gh-NNNN: Route float→string through _Py_fmt_dtoa

Swap the _Py_dg_dtoa / _Py_dg_freedtoa call pair at the three remaining
in-tree call sites (pystrtod.c's format_float_short and floatobject.c's
double_round) for _Py_fmt_dtoa / _Py_fmt_dtoa_free. Covers all three
cpython dtoa modes bit-exactly:

  * mode 0 (shortest / repr): fmt::detail::dragonbox::to_decimal
  * mode 2 (N sig digits):    fmt::detail::format_float, specs=exp
  * mode 3 (N frac digits):   fmt::detail::format_float, specs=fixed

fmt's format_float with specs=exp takes `precision` as total-digits-to-
emit, not "digits after the leading one" — so for mode 2 N sig digits we
pass N directly. For specs=fixed fmt's adjust_precision adds the decade
internally and handles negative precision natively, which is exactly
what CPython mode 3 with negative ndigits needs.

Python/dtoa.c is still in the build for _Py_dg_strtod (commit 6 handles
that swap) but _Py_dg_dtoa is now unreferenced.

Full float-formatting regression suite passes: 2,239 tests across
test_float, test_format, test_fstring, test_strtod, test_json, test_math,
test_cmath, test_complex, test_builtin, test_decimal, test_statistics,
test_string, test_str, test_reprlib, test_pprint, test_csv, test_struct,
test_long.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: eendebakpt

Objects/floatobject.c

@@ -900,7 +900,7 @@ float___ceil___impl(PyObject *self)
 
 #if _PY_SHORT_FLOAT_REPR == 1
 /* version of double_round that uses the correctly-rounded string<->double
-   conversions from Python/dtoa.c */
+   conversions from the fmt/wuffs shims in Python/_fmt/ and Python/_wuffs/. */
 
 static PyObject *
 double_round(double x, int ndigits) {
@@ -914,7 +914,7 @@ double_round(double x, int ndigits) {
 
     /* round to a decimal string */
     _Py_SET_53BIT_PRECISION_START;
-    buf = _Py_dg_dtoa(x, 3, ndigits, &decpt, &sign, &buf_end);
+    buf = _Py_fmt_dtoa(x, 3, ndigits, &decpt, &sign, &buf_end);
     _Py_SET_53BIT_PRECISION_END;
     if (buf == NULL) {
         PyErr_NoMemory();
@@ -951,7 +951,7 @@ double_round(double x, int ndigits) {
     if (mybuf != shortbuf)
         PyMem_Free(mybuf);
   exit:
-    _Py_dg_freedtoa(buf);
+    _Py_fmt_dtoa_free(buf);
     return result;
 }
 

Python/_fmt/fmt_dtoa.cc

@@ -103,8 +103,11 @@ char *_Py_fmt_dtoa(double d, int mode, int ndigits,
     bool fixed;
     if (mode == 2) {
         specs.set_type(fmt::presentation_type::exp);
-        // max(1, ndigits) significant digits, fmt wants precision = N - 1.
-        precision = (ndigits < 1 ? 1 : ndigits) - 1;
+        // fmt::detail::format_float with specs=exp emits exactly `precision`
+        // digits (no leading-digit adjustment — that's only for specs=fixed,
+        // where adjust_precision adds the decade). So for N significant
+        // digits we pass precision = N directly.
+        precision = (ndigits < 1 ? 1 : ndigits);
         fixed = false;
     } else {
         // Treat any unrecognised mode as 3 (matches _Py_dg_dtoa's fallthrough).

Python/pystrtod.c

@@ -980,11 +980,12 @@ format_float_short(double d, char format_code,
     Py_ssize_t decpt, digits_len, vdigits_start, vdigits_end;
     _Py_SET_53BIT_PRECISION_HEADER;
 
-    /* _Py_dg_dtoa returns a digit string (no decimal point or exponent).
-       Must be matched by a call to _Py_dg_freedtoa. */
+    /* fmt-backed drop-in for _Py_dg_dtoa. Returns a digit string (no
+       decimal point or exponent). Must be matched by a call to
+       _Py_fmt_dtoa_free. Covers all three modes bit-exactly. */
     _Py_SET_53BIT_PRECISION_START;
-    digits = _Py_dg_dtoa(d, mode, precision, &decpt_as_int, &sign,
-                         &digits_end);
+    digits = _Py_fmt_dtoa(d, mode, precision, &decpt_as_int, &sign,
+                          &digits_end);
     _Py_SET_53BIT_PRECISION_END;
 
     decpt = (Py_ssize_t)decpt_as_int;
@@ -1212,7 +1213,7 @@ format_float_short(double d, char format_code,
         assert(p-buf < bufsize);
     }
     if (digits)
-        _Py_dg_freedtoa(digits);
+        _Py_fmt_dtoa_free(digits);
 
     return buf;
 }