[chiliec]

Fixes #140

What

Context.Quantize ignored the rounding mode when the input had fewer significant digits than the target exponent (i.e. its magnitude was smaller than one unit in the last place). It truncated straight to zero regardless of the configured Rounder:

ctx := apd.Context{Precision: 100, Rounding: apd.RoundCeiling, MaxExponent: 1000, MinExponent: -1000, Traps: apd.DefaultTraps}
var d apd.Decimal
d.SetString("0.002")
ctx.Quantize(&d, &d, -1)
// got "0.0", want "0.1"

RoundCeiling(1.002, -1) already returned 1.1; only the sub-ULP case was wrong.

Fix

In the p < 0 branch of quantize, run the discarded coefficient through the context's Rounding.ShouldAddOne (the same mechanism the normal rounding path uses) instead of unconditionally zeroing it.

Tests

Added TestQuantizeRounding covering the reported case plus every rounding mode (ceiling/floor/up/down/half-up/half-even, positive and negative). Verified genuine RED→GREEN:

# with fix
$ go test -run TestQuantizeRounding .
ok  github.com/cockroachdb/apd/v3

# revert context.go, keep the test
$ git stash push context.go && go test -run TestQuantizeRounding .
--- FAIL: TestQuantizeRounding/0.002:_-1_ceiling
    expected: 0.1, got: 0.0
--- FAIL: TestQuantizeRounding/-0.002:_-1_floor
    expected: -0.1, got: -0.0
FAIL

Full suite (including the GDA conformance tests) is green:

$ go test ./...
ok  github.com/cockroachdb/apd/v3  1.524s

Happy to adjust.

[rafiss]

The fix is right. I checked the branch against Python's decimal for all eight rounding modes on sub-ULP inputs and signed zeros, including through RoundToIntegralValue and RoundToIntegralExact, and results and flags match. PostgreSQL numeric agrees on the cases it can express (half away from zero, trunc, ceil and floor), differing only in the sign of zero, which numeric does not have.

I pushed c3295ad on top. In this branch the coefficient has fewer digits than the exponent difference, so the discarded value is always below one half and discard.Cmp(decimalHalf) is always -1; the commit passes -1 directly with a comment saying why, which also makes it clear that only ceiling, floor, up and 05up can round to one unit here. It adds Round05Up and RoundHalfDown rows, asserts the Inexact and Rounded conditions, checks a separate destination, and adds a RoundToIntegral test since toIntegral reaches the same branch.

One small thing for the record: the branch fires for values below a tenth of a ULP, not below one ULP. 0.05 quantized to one decimal place already rounded correctly before this change.

[chiliec]

Thanks for the thorough cross-check and the follow-up commit — and good catch on the threshold: it's below a tenth of a ULP, not one ULP. Appreciate the merge.