feat(utils): port the interp and boost kernels to rust

[LoganAMorrison]

Summary

• Ports np.interp and the four live routines of hazma/_utils/boost.pyx — boost_beta, boost_gamma, boost_delta_function, boost_integrate_linear_interp — to PyO3-free Rust (rust/src/{interp,boost}.rs), with hazma._core.{interp,boost} registration-only probes so the tests can reach them. No kernel is swapped: nothing under hazma/ imports either module, the only change under hazma/ is a comment in hazma/_core.pyi, and the parity corpus still runs in bit-equality mode (rtol = 0, provenance → exact=True).
• The port reproduces the shipped Cython's fused multiply-adds, and that is load-bearing rather than pedantic. Clang defaults to -ffp-contract=on; the corpus's capturing platform (macOS/arm64) contracts eight distinct expressions across boost.pyx, and NumPy's arr_interp contracts too. Written the obvious unfused way the port misses the corpus by up to 3.6e-12 relative on the corpus's own grids — past the 1e-12 TABULATED budget in test/parity/tolerances.py — so the Phase 04 swap would have failed its own gate. With f64::mul_add at those sites it is bit-equal at every one of those points. The converse is the trap: boost_beta is deliberately not fused, because none of its ten inlining call sites contract 1 - (m/E)**2. Each site was established twice, by disassembling the shipped .so and by bisecting all 16 on/off combinations against the live kernel.
• No values move in this PR, but one drift is now known and lands with the Phase 04 swap: the Rust returns the contracted (arm64) numbers on every platform, so on a target whose C compiler does not contract — baseline x86-64, which is what the Linux wheels are built for — today's Cython returns values up to 3.6e-12 relative away from these. Past rules.md rule 3's 1e-12 threshold, so it is recorded in the project's "Numerical impact so far" for the Phase 04 PR to declare. The alternative, plain arithmetic, misses the corpus by that same amount on every platform.
• Found while porting: boost_integrate_linear_interp is wrong near threshold, in released hazma. Its interior sum stops one cell short of the upper partial cell, and when both integration bounds fall inside one cell the two partial-cell terms overlap instead — an over-count of cell width / window width, which diverges as beta → 0. All seven tabulated photon spectra therefore blow up instead of converging to their own rest-frame spectrum as the parent slows: dnde_photon_eta(54.79, m_eta) is 0.02313 MeV⁻¹ at rest against 767.2 one part in 1e12 above it, a factor of 33,000; the other six run 6,500×–9,800×. This PR reproduces the behavior rather than repairing it (rules.md rule 1 — the corpus pins these values, so a fix here would fail the gate), pins it in both languages, and files the repair, blocked until after Phase 06 Task 6.4 because it needs a declared corpus regeneration. Worth a look independently of this project's schedule.
• Two canonical patches, in the same PR as the code: the phase file's Task 3.4 block gains five "criteria added during execution" bullets (the oracle it named does not exist — Phase 01 captures top-level defs and these are all cdef, so the tests call the live Cython through __pyx_capi__ with ctypes instead), and references/numerics-replacements.md gains the measured block, since its own np.interp and boost prose is what a next reader would port from and both are incomplete in ways that change the numbers.

Project

projects/cython-to-rust/ — Task 3.4: Interpolation + boost kernels.

See projects/cython-to-rust/task-notes/phase-03/task-3.4-interp-boost.md
for implementation detail, decisions, the mutation tables, and verification.

Test plan

• scripts/agents/preflight.sh --paths "…" --md "…" → RESULT: PASS (ten rows PASS, one expected SKIP for the version bump)

PASS   black --check / isort --check-only / ruff check
PASS   cargo fmt --check / cargo clippy / cargo test        rust/
PASS   pytest                  1314 passed, 13 skipped, 5 warnings in 566.77s (0:09:26)
PASS   import hazma            version 2.1.0
PASS   markdownlint            (8 changed .md)
SKIP   version bump            not a closing PR (pass --closing)
PASS   forbidden tokens        none added
RESULT: PASS

• +102 on Task 3.3's 1212 passed, 13 skipped, all of them this task's tests. The skip count is unchanged at 13, which is how the parity suite reports it is still in bit-equality mode — confirmed directly rather than inferred:

$ python -c "...tolerances.provenance(manifest)..."
Provenance(exact=True, detail='')
$ python -c "...cases.rust_core_kernels()..."
[]

• pytest test/test_core_interp.py -q → 33 passed in 0.46s (6 classes); pytest test/test_core_boost.py -q → 69 passed in 0.91s (9 classes)
• cargo test --manifest-path rust/Cargo.toml --no-default-features → 67 passed (24 new)
• Bit-equality against the oracles, not a tolerance: zero mismatches on all seven live tables across six boost regimes × 400 energies, zero across 40,000 boost_delta_function draws at both live product masses, and zero on 20,204 np.interp abscissae per table.
• Test validity — 21 mutations, run sequentially behind a lock with a green baseline asserted before and after. 17 of the first 20 caught; all 21 after two tests were added. The three survivors shared one shape — each moved a branch boundary by a single double without touching any returned value, which no grid sample can see. The tests written for them bisect on the bit pattern. Tables in the task note.
• No public value changes: git diff origin/master -- hazma is one file, hazma/_core.pyi, and every line of the hunk is comment text.