Fix float8 precision loss during CDC replay

[teknogeek0]

Problem

During CDC replay, float8 values were formatted with %f, which truncates to 6 decimal places. This silently corrupted any value needing more precision — e.g. -216237.00000035969 was applied to the target as -216237.000000.

This is the issue reported in upstream dimitri/pgcopydb dimitri#968. A customer hit it on PG16 and worked around it with wal2json's --wal2json-numeric-as-string.

Fix

Emit the shortest decimal string that round-trips back to the same IEEE-754 double: try 15 significant digits, increasing to the 17 required for a guaranteed round-trip, stopping at the first precision that parses back to the same value.

for (int precision = 15; precision <= 17; precision++)
{
    sformat(string, sizeof(string), "%.*g", precision, value->val.float8);
    if (strtod(string, NULL) == value->val.float8)
        break;
}

Why not a fixed %.17g / %.17f?

A fixed high precision fixes the truncation but introduces a silent data-divergence regression. Under REPLICA IDENTITY FULL (tables without a primary key), the replayed value is used not only as an INSERT/UPDATE value but also as a WHERE-clause key for UPDATE/DELETE. A fixed %.17g expands 5.99 to "5.9900000000000002", which does not match the stored numeric value 5.99, so the replayed UPDATE/DELETE matches zero rows and the target silently diverges. (Upstream dimitri#968's %.17f has this same defect.)

The shortest-round-trip form keeps 5.99 as "5.99" (matches the stored value) while still emitting -216237.00000035969 in full when the precision is genuinely needed — avoiding both the truncation and the WHERE-match failure.

PostgreSQL accepts the %g output (including scientific notation for extreme magnitudes) as float input when the value is bound as a query parameter, which is how these values are applied.

Scope

This addresses float4/float8 columns. True arbitrary-precision numeric fidelity is a separate concern — wal2json coerces numeric to a C double upstream of this code, so --wal2json-numeric-as-string remains the correct option there.

Testing

The golden fixture tests/cdc-wal2json/000000010000000000000002.sql is updated to the corrected output (5.99 / 11.95 / 11.99).

Full CDC / follow / unit / base-clone suites pass on PG18. Notably the data-applying follow-defer-indexes and follow-defer-validate-fks suites pass — these exercise REPLICA IDENTITY FULL replay and are what surface the WHERE-match regression that a naive fixed-precision fix would introduce (the text-diff cdc-wal2json suite alone does not catch it).