docs(spec): split range bounds from counts; pin inline-for range semantics (0083)

specs.md lumped `inline for` / `for` range bounds in with counts (array
dimension, Vector lane count, generic value-param count) under the
count negative-rejection rule. A range bound is a range ENDPOINT, not a
count: negative endpoints are valid and an empty/inverted range runs zero
iterations. The compiler already implements this correctly (Agra ruling:
spec-text bug, no code change).

- specs.md: counts and range bounds are now described separately. Counts
  reject negatives; bounds accept any compile-time integer (negatives
  valid, integral floats fold) but still reject a non-integral float
  because the loop cursor must be an integer.
- examples/0612-comptime-inline-for-range-bounds.sx: `inline for -2..1`
  and `for -2..1` both sum -3; `inline for 0..(-2.0)` runs zero
  iterations (empty range). Runtime/comptime parity asserted.
- examples/1138-diagnostics-inline-for-non-integral-bound.sx: a
  non-integral float bound `inline for 0..4.5` is a clean diagnostic,
  exit 1 (must-be-integer still applies to bounds).

Count consumers (1132/1133/1134/1135) unchanged and green.

examples/0612-comptime-inline-for-range-bounds.sx

@@ -0,0 +1,24 @@
+// An `inline for` / `for` range bound is a range ENDPOINT, not a count, so the
+// count negative-rejection rule does NOT apply to it: negative endpoints are
+// valid and an empty/inverted range simply runs zero iterations.
+//
+// Regression (F0.4 attempt 11, Agra ruling): the spec wrongly lumped inline-for
+// bounds with counts (array dim / Vector lane / value-param), which reject
+// negatives. Bounds are exempt — `inline for -2..1` iterates -2,-1,0 and an
+// integral-float empty range `0..(-2.0)` runs zero iterations. Comptime and
+// runtime loops must agree.
+#import "modules/std.sx";
+
+main :: () {
+    s := 0;
+    inline for -2..1: (i) { s += i; }
+    print("inline for -2..1 sum = {}\n", s);   // -2 + -1 + 0 = -3
+
+    r := 0;
+    for -2..1: (i) { r += i; }
+    print("for -2..1 sum = {}\n", r);          // -2 + -1 + 0 = -3 (runtime parity)
+
+    e := 0;
+    inline for 0..(-2.0): (i) { e += i; }
+    print("inline for 0..(-2.0) sum = {}\n", e);  // empty range -> 0 iterations
+}

examples/1138-diagnostics-inline-for-non-integral-bound.sx

@@ -0,0 +1,14 @@
+// A NON-integral float (`4.5`) as an `inline for` range bound is a hard error:
+// the loop cursor must be a compile-time integer, so only an integral float
+// (`4.0`, `-2.0`) folds. Clean diagnostic + non-zero exit.
+//
+// Regression (F0.4 attempt 11, Agra ruling): range bounds are exempt from the
+// count negative-rejection (negatives are valid endpoints), but the
+// must-be-integer requirement still applies — `4.5` has no integer value.
+#import "modules/std.sx";
+
+main :: () {
+    s := 0;
+    inline for 0..4.5: (i) { s += i; }
+    print("unreachable: {}\n", s);
+}

examples/expected/0612-comptime-inline-for-range-bounds.exit

@@ -0,0 +1 @@
+0

examples/expected/0612-comptime-inline-for-range-bounds.stderr

@@ -0,0 +1 @@
+

examples/expected/0612-comptime-inline-for-range-bounds.stdout

@@ -0,0 +1,3 @@
+inline for -2..1 sum = -3
+for -2..1 sum = -3
+inline for 0..(-2.0) sum = 0

examples/expected/1138-diagnostics-inline-for-non-integral-bound.exit

@@ -0,0 +1 @@
+1

examples/expected/1138-diagnostics-inline-for-non-integral-bound.stderr

@@ -0,0 +1,5 @@
+error: inline for: range end is not a compile-time integer
+  --> examples/1138-diagnostics-inline-for-non-integral-bound.sx:12:19
+   |
+12 |     inline for 0..4.5: (i) { s += i; }
+   |                   ^^^

examples/expected/1138-diagnostics-inline-for-non-integral-bound.stdout

@@ -0,0 +1 @@
+

specs.md

@@ -651,11 +651,20 @@ Arrays can also be constructed programmatically with the `Array` builtin:
 MyArr :: Array(5, s32);   // equivalent to [5]s32
 ```
 
-An array dimension — and likewise a `Vector` lane count, a generic value-param
-count, and an `inline for` bound — accepts any compile-time numeric constant
-whose value is a positive integral number. An integral float (`4.0`, or a
-float-typed const `N : f64 : 4.0`) folds to its integer (`[4.0]s64` ≡ `[4]s64`);
-a non-integral float (`4.5`) or a negative value is rejected.
+A **count** — an array dimension, a `Vector` lane count, or a generic
+value-param count — accepts any compile-time numeric constant whose value is a
+positive integral number. An integral float (`4.0`, or a float-typed const
+`N : f64 : 4.0`) folds to its integer (`[4.0]s64` ≡ `[4]s64`); a non-integral
+float (`4.5`) or a negative value is rejected.
+
+A **range bound** — the start/end of an `inline for` or `for` range — is a
+range *endpoint*, not a count, so the count rules above do not apply. A bound
+accepts any compile-time **integer**, including a negative one; an integral
+float (`-2.0`) folds to its integer. A non-integral float (`4.5`) is still
+rejected, because the loop cursor must be a compile-time integer. Negative
+endpoints are valid: `inline for -2..1` iterates `-2, -1, 0`. An empty or
+inverted range (start ≥ end, e.g. `0..(-2.0)`) simply runs zero iterations
+rather than being an error.
 
 ### Slice Types
 A slice `[]T` is a fat pointer `{ptr, i64}` referencing a contiguous sequence of `T` elements. Same runtime layout as `string`.