Deterministic ephemeris: vendor musl transcendentals, pin FP contraction

[heirloomlogic]

Closes #28.

Problem

Ephemeris positions were not bit-reproducible across environments: the vendored astronomy.c called the host libm's transcendentals (sin ×137, cos ×97, atan2 ×14, …), whose last-ULP results differ between macOS point releases (26.4 vs 26.5.2) and vs glibc. Positions drifted by sub-arcminute amounts per environment, silently flipping ~6 borderline cases in AstrologyKit's expert-locked calibration corpus and making hosted CI impossible to keep green.

Fix

Vendor the math, pin the codegen, lock it with bit-exact tests.

• Sources/CLibAstronomy/detmath/ — musl 1.2.5 double-precision subset (sin cos tan asin acos atan atan2 exp log10 pow cbrt hypot + kernels: __rem_pio2, __rem_pio2_large, scalbn, exp/pow tables, error helpers). Every symbol is ak_-prefixed (including internals) so nothing can collide with or resolve to the host libm at link time. musl attribution preserved; MIT license added to THIRD_PARTY_NOTICES.
• Local patch Fix lint warnings, Swiftlint and README updates #6 to astronomy.c (one include line, documented in the vendoring note): ak_detmath.h redirects the 12 transcendentals to ak_* and opens with #pragma STDC FP_CONTRACT OFF, pinning FMA contraction off for the whole translation unit. Build logs confirm Xcode defaults to -ffp-contract=on, so the pragma is load-bearing. Source-level on purpose — unsafeFlags would make the package unconsumable as a dependency.
• pow.c #undefs __FP_FAST_FMA — musl's pow has an FMA path that is compiled on Apple Silicon but not x86; exactly the cross-arch divergence this PR eliminates.
• Swift layer: the 3 transcendental call sites (MoonPhase.illumination's cos, FixedStar's atan2/asin) now call ak_* via the new public include/ak_math.h. sqrt/fmod/floor/ceil/fabs deliberately stay on the host — they are IEEE-exact/correctly rounded everywhere and carry no drift risk.
• ReproducibilityTests — 158 bit-exact golden assertions (Double(bitPattern:), exact ==): ecliptic lon/lat/dist for 10 bodies at 4 instants (1980/2000/2026/2050), equatorial + topocentric coordinates, sunrise/sunset and full-moon search times, FixedStar and Moon.illumination Swift paths, Chiron. Identical bit patterns in debug and release. These running green on both the macOS and Ubuntu CI jobs is the cross-environment proof from the issue's acceptance criteria.
• CI nm guard on the Linux job: fails if any built object regains an undefined reference to a host transcendental (so a future call site can't silently escape the redirect).
• Docs: MAINTAINING.md gains patch Fix lint warnings, Swiftlint and README updates #6 in the re-apply checklist plus a "Deterministic math" section (musl update procedure, no-unsafeFlags constraint, Windows note: Swift-on-Windows compiles C via clang, which honors the STDC pragma).

Accuracy

Unchanged at the meaningful scale: all JPL Horizons validation suites (±1–1.5′) and audit validation suites pass untouched. The change is last-ULP only.

⚠️ Downstream impact (one-time)

Vendored math shifts last-ULP values relative to every current host libm, so AstrologyKit must re-baseline its calibration corpus once against this version. After that, the baseline holds permanently on every OS, runner image, and toolchain — which is the point.

Verification

• swift test (debug): 579/579 in 175 suites ✅
• swift test -c release: 579/579, same golden bit patterns ✅
• swift test --sanitize=thread: 579/579 ✅
• swift-format lint --strict (CI config): clean ✅
• nm audit: zero undefined host-transcendental refs in astronomy.o and detmath objects; no fma ✅
• CI on this PR: Ubuntu (glibc, x86) and macOS (arm64) jobs asserting the same 158 bit patterns = the cross-OS/cross-arch reproducibility proof.

🤖 Generated with Claude Code