Reject most shadow candidates in double before the exact test

[meshinspector-agent]

Supersedes #6593, which sits on closed #6592's branch and needs toDouble. This one is based on alpha-shape-triangle-shadow and needs nothing new — it recomputes the pre-filter's operands in double instead of converting the exact bignums.

• The port to FastInt dropped the pre-filter, so every candidate reached the exact path. Restoring it skips 79–91% of them.
• up/uq are hoisted out of the candidate loop; bp = dot( d, up ), bq = dot( d, uq ), and a negative sign of either rejects the candidate before any bignum work.
• Why no toDouble. The operands must approximate the value on the integer grid, where the predicates and the SoS ties live. Point differences are below 2^31, so Vector3d of one is exact and pp/qq/pq are ordinary double dot products — 0.7 bits worse than converting the exact bignum (max error 2^10.7 vs 2^10.0 over 200k pairs), against 2^33.3 if taken from the cloud's Vector3f.
• Error bound: those dot products are below 2^64, the components of up/uq below 2^96, their dot with d below 2^129; each rounding costs 2^-53 of the running magnitude, so the total stays below 2^79 and a double below -2^79 is negative exactly as well. The 2^84 tolerance keeps 32x that margin; candidates in between fall through to the exact path unchanged.
• Consequence: no dependency on Add toDouble for the FastInt family #6594, so this can land on the branch on its own.

The exact path itself is untouched — this only decides which candidates reach it.

Verification

Built in the ubuntu22 image mirroring build-test-ubuntu-x64.yml: clone_submodules_linux.sh --skip-prebuilt-thirdparty, prebuilt thirdparty linked as lib/include/share, Ninja Release, g++-12, -DMR_CXX_STANDARD=20. MRTest and its dependency graph — 640/640, no new warnings.

MRTest --gtest_filter=*AlphaShape*:*PrecisePredicates*:*InSphere*:*FastInt* → 16/16 pass.

DISABLED_AlphaShapeBench, pre-filter on:

cloud	triangle hash	shadowTests	exactShadowTests	rejected
sphere 40k	17403291826481596167	487416	54994	88.7%
grid 40k	635805191690857954	2058881	186448	90.9%
random 40k	11472133144093973002	54564333	11466269	79.0%

Every hash equals the filter-off baseline, and consideredTris, touchableTris, inBallTests (1461420 / 3055058 / 23736199) and shadowedNeis are unchanged — the pre-filter is output-preserving, as its bound claims. All of those, and all three hashes, are also identical to #6593's toDouble version: on these clouds the cheaper operands decide every candidate exactly the same way.

No timings on purpose — this box is not a GitHub-hosted runner, so its milliseconds are not comparable to the reported set. The counters above are exact and platform-independent; the wall-clock payoff comes from the branch run.

Not exercised: the other MRTest configs (MSVC, arm64, wasm, macOS) — a PR based on alpha-shape-triangle-shadow gets no checks of its own. No UI surface (library-internal predicate, no MCP-exposed tool and no viewer widget), so nothing to drive over MCP and no screenshots; the three structurally different clouds are the state coverage instead.

CI labels: none.