Fix toString(base) under-rounding for odd bases (#412)

[youdie006]

Fixes #412.

Problem

toString(base) under-rounds for odd bases. Under { DECIMAL_PLACES: 0, ROUNDING_MODE: ROUND_HALF_UP }:

new BigNumber('0.6').toString(3)   // '0'  (should be '1')
new BigNumber('2.6').toString(3)   // '2'  (should be '10')

.toString(10) is correct, so base 10 is the oracle. The values are exact (string constructor), so this is not float error.

Root cause

In convertBase the round-up decision compared the single rounding digit i against k = baseOut / 2. For an odd baseOut, k is fractional (1.5 for base 3), so i == k is unreachable and the remainder flag is never consulted: a discarded tail whose leading digit is floor(baseOut/2) was treated as "< half" even when the full tail was >= half a ULP.

A local tweak isn't sufficient: for an odd base, half a ULP is 0.mmm... (m = (baseOut-1)/2), a non-terminating fraction, so the tail must be compared against a repeating m -- but div only produced ~2 guard digits, giving byte-identical output for values that must round differently (e.g. base-3 0.49->0 vs 0.51->1 both yield [1,1,1]).

Fix

1. For an odd output base, compute enough extra base-baseOut digits that the first tail digit differing from m is always available. The leading run of m digits before the tail is decided is at most fracLen * log_baseOut(10).
2. Replace the round-up decision with a base-parity-aware one:
   • directed modes (UP/DOWN/CEIL/FLOOR) are parity-independent -- unchanged;
   • even bases keep the exact original logic (byte-identical output);
   • odd bases compare the tail digit-by-digit against m to get its position relative to half. For ROUND_HALF_EVEN on an exact half, the even neighbour has an even mantissa, which (baseOut being odd) means an even sum of kept digits -- generalising the even-base xc[d-1] & 1 test.

Even-base and decimal output are unchanged.

Tests

• The two reported cases plus an inline exact BigInt oracle cross-check over odd bases {3,5,7,9} (and even controls {2,4,8,16}) for every rounding mode {0..8}.
• Full suite: 71133 of 71133 pass (node 22).
• Cross-checked against an independent exact-rational oracle over ~160k cases (random, half-boundary adversarial, large fraction lengths) -- 0 mismatches, and 0 changes to even-base output.
• Four auto-generated odd-base fixtures were each off by one ULP (they had recorded the old buggy output); corrected to the exact values.

dist/ was rebuilt with node build.js.

---

This fix was prepared with AI assistance (Claude) and reviewed by me before submission.