test(suite): declare the interp oracle's platform scope

[LoganAMorrison]

Summary

• test/test_core_interp.py scoped its NumPy oracle with a probe, and that probe was resolving the wrong way. Built for linux/amd64 it returned False, so all nine of the module's cross-implementation claims skipped and hazma._core.interp was checked against nothing but its own clamping contract, quirks and error paths on every CI entry but macOS. Same mechanism Task 4.1 retired in test_core_positron_muon.py and the 2026-08-12 rewrite retired in test_core_boost.py. The module now runs 42 passed, 0 skipped on both platforms (was 24 passed, 9 skipped off macOS). Resolves interp-oracle-scoped-by-an-unsound-probe.md.
• The mode is declared, and the budget was measured. ON_THE_CAPTURING_PLATFORM is read from test/parity/data/manifest.json so this module's scope cannot drift from the corpus's. Off it, the tree was built for linux/amd64 (Debian bookworm, glibc 2.36, CPython 3.12.13, NumPy 2.5.1) and compared directly over 1,154,010 abscissae: the port and that build's np.interp differ at 311,501 points by up to 4.0e-02 pointwise relative but only 2.2e-16 of the peak, with no non-finite value on either side (0 of 2,308,020) and no disagreement anywhere about which abscissae return exactly zero. Hence OFF_PLATFORM_BUDGET = 1e-12 scaled to the peak — two orders tighter than the boost module's 1e-10, not looser as the follow-up guessed, because peak scaling collapses the very cancellation population the pointwise reading blows up.
• TestFusedArithmetic no longer discriminates against NumPy. Its reference is interp_reference, a full Python transcription of rust/src/interp.rs (bisection, clamps, exact-node short circuit, NumPy's two-step NaN rescue) parameterised by its multiply-add, asserted bit-for-bit on every platform. That makes it the stronger of the two off-platform gates, since peak scaling is blind to a defect confined to a value far below the peak. The discrimination is direct, not assumed: np.interp is bit-equal to mul_add=fma on all seven live tables on macOS/arm64 and bit-equal to mul_add=unfused on all seven on linux/amd64. TestOffPlatformBudget guards both the tolerance and the mode dispatch.
• Two supporting fixes the measurement forced. rng.uniform(lo, hi, n) is not platform-independent — NumPy computes lo + (hi - lo) * u in C and macOS/arm64 contracts that, so 6,532 of 20,000 draws over the eta grid landed on different doubles than the same seed gives on Linux; the new spread() helper writes it as separate ufunc calls. And the per-table seed moved from hash(name) (randomised per process unless PYTHONHASHSEED is set) to zlib.crc32. Without both, every count recorded here would have been a per-platform anecdote rather than a reproducible measurement.
• rust/src/interp.rs is comment-only (verified: git diff origin/master -- rust/src/interp.rs filtered to non-//! lines is empty). Its module docs carried the now-measurably-wrong claim that an unfused NumPy "would differ from this function by ≤1.1e-13 — an order inside the 1e-12 budget test/parity/tolerances.py sets"; the pointwise figure is 5.9e-05 on the live tables, and the corpus is skipped off-platform anyway.

No public value moves. The diff is one test module, comment-only Rust edits, and four docs; no public function changed.

Project

projects/cython-to-rust/ — not a PLAN task. This resolves the
cross-cutting follow-up
interp-oracle-scoped-by-an-unsound-probe.md,
which was carved out of the test_core_boost.py probe removal
(1ed4eda), so the branch is ad-hoc (claude/<short-description>) rather
than claude/<project-slug>/<task-slug> and there is no
task-notes/task-N-<slug>.md for it. The follow-up file carries the
Resolution, Verification and stale-state sweep that a task note would.

The module under change was built by Task 3.4 (projects/cython-to-rust/task-notes/phase-03/task-3.4-interp-boost.md,
shipped as #61); this PR supersedes that task's platform-scoping remedy
and patches the two inbound links plus a sweep-size slip in its records.
It does not reopen Task 3.4, whose status stays Complete.

Test plan

• scripts/agents/preflight.sh --paths "test/test_core_interp.py" --md "…" — RESULT: PASS, all eleven rows:

PASS   black --check           test/test_core_interp.py
PASS   isort --check-only      test/test_core_interp.py
PASS   ruff check              test/test_core_interp.py
PASS   cargo fmt --check       rust/
PASS   cargo clippy            rust/
PASS   cargo test              rust/
PASS   pytest                  1444 passed, 14 skipped, 5 warnings in 594.22s (0:09:54)
PASS   import hazma            version 2.1.0
PASS   markdownlint            docs/followups/README.md docs/followups/done/… projects/…
SKIP   version bump            not a closing PR (pass --closing)
PASS   forbidden tokens        none added
RESULT: PASS

• pytest test/test_core_interp.py -q on macOS/arm64 → 42 passed in 0.94s

• the same command inside the linux/amd64 container built from this tree → 42 passed in 2.61s

• cargo test --manifest-path rust/Cargo.toml --no-default-features → 80 passed

• Test validity, stash-proof. slope.mul_add(x - xp[j], fp[j]) in rust/src/interp.rs replaced with slope * (x - xp[j]) + fp[j], rebuilt with pip install -e ., module re-run on both platforms:

  • macOS/arm64 → 15 failed, 27 passed (all seven TestAgainstNumpy table cases, the random grid, all seven TestFusedArithmetic cases)
  • linux/amd64 → 7 failed, 35 passed (the seven TestFusedArithmetic cases only)

  That second line is the point in both directions: off-platform TestAgainstNumpy passes the mutant, because there np.interp is the unfused form — the budget genuinely cannot see this defect, and the fused reference is what catches it. Under the old module the off-platform count would have been 0 failed. Mutation reverted and the tree rebuilt before the gate was run.

The 14th skip is test/parity/test_parity.py in budget rather than bit-equality mode (hazma._core serves 1 kernel(s)) — Task 4.1's positron-muon swap, present on master before this change and untouched by it.

• Confirmed on CI, which is where the skips lived. Linux py3.12, pytest --ignore=test/parity:

  run	result
  master 707b07c (run 31646076820)	797 passed, 22 skipped
  this PR (run 31651543376)	815 passed, 13 skipped

  −9 skipped, +18 passed — the 9 are exactly the module's previously-voided cross-implementation claims, and the +18 is those 9 plus the 9 net-new tests (33 → 42). All eight checks green, including all five Ubuntu entries py3.10–py3.14 and macOS py3.14.