perf(assert): compare assert_within_delta in fixed point, not via bc

[Chemaclass]

🤔 Background

assert_within_delta called bashunit::math::calculate twice, and each call is a subshell wrapping a bc (or awk) process — four forks per assertion, on a per-assertion path.

	before	after
assert_within_delta ×200	1092 ms	165 ms
fork-free floor	108 ms	108 ms

💡 Changes

The comparison is |expected - actual| <= delta, which needs no floating point at all. All three operands are padded to one decimal scale and compared as integers, in pure bash.

The fast path is deliberately narrow and refuses what it cannot represent exactly — exponent notation, a sign anywhere but the front, or enough digits to risk 64-bit overflow — so those fall through to the existing bc/awk chain rather than getting a quietly wrong answer.

bashunit::math::is_le gets the same fast path; it had the identical bc > awk > strip-decimals chain for the same reason.

🐛 Also fixes a bug, rather than moving it

_is_numeric accepts a leading +, but bc cannot parse one: +5 - 5 returned an empty string, which compared unequal to "1", so assert_within_delta +5 5 1 failed. The sign is now stripped once before either path — so the fallback is fixed too, not just bypassed.

✅ How agreement was verified

Forcing the fixed-point path to always refuse leaves the entire numeric suite green on the bc/awk path. That's the check that the two paths agree, rather than just that the new one passes.

Slots rather than echoes for the three helpers — not stylistic: the caller needs the decimal count three times per assertion, and three $( ) captures would cost more than the two bc forks this change exists to remove.

🔒 Verification

Bash 3.0 safe — parameter expansion, case and integer arithmetic only; compat gate green (14/14). Fork budgets unchanged. make sa · make lint · bash build.sh bin -v → ✅ Build verified ✅ · 1661 sequential / 1620 parallel-simple-strict.

[Chemaclass]

Reopening to re-trigger CI; no pull_request runs were created.