Revision history for Struct::Codec

0.05	2026-09-18
        - FIXES a heap corruption decoding a hash key of length zero that
          carries the UTF-8 flag. 
        - t/25-wide-nv.t used perl's numifier to build -1.797...e+308, which
          overflows to -Inf below 5.30, and an infinity compares against
          nothing. The exponent is divided out and the mantissa named.
        - t/14-corrupt.t asserted two refusals that only a 64-bit UV gives:
          an eleven-byte varint is refused for its width, not its length,
          where a UV is 32 bits, and the descriptor fixture is now a value
          past INT_MAX that still fits one.

0.04	2026-09-18
        - A float on a perl built -Dusequadmath kept one significant digit: a
          third came back 0.3, and on a 32-bit one perl died "panic:
          my_snprintf buffer overflow". The precision was passed as "%.*",
          which my_snprintf never hands to quadmath_snprintf. It is written
          into the format now, as perl's own sv.c does.
        - A float on a perl built -Duselongdouble below 5.30 came back a ULP
          or two out: the digits were read with Atof, which is perl's own
          numifier and was not correctly rounded until 5.30 (perl RT #41202).
          The decoder names strtold, strtoflt128 or strtod for the build.
        - Those digits now use '.' for a radix whatever the locale is.
        - t/25-wide-nv.t asserts the digit count on the bytes, and no longer
          uses perl's numifier as an oracle on a perl whose numifier is the
          thing at fault.

0.03	2026-09-17
        - FIXES a heap corruption when encoding an anonymous sub. The
          deparser is loaded with load_module, which runs perl and can
          reallocate the argument stack; the stack pointer taken before it
          was then stale, and pushing the deparse arguments through it wrote
          to freed memory. It surfaced only when the load triggered a
          realloc, so it depended on how deep the stack already was: a bare
          encode was usually fine and a sub encoded from inside a larger
          program crashed at destruction. SPAGAIN after the load. Seen as a
          SIGABRT/SIGSEGV in t/32-kinds.t on threaded 5.18.
        - t/16-leak.t built its probe list with a regexp unconditionally,
          which the codec refuses below 5.12, so the file died there on
          5.10. The regexp joins the list only where it can be encoded, as
          t/32 already skips it.
        - A float on a perl built -Duselongdouble or -Dusequadmath kept only
          what a double holds. Such a value now goes out as decimal digits
          under tag 0x38, and comes back the number it was. A perl whose NV
          is a double writes the same bytes as before.
        - t/25-wide-nv.t.

0.01	2026-09-12
        - First version.