INFR: Coverage measurement excludes all numba-jitted code and counts the test tree #871

Problem

.coveragerc lists @jit, @jit\(.*nopython=True, @njit, @overload and @guvectorize\(.*nopython=True under [report] exclude_lines. The intent is sound — jit-compiled code is not traced by CPython, so its lines would otherwise read as uncovered. But coverage.py excludes the entire decorated function when a decorator line matches an exclusion pattern, not just the decorator line. The result is that the hottest, most algorithmically intricate code in the library is not merely uncounted — it is unmeasurable.

Measured by walking the AST and comparing against what coverage report attributes to each file:

The entire quantecon.optimize subpackage — 2,396 LOC of numerical optimisation — reports 100% while measuring almost nothing. The code is not untested (it has 1,052 LOC of tests); it is unmeasured: no instrument can tell which branches of the numerical kernels are exercised, or detect regression in that coverage.

A second, independent defect compounds it: [run] source = quantecon includes quantecon/**/tests/, so the test files count toward the percentage. CI computes and Coveralls displays 94%; source-only coverage is 87%.

Proposed fix

Run the suite twice in CI: once normally, and once with NUMBA_DISABLE_JIT=1, which makes jitted functions execute as ordinary Python and become fully traceable; combine with coverage combine. This is the standard approach for numba projects and requires no library-code changes. At the same time, narrow source/omit so the test tree is excluded from the headline figure.

Expect the reported percentage to fall when this lands — that is the point; the new number will be true and becomes the baseline for future coverage work.

Acceptance criteria

• CI runs the suite a second time under NUMBA_DISABLE_JIT=1 and combines the results
• The jit-decorator patterns are removed from exclude_lines, or scoped to the non-instrumented pass only
• source/omit excludes quantecon/**/tests/ from the reported figure
• The new baseline percentage is recorded in this thread
• quantecon/optimize/* reports statement counts consistent with AST statement counts

From the July 2026 technical-debt audit (AI-assisted; claims verified against 28d4b3b on 2026-07-25).