Fix powi spurious overflow for bases near 1 on 32-bit (#95)

[cmpute]

Summary

Fixes #95 — high-precision base-2 → base-10 conversion (FBig::with_base) panicked on 32-bit targets (wasm32, i686) with arithmetic operations with the infinity are not allowed!.

Root cause

with_base evaluates exp(r) internally as sum^(B^n) via Context::powi with a huge integer exponent B^n on a base (sum, the Maclaurin-series result) that is very close to 1.

powi's overflow guard estimated log2(base) with log2_est. For a base ≈ 1 (a large significand with a large negative exponent), log2(base) is the difference of two ~1e3-magnitude terms and catastrophically cancels to ~1e-4 of f32 noise instead of ~0. Scaled by the large exponent, that noise crossed the overflow threshold — which on 32-bit is only isize::MAX·log2(B) ≈ 7e9 (vs ≈3e19 on 64-bit), so it only manifested on 32-bit:

base_log2 = 1.22e-4   (true log2(sum) ≈ 3e-17)
exp_f64 * base_log2 = 1e16 * 1.22e-4 = 1.22e12  >  threshold 7.13e9   → spurious Overflow

powi returned ±inf, and the subsequent v << s panicked (shifting an infinity).

Fix

Switch the guard to the bit-length-based log2_bounds, which doesn't cancel:

• declare overflow only on the lower bound of log2(base) (no false positives),
• declare underflow only on the upper bound (no false underflows, and excluded for a zero base),
• anything in between (e.g. a base that straddles 1) is computed.

Verification

Reproduced on real 32-bit (i686-unknown-linux-gnu, usize::BITS = 32) before the fix, confirmed gone after:

• Reporter's exact input (precision 586) now converts to the same value as 64-bit.
• New regression test convert::tests::with_base_high_precision_no_overflow reliably fails on the buggy code and passes on the fix (verified on i686).
• dashu-float tests pass on i686 (77) and x86_64 (83); workspace tests + clippy (-D warnings) + fmt --check + cargo check --all-features --tests all green.

The crash was codegen-sensitive (it depended on how LLVM folded the cancelling f32 ops, so it surfaced via the dashu meta-crate / default features but not a bare no_std build) — the fix is at the algorithm level, so it's robust to all of those.

🤖 Generated with Claude Code