0000000..0c1520d
--- /dev/null
+++ b/issues/0092-raw-reserved-value-numeric-limit-shadow.md
@@ -0,0 +1,77 @@
+> **RESOLVED** (NL.2 attempt 3). Root cause: the numeric-limit accessor
+> intercept treated ANY receiver whose text matched a builtin numeric type
+> name as a TYPE receiver, without first checking whether that identifier
+> resolved to an in-scope VALUE binding. An F0.6 backtick raw identifier
+> (`` `f64 := … ``) binds a local under the stripped name `f64`; field access
+> on it (`` `f64.epsilon ``) parses as an `.identifier` receiver, which the
+> intercept silently folded to the type's numeric limit — a silent-wrong-value
+> bug.
+>
+> Fix (value-binding precedence for `.identifier` receivers; `.type_expr`
+> receivers are unambiguous types and never shadowed):
+> - `src/ir/lower.zig` — `lowerNumericLimit`: after confirming the receiver is
+>   a builtin numeric type name and the field is a limit accessor, return null
+>   (defer to ordinary field lowering) when `fa.object` is an `.identifier`
+>   that `Scope.lookup` resolves to a value binding.
+> - `src/ir/expr_typer.zig` — numeric-limit inference arm: mirror the same
+>   guard so inferred types match lowering (avoids the issue-0083 two-resolver
+>   desync).
+>
+> Bare `f64.epsilon` / `s32.max` (no shadowing binding) still fold — the parser
+> classifies a bare builtin name as a `.type_expr` (parser.zig:2743), so the
+> bare receiver is never value-shadowed even in a scope where `` `f64 `` is
+> bound. Float-only-on-int and non-numeric-receiver errors are unchanged.
+>
+> Regression: `examples/0161-types-numeric-limit-value-shadow.sx` (raw
+> `` `f64 ``/`` `s32 ``/`` `u8 `` value reads coexisting with bare folds) +
+> unit test in `src/ir/expr_typer.test.zig`. NL.1 (`examples/0148`) / NL.2
+> (`examples/0159`, `examples/0160`) unregressed.
+
+# 0092 — numeric-limit intercept hijacks raw reserved-spelled value receivers
+
+## Symptom
+
+Field access on a raw reserved-spelled value binding is interpreted as a builtin
+type numeric-limit access instead of an ordinary value field access. Observed:
+the repro prints `0.000000 2147483647` (`f64.epsilon` / `s32.max`). Expected:
+it prints `12 78` from the `Box` fields.
+
+## Reproduction
+
+```sx
+#import "modules/std.sx";
+
+Box :: struct { epsilon: s64; max: s64; }
+
+main :: () -> s32 {
+    `f64 := Box.{ epsilon = 12, max = 34 };
+    `s32 := Box.{ epsilon = 56, max = 78 };
+    print("{} {}\n", `f64.epsilon, `s32.max);
+    return 0;
+}
+```
+
+## Investigation prompt
+
+Investigate issue 0092: raw reserved-spelled value receivers are being captured
+by the numeric-limit accessor intercept. In `src/ir/lower.zig`, start at
+`Lowering.lowerFieldAccess` and `Lowering.lowerNumericLimit` (currently around
+`lower.zig:4826` / `lower.zig:4923`). The intercept treats any
+`.identifier`/`.type_expr` receiver whose text is a builtin type name as a type
+receiver, without first checking whether the identifier resolves to a value
+binding in the current lexical scope. Mirror the fix in `src/ir/expr_typer.zig`
+around the numeric-limit inference arm so inferred types match lowering.
+
+Likely fix: for `.identifier` receivers, prefer an in-scope value binding
+(`Scope.lookup`) over the builtin-type numeric-limit intercept; keep the
+intercept for actual type receivers (`.type_expr`, and bare reserved integer
+names with no value binding). If raw provenance is available in the AST, using it
+to disambiguate is also acceptable, but the observable rule must be that
+`` `f64.epsilon `` reads the value field when `` `f64 `` is a value binding,
+while bare `f64.epsilon` / `s32.max` still fold as numeric limits.
+
+Verification: pin a regression test from the repro above. It should print
+`12 78`. Also verify the existing numeric-limit examples still pass:
+`examples/0148-types-int-numeric-limits.sx`,
+`examples/0159-types-float-numeric-limits.sx`, and the negative diagnostics in
+`examples/0160-types-float-numeric-limits-errors.sx`.
diff --git a/readme.md b/readme.md
index a6b70d9..ef80ca0 100644
--- a/readme.md
+++ b/readme.md
@@ -96,7 +96,10 @@ C's `DBL_MIN`) plus the float-only `.epsilon` (ULP of 1.0, not C#'s denormal
 `Epsilon`), `.min_positive` (smallest normal = C `DBL_MIN`), `.true_min` (smallest
 subnormal — beware flush-to-zero CPU modes), `.inf`, and `.nan`. A float-only
 accessor on an integer (`s32.epsilon`), or any accessor on a non-numeric type, is
-a clean compile error. See `specs.md` → Numeric Limits.
+a clean compile error. The fold applies only to a bare type-name receiver: a raw
+identifier that binds a value shadowing a type name (`` `f64 := … `` then
+`` `f64.epsilon ``) reads the value's field, not the limit. See `specs.md` →
+Numeric Limits.
 
 ### Declarations
 
diff --git a/specs.md b/specs.md
index 506c6ea..2056a43 100644
--- a/specs.md
+++ b/specs.md
@@ -340,6 +340,14 @@ qn  := f64.nan;           // a quiet NaN
   `true_min = 0x0000000000000001`, `inf = 0x7FF0000000000000`; the `f32` set is
   `0x7F7FFFFF` / `0xFF7FFFFF` / `0x34000000` / `0x00800000` / `0x00000001` /
   `0x7F800000`.
+- **Type receiver vs. a shadowing value binding.** A numeric-limit access folds
+  only when the receiver is a builtin numeric **type name** (`f64.epsilon`,
+  `s32.max`, `u8.max`). A backtick raw identifier that binds a *value* whose
+  spelling shadows a type name (`` `f64 := … ``, F0.6) is an ordinary value:
+  `` `f64.epsilon `` reads that value's `epsilon` field — it does **not** fold to
+  the limit. The two never collide, even in the same scope — a bare builtin name
+  in expression position is always a type, and only the raw `` `…` `` spelling
+  can bind a value under it.
 
 ### Enum Types
 User-defined sum types with named variants. Variants may optionally carry typed data (tagged unions). Internally, payload-less enums are represented as `i64` (variant index). Enums with payloads are represented as `{ i64, [max_payload_size x i8] }` (tag + data).
diff --git a/src/ir/expr_typer.test.zig b/src/ir/expr_typer.test.zig
index 46db0f0..70e0435 100644
--- a/src/ir/expr_typer.test.zig
+++ b/src/ir/expr_typer.test.zig
@@ -9,7 +9,9 @@ const Node = ast.Node;
 
 const ir_mod = @import("ir.zig");
 const TypeId = ir_mod.TypeId;
+const Ref = ir_mod.Ref;
 const Lowering = ir_mod.Lowering;
+const Scope = @import("lower.zig").Scope;
 
 fn node(data: ast.Node.Data) Node {
     return .{ .span = .{ .start = 0, .end = 0 }, .data = data };
@@ -73,3 +75,46 @@ test "expr_typer: deref of a non-pointer is unresolved" {
     var deref_n = node(.{ .deref_expr = .{ .operand = &i } });
     try std.testing.expectEqual(TypeId.unresolved, l.inferExprType(&deref_n));
 }
+
+// issue 0092: a raw `` `f64 `` value binding shadows the builtin numeric type
+// name — `` `f64.epsilon `` (an `.identifier` receiver) must type as the value's
+// field, NOT the float numeric-limit fold. A bare `f64.epsilon` (a `.type_expr`
+// receiver, never shadowed) still folds to the queried float type. This pins the
+// two-resolver agreement: expr_typer's inference must match lowerNumericLimit.
+test "expr_typer: raw value binding shadows numeric-limit, bare type still folds" {
+    // Arena-backed (like calls.test.zig's scope tests): the `.identifier`
+    // field-access path interns a `obj.field` probe string into `l.alloc`,
+    // which the production compiler owns via an arena — an explicit free would
+    // not match the real lifetime.
+    var arena = std.heap.ArenaAllocator.init(std.testing.allocator);
+    defer arena.deinit();
+    const alloc = arena.allocator();
+    var module = ir_mod.Module.init(alloc);
+    defer module.deinit();
+    var l = Lowering.init(&module);
+
+    // A struct `Box { epsilon: s64 }` bound to the raw name `` `f64 ``.
+    const box_fields = [_]ir_mod.types.TypeInfo.StructInfo.Field{
+        .{ .name = module.types.internString("epsilon"), .ty = .s64 },
+    };
+    const box_ty = module.types.intern(.{ .@"struct" = .{
+        .name = module.types.internString("Box"),
+        .fields = &box_fields,
+    } });
+
+    var scope = Scope.init(alloc, null);
+    defer scope.deinit();
+    l.scope = &scope;
+    scope.put("f64", .{ .ref = Ref.none, .ty = box_ty, .is_alloca = false });
+
+    // `` `f64.epsilon `` — identifier receiver resolving to the value binding →
+    // ordinary field read, types as s64 (the field), not f64.
+    var raw_recv = node(.{ .identifier = .{ .name = "f64", .is_raw = true } });
+    var raw_fa = node(.{ .field_access = .{ .object = &raw_recv, .field = "epsilon" } });
+    try std.testing.expectEqual(TypeId.s64, l.inferExprType(&raw_fa));
+
+    // bare `f64.epsilon` — type_expr receiver, never shadowed → folds to f64.
+    var type_recv = node(.{ .type_expr = .{ .name = "f64" } });
+    var type_fa = node(.{ .field_access = .{ .object = &type_recv, .field = "epsilon" } });
+    try std.testing.expectEqual(TypeId.f64, l.inferExprType(&type_fa));
+}
diff --git a/src/ir/expr_typer.zig b/src/ir/expr_typer.zig
index 1b8d8a9..45692d9 100644
--- a/src/ir/expr_typer.zig
+++ b/src/ir/expr_typer.zig
@@ -140,8 +140,15 @@ pub const ExprTyper = struct {
                         else => null,
                     };
                     if (type_name) |tn| {
-                        if (TypeResolver.integerLimitFor(tn, fa.field) != null or
-                            TypeResolver.floatLimitFor(tn, fa.field) != null)
+                        // Skip the fold when a raw value binding shadows the
+                        // builtin type name (`` `f64 := … ``) — mirrors the
+                        // lowerNumericLimit guard so inference matches lowering
+                        // (issue 0092). A `.type_expr` receiver is never shadowed.
+                        const shadowed = fa.object.data == .identifier and
+                            (if (self.l.scope) |scope| (scope.lookup(tn) != null) else false);
+                        if (!shadowed and
+                            (TypeResolver.integerLimitFor(tn, fa.field) != null or
+                                TypeResolver.floatLimitFor(tn, fa.field) != null))
                         {
                             if (TypeResolver.resolveBuiltinName(tn, &self.l.module.types)) |t| return t;
                         }
diff --git a/src/ir/lower.zig b/src/ir/lower.zig
index 81611c2..bfd37b1 100644
--- a/src/ir/lower.zig
+++ b/src/ir/lower.zig
@@ -4928,6 +4928,19 @@ pub const Lowering = struct {
         };
         if (!TypeResolver.isLimitField(fa.field)) return null;
         const ty = TypeResolver.resolveBuiltinName(name, &self.module.types) orelse return null;
+
+        // A backtick raw identifier (F0.6) can bind a value whose spelling
+        // shadows a builtin type name (`` `f64 := … ``). Field access on that
+        // value is an ordinary field read, not a numeric-limit fold — defer to
+        // the normal field-access path when the receiver identifier resolves to
+        // an in-scope value binding (issue 0092). A `.type_expr` receiver is
+        // unambiguously a type and can never be value-shadowed.
+        if (fa.object.data == .identifier) {
+            if (self.scope) |scope| {
+                if (scope.lookup(name) != null) return null;
+            }
+        }
+
         if (TypeResolver.integerLimitFor(name, fa.field)) |value| {
             return self.builder.constInt(value, ty);
         }

From 6478ccbe3c37aac1a0760d391b7163493b60c335 Mon Sep 17 00:00:00 2001
From: agra <alex@swipelab.co>
Date: Fri, 5 Jun 2026 00:21:32 +0300
Subject: [PATCH 4/5] fix(lang): numeric-limit shadow guard covers all 3 value
 sources [NL.2]
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

The issue-0092 fix guarded the numeric-limit accessor intercept against
raw value shadowing using only lexical Scope.lookup. The ordinary
identifier field-access path resolves a value through THREE sources
(scope / program_index.global_names / program_index.module_const_map),
so a backtick raw identifier bound at module scope — a global
`` `f64 := Box.{…} `` or a module constant `` `f64 :: Box.{…} `` — still
folded `` `f64.epsilon `` to the numeric limit instead of reading the
value's field (issue 0093, plus the module-const variant: same root
cause, same fix).

Fix: a single shared helper Lowering.identifierBindsValue(name) that
returns true when the name resolves through scope OR global_names OR
module_const_map. Used in BOTH lowerNumericLimit (lower.zig) and the
numeric-limit inference arm (expr_typer.zig) so the two resolvers can't
desync (issue-0083 class). A bare `f64.epsilon` / `s32.max` (a
.type_expr receiver) still folds even when a raw value of the same
spelling is bound — the bare receiver is never value-shadowed.

- examples/0161: extended to exercise all three binding kinds — a
  GLOBAL `` `f32 ``, a MODULE-CONST `` `s16 ``, and LOCAL
  `` `f64 ``/`` `s32 ``/`` `u8 `` — each reading its field while the
  bare spelling still folds.
- src/ir/expr_typer.test.zig: unit test pinning the global +
  module-const sources of the shared guard.
- issues/0093: RESOLVED banner (3-source root cause + fix, module-const
  variant folded in).
- specs.md / readme.md: numeric-limit shadow note now source-agnostic
  (local / global / module-const).
---
 .../0161-types-numeric-limit-value-shadow.sx  | 50 ++++++++---
 ...61-types-numeric-limit-value-shadow.stdout | 10 ++-
 ...3-global-raw-value-numeric-limit-shadow.md | 84 +++++++++++++++++++
 readme.md                                     |  4 +-
 specs.md                                      | 10 ++-
 src/ir/expr_typer.test.zig                    | 59 +++++++++++++
 src/ir/expr_typer.zig                         |  6 +-
 src/ir/lower.zig                              | 32 +++++--
 8 files changed, 225 insertions(+), 30 deletions(-)
 create mode 100644 issues/0093-global-raw-value-numeric-limit-shadow.md

diff --git a/examples/0161-types-numeric-limit-value-shadow.sx b/examples/0161-types-numeric-limit-value-shadow.sx
index 8e7dc20..dc6f1da 100644
--- a/examples/0161-types-numeric-limit-value-shadow.sx
+++ b/examples/0161-types-numeric-limit-value-shadow.sx
@@ -5,30 +5,52 @@
 // intercept (NL.1 integer `.min`/`.max`, NL.2 float `.epsilon`/… ) must NOT
 // hijack it. An adjacent BARE `f64.epsilon` / `s32.max` / `u8.max` — which the
 // parser classifies as a type receiver, not the raw value — STILL folds to the
-// numeric limit. Both behaviors coexist in one scope: the raw receiver reads
-// the value, the bare receiver folds the limit.
+// numeric limit. Both behaviors coexist: the raw receiver reads the value, the
+// bare receiver folds the limit.
 //
-// Regression (issue 0092): the intercept previously treated any identifier
-// whose text matched a builtin numeric type name as a TYPE receiver, silently
-// shadowing the in-scope value binding (`` `f64.epsilon `` folded to 2^-52,
-// `` `s32.max `` folded to 2147483647 — a silent wrong value).
+// A raw value binding can reach the intercept through THREE sources, exactly
+// mirroring the ordinary identifier field-access path (scope / globals / module
+// consts). This example exercises all three: a GLOBAL `` `f32 ``, a MODULE-CONST
+// `` `s16 ``, and LOCAL `` `f64 ``/`` `s32 ``/`` `u8 `` — each reads its field,
+// and the bare spelling of each STILL folds.
+//
+// Regression (issues 0092 local, 0093 global + module-const): the intercept
+// previously treated any identifier whose text matched a builtin numeric type
+// name as a TYPE receiver, silently shadowing the in-scope value binding
+// (`` `f64.epsilon `` folded to 2^-52, `` `s32.max `` folded to 2147483647 —
+// a silent wrong value). The attempt-3 fix guarded only lexical scope, so
+// GLOBAL and MODULE-CONST raw bindings still folded (issue 0093).
 #import "modules/std.sx";
 
 FBox :: struct { epsilon: s64; max: s64; min_positive: s64; }
 IBox :: struct { max: s64; min: s64; }
 UBox :: struct { max: s64; }
 
+// GLOBAL raw value binding whose spelling shadows the builtin `f32`. Reachable
+// via `program_index.global_names`, not lexical scope (issue 0093).
+`f32 := FBox.{ epsilon = 44, max = 55, min_positive = 66 };
+
+// MODULE-CONST raw value binding whose spelling shadows the builtin `s16`.
+// Reachable via `program_index.module_const_map` (issue 0093, const variant).
+`s16 :: IBox.{ max = 99, min = -99 };
+
 main :: () -> s32 {
-    // Raw value bindings whose spelling shadows a builtin numeric type name.
+    // LOCAL raw value bindings whose spelling shadows a builtin numeric type name.
     `f64 := FBox.{ epsilon = 11, max = 22, min_positive = 33 };
     `s32 := IBox.{ max = 78, min = -78 };
     `u8  := UBox.{ max = 7 };
 
     // Raw receiver → ordinary field READ (the value), never the numeric limit.
-    print("val f64: epsilon={} max={} min_positive={}\n",
-          `f64.epsilon, `f64.max, `f64.min_positive);   // 11 22 33
-    print("val s32: max={} min={}\n", `s32.max, `s32.min); // 78 -78
-    print("val u8: max={}\n", `u8.max);                    // 7
+    print("local f64: epsilon={} max={} min_positive={}\n",
+          `f64.epsilon, `f64.max, `f64.min_positive);    // 11 22 33
+    print("local s32: max={} min={}\n", `s32.max, `s32.min); // 78 -78
+    print("local u8: max={}\n", `u8.max);                    // 7
+
+    // GLOBAL raw receiver → ordinary field READ (issue 0093).
+    print("global f32: epsilon={} max={} min_positive={}\n",
+          `f32.epsilon, `f32.max, `f32.min_positive);    // 44 55 66
+    // MODULE-CONST raw receiver → ordinary field READ (issue 0093).
+    print("const s16: max={} min={}\n", `s16.max, `s16.min); // 99 -99
 
     // The value-field read carries the field type (s64 here): round-trips
     // through a typed binding, so a mistyped/boxed read would not type-check.
@@ -36,15 +58,19 @@ main :: () -> s32 {
     print("typed val e={}\n", e);                          // 11
 
     // Bare receiver (a type receiver, NOT the raw value) → STILL folds to the
-    // numeric limit, even though `s32`/`u8`/`f64` are bound in this same scope.
+    // numeric limit, even though a LOCAL (`s32`/`u8`/`f64`), GLOBAL (`f32`), or
+    // MODULE-CONST (`s16`) value of the same spelling is bound. The bare receiver
+    // is never blocked by any of the three value sources.
     print("lim s32.max={} s32.min={}\n", s32.max, s32.min); // 2147483647 -2147483648
     print("lim u8.max={}\n", u8.max);                       // 255
+    print("lim s16.max={} s16.min={}\n", s16.max, s16.min); // 32767 -32768
 
     // Bare float accessors still fold; the formatter is crude (issue 0090), so
     // pin the values by their defining properties rather than by printing.
     print("lim (1.0+f64.epsilon)!=1.0: {}\n", (1.0 + f64.epsilon) != 1.0); // true
     print("lim f64.inf > f64.max: {}\n", f64.inf > f64.max);               // true
     print("lim f64.min == -f64.max: {}\n", f64.min == -f64.max);           // true
+    print("lim f32.inf > f32.max: {}\n", f32.inf > f32.max);               // true
 
     return 0;
 }
diff --git a/examples/expected/0161-types-numeric-limit-value-shadow.stdout b/examples/expected/0161-types-numeric-limit-value-shadow.stdout
index e54c05a..0703721 100644
--- a/examples/expected/0161-types-numeric-limit-value-shadow.stdout
+++ b/examples/expected/0161-types-numeric-limit-value-shadow.stdout
@@ -1,9 +1,13 @@
-val f64: epsilon=11 max=22 min_positive=33
-val s32: max=78 min=-78
-val u8: max=7
+local f64: epsilon=11 max=22 min_positive=33
+local s32: max=78 min=-78
+local u8: max=7
+global f32: epsilon=44 max=55 min_positive=66
+const s16: max=99 min=-99
 typed val e=11
 lim s32.max=2147483647 s32.min=-2147483648
 lim u8.max=255
+lim s16.max=32767 s16.min=-32768
 lim (1.0+f64.epsilon)!=1.0: true
 lim f64.inf > f64.max: true
 lim f64.min == -f64.max: true
+lim f32.inf > f32.max: true
diff --git a/issues/0093-global-raw-value-numeric-limit-shadow.md b/issues/0093-global-raw-value-numeric-limit-shadow.md
new file mode 100644
index 0000000..cc412d0
--- /dev/null
+++ b/issues/0093-global-raw-value-numeric-limit-shadow.md
@@ -0,0 +1,84 @@
+> **RESOLVED** (NL.2 attempt 4). Root cause: the issue-0092 fix guarded the
+> numeric-limit intercept against value shadowing using ONLY lexical
+> `Scope.lookup`. But the ordinary identifier field-access path resolves a
+> value through THREE sources (`expr_typer.zig` `.identifier` arm): lexical
+> `scope` → program `global_names` → module value constants
+> `module_const_map`. A backtick raw identifier bound at MODULE scope
+> (`` `f64 := Box.{…} ``, a global, or `` `f64 :: Box.{…} ``, a module const)
+> is registered in `global_names` / `module_const_map`, NOT in `Scope`, so the
+> scope-only guard missed it and the intercept still folded `` `f64.epsilon ``
+> to the numeric limit — the same silent-wrong-value bug as 0092, one source
+> deeper. The module-const variant has the same root cause and is covered by
+> the same fix (no separate issue).
+>
+> Fix (close ALL THREE value-binding sources in one pass): a single shared
+> helper `Lowering.identifierBindsValue(name)` returns true when `name`
+> resolves through `scope.lookup` OR `program_index.global_names` OR
+> `program_index.module_const_map`. Used in BOTH resolvers so they cannot
+> desync (issue-0083 two-resolver class):
+> - `src/ir/lower.zig` — `lowerNumericLimit`: defer to ordinary field lowering
+>   (return null) when an `.identifier` receiver `identifierBindsValue`.
+> - `src/ir/expr_typer.zig` — numeric-limit inference arm: the `shadowed`
+>   check now calls the same helper.
+>
+> A bare `f64.epsilon` / `s32.max` (a `.type_expr` receiver, never an
+> `.identifier`) still folds, even when a global or module-const raw value of
+> the same spelling exists — the bare receiver is never value-shadowed.
+> Float-only-on-int and non-numeric-receiver errors are unchanged.
+>
+> Regression: `examples/0161-types-numeric-limit-value-shadow.sx` now exercises
+> all three binding kinds — a GLOBAL `` `f32 ``, a MODULE-CONST `` `s16 ``, and
+> LOCAL `` `f64 ``/`` `s32 ``/`` `u8 `` — each reading its field while the bare
+> spelling still folds. Unit test `src/ir/expr_typer.test.zig` pins the global
+> + module-const sources. NL.1 (`examples/0148`) / NL.2 (`examples/0159`,
+> `examples/0160`) unregressed.
+
+# 0093 — numeric-limit intercept hijacks global raw reserved-spelled value receivers
+
+## Symptom
+
+Field access on a **global** raw reserved-spelled value binding is interpreted as
+a builtin type numeric-limit access instead of an ordinary value field access.
+Observed: the repro prints `0.000000 2147483647` (`f64.epsilon` /
+`s32.max`). Expected: it prints `12 78` from the `Box` fields.
+
+## Reproduction
+
+```sx
+#import "modules/std.sx";
+
+Box :: struct { epsilon: s64; max: s64; }
+
+`f64 := Box.{ epsilon = 12, max = 34 };
+`s32 := Box.{ epsilon = 56, max = 78 };
+
+main :: () -> s32 {
+    print("{} {}\n", `f64.epsilon, `s32.max);
+    return 0;
+}
+```
+
+## Investigation prompt
+
+Investigate issue 0093: the issue-0092 value-binding precedence fix covers
+lexical locals but misses global raw value bindings. In `src/ir/lower.zig`, start
+at `Lowering.lowerNumericLimit` and the new issue-0092 guard around
+`Scope.lookup`. That guard returns `null` for a shadowing local, but global raw
+bindings are registered in `ProgramIndex.global_names` (and module constants in
+`ProgramIndex.module_const_map`), not in `Scope`, so an `.identifier` receiver
+whose text is `f64` / `s32` still folds to a numeric limit before ordinary
+global field lowering can read the value. Mirror the same rule in
+`src/ir/expr_typer.zig` so inferred types match lowering.
+
+Likely fix: for `.identifier` numeric-limit receivers, prefer any in-scope value
+binding source over the builtin-type fold: lexical `Scope.lookup`, global values
+(`program_index.global_names`), and module value constants where applicable.
+Keep `.type_expr` receivers folding as type receivers, so bare `f64.epsilon` and
+`s32.max` still fold even when a raw global value of the same spelling exists.
+
+Verification: pin the repro above as a regression. It should print `12 78`.
+Also verify the existing numeric-limit examples still pass:
+`examples/0148-types-int-numeric-limits.sx`,
+`examples/0159-types-float-numeric-limits.sx`,
+`examples/0160-types-float-numeric-limits-errors.sx`, and
+`examples/0161-types-numeric-limit-value-shadow.sx`.
diff --git a/readme.md b/readme.md
index ef80ca0..6bed385 100644
--- a/readme.md
+++ b/readme.md
@@ -98,8 +98,8 @@ subnormal — beware flush-to-zero CPU modes), `.inf`, and `.nan`. A float-only
 accessor on an integer (`s32.epsilon`), or any accessor on a non-numeric type, is
 a clean compile error. The fold applies only to a bare type-name receiver: a raw
 identifier that binds a value shadowing a type name (`` `f64 := … `` then
-`` `f64.epsilon ``) reads the value's field, not the limit. See `specs.md` →
-Numeric Limits.
+`` `f64.epsilon ``) reads the value's field, not the limit — for a local, global,
+or module-constant binding alike. See `specs.md` → Numeric Limits.
 
 ### Declarations
 
diff --git a/specs.md b/specs.md
index 2056a43..852fea4 100644
--- a/specs.md
+++ b/specs.md
@@ -343,10 +343,12 @@ qn  := f64.nan;           // a quiet NaN
 - **Type receiver vs. a shadowing value binding.** A numeric-limit access folds
   only when the receiver is a builtin numeric **type name** (`f64.epsilon`,
   `s32.max`, `u8.max`). A backtick raw identifier that binds a *value* whose
-  spelling shadows a type name (`` `f64 := … ``, F0.6) is an ordinary value:
-  `` `f64.epsilon `` reads that value's `epsilon` field — it does **not** fold to
-  the limit. The two never collide, even in the same scope — a bare builtin name
-  in expression position is always a type, and only the raw `` `…` `` spelling
+  spelling shadows a type name (F0.6) is an ordinary value: `` `f64.epsilon ``
+  reads that value's `epsilon` field — it does **not** fold to the limit. This
+  holds for **every** value-binding kind — a `` `f64 := … `` local, a module-scope
+  global, or a `` `f64 :: … `` module constant — so the fold can never silently
+  hijack a raw value, whatever its scope. The two never collide: a bare builtin
+  name in expression position is always a type, and only the raw `` `…` `` spelling
   can bind a value under it.
 
 ### Enum Types
diff --git a/src/ir/expr_typer.test.zig b/src/ir/expr_typer.test.zig
index 70e0435..b267ada 100644
--- a/src/ir/expr_typer.test.zig
+++ b/src/ir/expr_typer.test.zig
@@ -118,3 +118,62 @@ test "expr_typer: raw value binding shadows numeric-limit, bare type still folds
     var type_fa = node(.{ .field_access = .{ .object = &type_recv, .field = "epsilon" } });
     try std.testing.expectEqual(TypeId.f64, l.inferExprType(&type_fa));
 }
+
+// issue 0093: a raw value binding can shadow a builtin numeric type name through
+// any of three sources — lexical scope (issue 0092), program globals, or module
+// value constants. The shared `identifierBindsValue` guard consults all three,
+// so a global `` `f32 := Box.{…} `` and a module-const `` `s16 :: Box.{…} `` each
+// read the value's field (NOT the numeric-limit fold), while a bare `f32.max` /
+// `s16.max` (a `.type_expr` receiver) still folds. Pins the guard across the two
+// non-lexical sources the attempt-3 scope-only fix missed.
+test "expr_typer: global and module-const raw bindings shadow numeric-limit" {
+    var arena = std.heap.ArenaAllocator.init(std.testing.allocator);
+    defer arena.deinit();
+    const alloc = arena.allocator();
+    var module = ir_mod.Module.init(alloc);
+    defer module.deinit();
+    var l = Lowering.init(&module);
+
+    // `Box { max: s64 }` — the struct both raw bindings resolve to.
+    const box_fields = [_]ir_mod.types.TypeInfo.StructInfo.Field{
+        .{ .name = module.types.internString("max"), .ty = .s64 },
+    };
+    const box_ty = module.types.intern(.{ .@"struct" = .{
+        .name = module.types.internString("Box"),
+        .fields = &box_fields,
+    } });
+
+    // GLOBAL raw binding `` `f32 := Box.{…} `` — registered in global_names.
+    try l.program_index.global_names.put("f32", .{ .id = @enumFromInt(0), .ty = box_ty });
+    // MODULE-CONST raw binding `` `s16 :: Box.{…} `` — registered in module_const_map.
+    var const_val = node(.{ .int_literal = .{ .value = 0 } });
+    try l.program_index.module_const_map.put("s16", .{ .value = &const_val, .ty = box_ty });
+
+    // The shared guard sees both non-lexical bindings, but not an unbound spelling.
+    try std.testing.expect(l.identifierBindsValue("f32"));
+    try std.testing.expect(l.identifierBindsValue("s16"));
+    try std.testing.expect(!l.identifierBindsValue("u8"));
+
+    // `` `f32.max `` — global raw receiver → ordinary field read, types as s64
+    // (the field), not f32 (the fold).
+    var g_recv = node(.{ .identifier = .{ .name = "f32", .is_raw = true } });
+    var g_fa = node(.{ .field_access = .{ .object = &g_recv, .field = "max" } });
+    try std.testing.expectEqual(TypeId.s64, l.inferExprType(&g_fa));
+
+    // `` `s16.max `` — module-const raw receiver → ordinary field read, types as s64.
+    var c_recv = node(.{ .identifier = .{ .name = "s16", .is_raw = true } });
+    var c_fa = node(.{ .field_access = .{ .object = &c_recv, .field = "max" } });
+    try std.testing.expectEqual(TypeId.s64, l.inferExprType(&c_fa));
+
+    // bare `f32.max` — type_expr receiver, never shadowed → folds to f32, even
+    // though a global `` `f32 `` value is bound.
+    var bare_f32 = node(.{ .type_expr = .{ .name = "f32" } });
+    var bare_f32_fa = node(.{ .field_access = .{ .object = &bare_f32, .field = "max" } });
+    try std.testing.expectEqual(TypeId.f32, l.inferExprType(&bare_f32_fa));
+
+    // bare `s16.max` — type_expr receiver, never shadowed → folds to the s16
+    // type, even though a module-const `` `s16 `` value is bound.
+    var bare_s16 = node(.{ .type_expr = .{ .name = "s16" } });
+    var bare_s16_fa = node(.{ .field_access = .{ .object = &bare_s16, .field = "max" } });
+    try std.testing.expectEqual(TypeId.s16, l.inferExprType(&bare_s16_fa));
+}
diff --git a/src/ir/expr_typer.zig b/src/ir/expr_typer.zig
index 45692d9..f9a1803 100644
--- a/src/ir/expr_typer.zig
+++ b/src/ir/expr_typer.zig
@@ -143,9 +143,11 @@ pub const ExprTyper = struct {
                         // Skip the fold when a raw value binding shadows the
                         // builtin type name (`` `f64 := … ``) — mirrors the
                         // lowerNumericLimit guard so inference matches lowering
-                        // (issue 0092). A `.type_expr` receiver is never shadowed.
+                        // (issues 0092, 0093). The shared helper consults all
+                        // three value sources (scope / globals / module consts);
+                        // a `.type_expr` receiver is never shadowed.
                         const shadowed = fa.object.data == .identifier and
-                            (if (self.l.scope) |scope| (scope.lookup(tn) != null) else false);
+                            self.l.identifierBindsValue(tn);
                         if (!shadowed and
                             (TypeResolver.integerLimitFor(tn, fa.field) != null or
                                 TypeResolver.floatLimitFor(tn, fa.field) != null))
diff --git a/src/ir/lower.zig b/src/ir/lower.zig
index bfd37b1..cbaf751 100644
--- a/src/ir/lower.zig
+++ b/src/ir/lower.zig
@@ -4904,6 +4904,27 @@ pub const Lowering = struct {
         return self.lowerFieldAccessOnType(obj, obj_ty, fa.field, span);
     }
 
+    /// True when an `.identifier` receiver text resolves to an in-scope VALUE
+    /// binding rather than a builtin type. A backtick raw identifier (F0.6) can
+    /// bind a value whose spelling shadows a builtin type name (`` `f64 := … ``);
+    /// such a value is reachable through the same three sources the ordinary
+    /// identifier field-access path consults (see `expr_typer` `.identifier`
+    /// arm): lexical `scope`, program `global_names`, and module value
+    /// constants `module_const_map`. The numeric-limit intercept must defer to
+    /// ordinary field access whenever ANY of the three binds the name, so a
+    /// raw value field read is never hijacked into a numeric-limit fold
+    /// (issues 0092 local / 0093 global + module-const). A single helper used
+    /// by both lowering and inference keeps the two resolvers in lockstep
+    /// (issue-0083 two-resolver defect class).
+    pub fn identifierBindsValue(self: *Lowering, name: []const u8) bool {
+        if (self.scope) |scope| {
+            if (scope.lookup(name) != null) return true;
+        }
+        if (self.program_index.global_names.get(name) != null) return true;
+        if (self.program_index.module_const_map.get(name) != null) return true;
+        return false;
+    }
+
     /// Numeric-limit accessor intercept (`<Type>.min`/`.max`/`.epsilon`/
     /// `.min_positive`/`.true_min`/`.inf`/`.nan`), a sibling of the `error.X` /
     /// `Struct.CONST` / pack-arity identifier-receiver intercepts in
@@ -4933,13 +4954,10 @@ pub const Lowering = struct {
         // shadows a builtin type name (`` `f64 := … ``). Field access on that
         // value is an ordinary field read, not a numeric-limit fold — defer to
         // the normal field-access path when the receiver identifier resolves to
-        // an in-scope value binding (issue 0092). A `.type_expr` receiver is
-        // unambiguously a type and can never be value-shadowed.
-        if (fa.object.data == .identifier) {
-            if (self.scope) |scope| {
-                if (scope.lookup(name) != null) return null;
-            }
-        }
+        // a value binding through any of scope / globals / module consts
+        // (issues 0092, 0093). A `.type_expr` receiver is unambiguously a type
+        // and can never be value-shadowed.
+        if (fa.object.data == .identifier and self.identifierBindsValue(name)) return null;
 
         if (TypeResolver.integerLimitFor(name, fa.field)) |value| {
             return self.builder.constInt(value, ty);

From bdf6433e7242f262d4bc7169aff23972314a58db Mon Sep 17 00:00:00 2001
From: agra <alex@swipelab.co>
Date: Fri, 5 Jun 2026 00:35:52 +0300
Subject: [PATCH 5/5] test(lang): pin f64.min via signed union bit view in
 example 0159 [NL.2]

---
 examples/0159-types-float-numeric-limits.sx         | 13 +++++++++----
 .../expected/0159-types-float-numeric-limits.stdout |  1 +
 2 files changed, 10 insertions(+), 4 deletions(-)

diff --git a/examples/0159-types-float-numeric-limits.sx b/examples/0159-types-float-numeric-limits.sx
index f2753ab..dc82d7c 100644
--- a/examples/0159-types-float-numeric-limits.sx
+++ b/examples/0159-types-float-numeric-limits.sx
@@ -25,10 +25,11 @@
 #import "modules/std.sx";
 
 // `bits` mirrors each float's raw IEEE-754 storage. f64 needs 64 bits, f32 32.
-// The f64 union uses a `u64` view so the all-ones-ish positive patterns read as
-// their true magnitude; the negative `f64.min` pattern (0xFFEF…) overflows the
-// i64 literal parser, so it is pinned by the `min == -max` property instead.
-Uf64 :: union { f: f64; bits: u64; }
+// The f64 union's `bits` (u64) view reads the all-ones-ish positive patterns as
+// their true magnitude; its `s` (s64) view pins the negative `f64.min` pattern
+// (0xFFEF…), whose unsigned form overflows the u64 literal parser, by comparing
+// the signed reinterpret to -4503599627370497.
+Uf64 :: union { f: f64; bits: u64; s: s64; }
 Uf32 :: union { f: f32; bits: u32; }
 
 main :: () -> s32 {
@@ -43,6 +44,10 @@ main :: () -> s32 {
 
     o.f = f64.max;
     print("f64.max          {}\n", o.bits == 0x7FEFFFFFFFFFFFFF);  // true
+    // f64.min = -max; its bit pattern 0xFFEFFFFFFFFFFFFF overflows an unsigned u64
+    // literal, so it is pinned directly via the SIGNED s64 view: -4503599627370497.
+    o.f = f64.min;
+    print("f64.min          {}\n", o.s == -4503599627370497);     // true (bits 0xFFEFFFFFFFFFFFFF)
     o.f = f64.epsilon;
     print("f64.epsilon      {}\n", o.bits == 0x3CB0000000000000);  // true
     o.f = f64.min_positive;
diff --git a/examples/expected/0159-types-float-numeric-limits.stdout b/examples/expected/0159-types-float-numeric-limits.stdout
index f562d64..b6c69f5 100644
--- a/examples/expected/0159-types-float-numeric-limits.stdout
+++ b/examples/expected/0159-types-float-numeric-limits.stdout
@@ -1,5 +1,6 @@
 f64.true_min     true
 f64.max          true
+f64.min          true
 f64.epsilon      true
 f64.min_positive true
 f64.inf          true