Adopt prior-art implementation patterns; fix fp-store overflow terminate

[NiceAndPeter]

Comparison pass over the resources.md libraries, adopting the implementation patterns bound was missing, plus a crash fix the new accuracy harness surfaced on its first run.

Adopted patterns

• Compile-fail diagnostics suite (tests/fail/, from bounded::integer/CNL): each deliberately ill-formed TU names its expected static_assert message in a // EXPECT: line; CMake passes the test only when the build fails with that message. A control TU proves the harness builds valid code.
• Accuracy characterization (tests/accuracy.cpp + committed docs/accuracy.md, fpm-style): all 16 one-arg transcendentals × 3 engines swept against a long-double reference; accuracy_report target regenerates the doc. dbl is correctly rounded everywhere (max 0.5 notch); flt degrades to ~3 notches on sinh/cosh/exp; cordic sits at 0.5–0.9.
• detail/rep.hpp: fp representation propagation deduped out of addition/multiplication/division into one fp_rep trait (codegen-neutral, gates pass).
• Grid hull + std::common_type + mixed-grid min/max (bounded::integer): hull(grid, grid) = interval hull + notch gcd (valid by construction via the lattice anchor), common_bound_t, std::common_type specialisation in numeric_limits.hpp.
• Docs: "implementation patterns compared" section in resources.md (adopted/rejected with reasons); roadmap notes for type-level interval unions and the 128-bit rounded store.

Bug fix

Storing a full-mantissa double-derived rational onto a snap grid with large |Lower| (e.g. math::dbl::tan's auto output grid) overflowed the exact 64-bit (rhs − Lower)/Notch rational and terminated through the noexcept engine as bad_optional_access. Now:

• rational::add_impl rescues mixed-sign cross-product overflow in 128-bit when the difference fits umax
• store_checked reports errc::overflow instead of dereferencing nullopt
• round_quotient is total: overflow-safe q/r offset rule, used as fallback when the signed value-index rebuild cannot fit 64 bits

Same-sign exact offsets needing >64 bits still report errc::overflow; the full fix (compute the rounded index in 128-bit like to_fixed) is on the roadmap. Regression tests in test_rational.cpp / test_storage_bugs.cpp.

Verification

428/428 ctest (unit, examples, compile-fail, amalgamate-up-to-date, codegen guard) · 418/418 under BND_MATH_FIXED · fuzz seed 42 × 20k iters clean (110M property checks) · clang-tidy clean under the CI header filter · amalgamation regenerated.

🤖 Generated with Claude Code