data.drivers._pelt._codecs_numeric

E2 codec family: arbitrary-precisionnumeric / decimal ↔ Decimal.

PostgreSQL'snumericis a base-10000 ("NBASE") big-decimal: a sign, a base-10000 exponent (weight), a display scale (dscale— the number of fractional decimal digits to render), and a run of base-10000 digit groups most-significant first. The binary wire layout produced bynumeric_send / consumed by numeric_recvis::

int16  ndigits   -- count of base-10000 digit groups that follow
int16  weight    -- base-10000 exponent of the FIRST group (group value = digit * 10000^weight)
uint16 sign      -- 0x0000 positive, 0x4000 negative,
                    0xC000 NaN, 0xD000 +Infinity, 0xF000 -Infinity (PG14+)
int16  dscale    -- display scale: number of fractional decimal digits
uint16[ndigits]  -- base-10000 digit groups, 0..9999, most-significant first

Decoding reconstructs the exactDecimal, preserving dscaleas the result's scale (soDecimal('1.50') round-trips as 1.50, not 1.5). Encoding goes the other way: it splits aDecimal's decimal digits into base-10000 groups aligned on the decimal point, normalizes off leading/trailing zero groups, and carries the original scale throughdscale. The special-value sign codes carry Decimal NaN/ Infinitywith no digit groups.

The text format defers toDecimal(str)— PostgreSQL emits the canonical decimal string. prefers_binary stays Trueso the hot path takes the exact base-10000 route and never round-trips through a lossy float. This module touches no socket and no anyio: bytes in, Decimalout.