Bench the alpha-shape shadow filter against the unfiltered search

[meshinspector-agent]

Closed unmerged — and the table below is superseded, not merely stale. Do not read it as a verdict on the shadow filter. It was measured when every shadow test was exact, so it reads as the filter losing on the sphere and random clouds, which is no longer the case. The floating-point pre-filters are since restored — #6596 (wedge) and #6595 (query filter) — and they pay: on an 18820-point cloud at r=0.5, one binary per variant, interleaved, hashes equal, removing the wedge filter costs +30% and removing the query filter +41%, and the branch runs 4.46 s against master's 10.02 s. AlphaShapeData::shadowFilter never landed anywhere; DISABLED_AlphaShapeBench stays on the branch for per-variant numbers. See the closing comment.

The shadow filter is output-preserving, so switching it off is the master baseline: one binary now times both variants on the same runner, same cloud, same tree — the only way they are comparable on shared CI vCPUs.

• AlphaShapeData::shadowFilter — new flag, no public signature change.
• Min of 5 reps per variant, capped at 20 s of wall clock; cloud drops to 4k points where NDEBUG is off or under emscripten (MSVC iterator-debug, brew-llvm Debug and wasm share the 10-minute Unit Tests cap).
• Ball radius scales with each cloud's sampling step, so a smaller cloud is not a sparser one and the filter-on/off ratio stays comparable across configs.
• One greppable line per variant: [alpha-bench] <cloud> … filter=on|off reps= min= med= max= tris= hash= <counters>, plus a speedup= / inBallTestsRatio= line.
• Asserted, not just printed: the triangle hash and count must agree filter-on vs filter-off.
• Last commit drops the DISABLED_ prefix so every MRTest job of the next run reports numbers; it is restored before this branch reaches master.

First numbers — the agent's own self-hosted build box (8 vCPU, EPYC 7R13), not a GitHub-hosted runner (public repos get 4-core, macos-14 3-core); GCC 12.3 Release, 40k points, min of 5:

cloud	filter off	filter on	speedup	inBallTests off → on
sphere	40.8 ms	42.2 ms	0.97×	1.98M → 1.46M
grid	401 ms	348 ms	1.15×	6.42M → 3.06M
random	1412 ms	1695 ms	0.83×	62.2M → 23.7M

The filter removes the work it is meant to remove — 1.4–2.6× fewer ball tests, ~3× fewer considered triangles — but every shadow test is exact right now (54.5M of them on the random cloud), and that currently costs more than it saves except on the grid. That is the case for restoring the floating-point pre-filters, which the next increments measure the same way.

Verification

• Built in the ubuntu22 image mirroring build-test-ubuntu-x64.yml: prebuilt thirdparty linked as lib/include/share, clone_submodules_linux.sh --skip-prebuilt-thirdparty, Ninja Release, g++-12, -DMR_CXX_STANDARD=20. MRTest target and its whole dependency graph built clean.
• MRTest --no-python-tests --gtest_filter=*AlphaShape* → 4/4 pass, including the new AlphaShapeBench (19.8 s at 40k points; the table above is its output). AlphaShape, AlphaShapeSquare and AlphaShapeCrossingGrids are unchanged and still pass, so the flag defaulting to true keeps the current behaviour. --no-python-tests only because the mrbind modules are not built in this checkout.
• Not exercised here: the smaller non-NDEBUG / emscripten cloud sizes, and the other 18 MRTest configs — that is what the next run on this branch is for. Those run on 3–4 vCPU, so their absolute ms will be higher than the table above; the filter-on/off ratio and the exact counters are the comparable part.
• No UI surface (gtest-only harness plus one struct field), so nothing to drive in the viewer and no screenshots.

CI labels: none — this PR is based on alpha-shape-triangle-shadow, so it gets no checks of its own by design; the numbers come from the branch run after it lands there.

[Fedr]

Closing: the decision this harness was built to inform has been made.

The floating-point pre-filters it was meant to measure are back in alpha-shape-triangle-shadow
(#6596 for the wedge, #6595 for the query filter), and they clearly pay. Measured on a real 18820-point
cloud at r=0.5, one binary per variant, interleaved, hashes equal: removing the wedge filter costs +30%
and removing the query filter +41%, and the branch is 4.46 s against master's 10.02 s.

That also makes the table above misleading rather than merely stale — it was measured when every shadow
test was exact, so it reads as the filter losing on the sphere and random clouds, which is no longer the
case.

What remains is not worth its price: AlphaShapeData::shadowFilter is a public, Python-bound field whose
only use is benchmarking, since the filter is output-preserving and no caller has a reason to switch it
off. The branch keeps DISABLED_AlphaShapeBench for per-variant numbers.

[meshinspector-agent]

Agreed on all three — nothing to reopen.

• "The decision has been made; the pre-filters pay." — Confirmed: Reject most shadow candidates in double before the exact test #6596 and Restore the double query filter in FastInSphereTesterSoS #6595 are both squash-merged into alpha-shape-triangle-shadow (16 Aug, 10ae314a / 937df58b), so the harness has no question left to answer.
• "The table is misleading rather than merely stale." — Fair, and the framing was mine: measured with every shadow test exact, it reads as a verdict on the filter when it was really a verdict on the exact path. Annotated the body with a superseded banner carrying your numbers, and left the table itself in place as the record.
• "AlphaShapeData::shadowFilter is not worth its price." — Agreed, and dropping it costs nothing: the field only ever existed on this PR's branch. MRAlphaShape.h at the branch tip has no shadowFilter, just the counters — so there is nothing to strip.

One loose end that closes with it: the "restore the DISABLED_ prefix before this reaches master" follow-up is moot. The un-disabling commit lived only here, so the branch tip still reads TEST( MRMesh, DISABLED_AlphaShapeBench ) (source/MRTest/MRAlphaShapeTests.cpp:193) — no always-on bench ships to master, and per-variant numbers stay available the way you took them: one binary per variant, --gtest_also_run_disabled_tests.

No code changed, so nothing was built or driven for this reply.