Expand test suite

[Copilot]

Cleans up the test suite per scope constraints: no src changes, no cast-mirrors-conv tautology tests, minimal array/tuple coverage, use std consts where appropriate.

Changes

• Remove tests/common/mod.rs — drop assert_ok_eq, assert_range, assert_inexact wrappers; use assert_eq! with Ok(...) / Err(Error::Range) / Err(Error::Inexact) directly
• Remove cast_approx_mirrors_conv_approx test from conv_approx.rs — testing that cast mirrors conv is a tautology since Cast is a blanket impl over Conv
• Improve float_approx.rs — replace raw bit-pattern construction with std consts:
  • f64::MIN_POSITIVE / 2.0 for subnormal, (f32::MIN_POSITIVE as f64) / 2.0 for small-normal
  • Derive negatives with unary - instead of bit manipulation
  • Inline single-use result variables (positive_zero, etc.)
  • Remove unnecessary parentheses: (1151u64) << 52 → 1151u64 << 52
• Revert src/impl_float.rs to master — a prior commit restructured the if/else if in ConvApprox<f64> for f32; reverted without touching anything else

Before / After (representative)

// Before
assert_ok_eq(i8::try_conv_trunc(f32::from(i8::MIN)), i8::MIN);
assert_range(i8::try_conv_trunc(f32::from(i8::MAX) + 1.0));

// After
assert_eq!(i8::try_conv_trunc(f32::from(i8::MIN)), Ok(i8::MIN));
assert_eq!(i8::try_conv_trunc(f32::from(i8::MAX) + 1.0), Err(Error::Range));

[dhardy]

So the purpose of this was to address (or at least chip away at) the last limitation listed in the readme:

You want a thoroughly tested library (we're not quite there yet)

This work is rather sloppy (over-zealously checking some parts of the API surface while still only checking a small number of actual conversions), but it did find something at least: impl ConvApprox<f64> for f32 doesn't handle small (normal) f64 values which could be converted to sub-normal f32 values.

I considered writing a test suite using proptest. The main issue I see is that the bounds used in the tests would need much of the same code as the conversions use, making the test suite prone to the same bugs it is trying to catch.

(Also: should ConvApprox<f64> for f32 convert f64::NAN to f32::NAN or continue to return Error::Range? The fact that Conv<f32> for f64 is derived from From and does convert NAN makes this incongruous. And should too large f64 values convert to f32::INFINITY or result in a range error?)

[dhardy]

Notes for myself.

These tests are simultaneously repetitive, inexhaustive and inconsistent in style. And mostly cover trivial stuff. But with a few tweaks they have some value.