[electricessence]

Make the two allocation tests deterministic by reading their rows from the existing TypedReader instead of a DataTableReader, which boxed an int per row and made them fail at random.

The fixture read ints through DataTableReader.GetInt32, which allocates a 24 B box per row. Tiered JIT removes that box part way through a run, so when it lands between the two back-to-back measurements of MappedQuerySelectTests.TheAsyncProjection_AllocatesNothingPerRow or SingleValueAllocationTests, one side is 24 B per row cheaper and the comparison fails; the library allocates nothing per row.

What changed

• Ints (in MappedQuerySelectTests) and Rows (in SingleValueAllocationTests) build a TypedReader (TypedReader.Create with ColumnDef columns) instead of a DataTable and DataTableReader. The cells are boxed once, when the rows are built, outside the measured region; the typed getters only unbox them, so a row costs nothing.
• Thresholds, slack, warm-up and the assertions are unchanged. Nothing under Source/ is touched.
• The branch has two commits: the first added a purpose-built ArrayReader; the second removes it in favor of the TypedReader the test project already has (the net result is only the two helper changes).

Verification

• Full Core test project, net11.0 Release, 10 runs: 10 of 10 passed, 10200 tests each.
• TheAsyncProjection_AllocatesNothingPerRow, net11.0 Release, filtered, 10 runs: 10 of 10 passed.
• SingleValueAllocationTests, net11.0 Debug, class alone, 10 runs: 10 of 10 passed (12 tests each).
• Full project once each: net11.0 Debug, net10.0 Release, net10.0 Debug: 10200 passed each.
• Measured per row (1,000-row ToListAsync, plain vs projected): 8.0 B per row for both (8712 B plain and 9032 B projected at 1,000 rows; 1472 B and 1792 B at 100), the same filtered, with tiered compilation off, and inside the full run. The projection adds a flat 320 B per call and the 24 B per row gap is gone.
• The guard still guards: with a temporary per-row allocation in the projection path, TheAsyncProjection_AllocatesNothingPerRow fails ("costs 2720 B per call over 100 rows and 24320 B over 1000"); with one in the single-value mapper, 4 of the 12 SingleValueAllocationTests fail (for example "ToArray over 100 rows: 2864 B per call through the terminal, 488 B by hand"). Both changes were reverted and never committed.
• Release build and style/analyzer format check clean.

Not done

• RowMapConverterTests.AConverter_AllocatesNothingBeyondItself_PerRow still reads through a DataTableReader and compares bytes of two calls (a converter against none), so it has the same exposure in principle; it did not fail in any run here. AllocationQuickWinTests also uses DataTableReader but only as a positioned record.
• No change to CI; enabling the tests there is a separate step.

🤖 Generated with Claude Code