An imported module's ** is computed by the fallback, not by the RTL's Power

Found while running lekkerzeilen, which is how the severity got measured correctly: the symptom was a crash, the crash was somebody else's bug, and what is left here is a last ulp.

Defect 1 — the scan runs before the imports

compiler/pyparser.inc:41693:

  pyWantsPow := False;
  for i := 0 to TokCount - 2 do
    if (Tokens[i].Kind = tkStar) and (Tokens[i + 1].Kind = tkStar) then

PyParseImportRun is at 41769 and the install at 41796. So the scan answers about the main module's tokens only.

A contained fix is possible at 41796, because by then the imports have been parsed and the tokens exist — set pyWantsPow there too rather than moving the scan. What makes it not purely mechanical is the ParseUsesUnit('math') pull that sits INSIDE the scan at 41708: the comment above 41713 records that math must be pulled BEFORE ParseUsesUnitAmbient('pylib') because the last unit named wins a name and math's integer Abs otherwise hides pylib's float one (measured: abs(-1.5) stopped resolving). So the scan cannot simply move down, and the install site cannot simply pull math.

Defect 2 — unit initialisation runs before the main body

The assignment is emitted as the program's FIRST statement (41796 onward, mainNode := PySeqAppend(mainNode, powAsgn)). An imported .py is a unit, and Pascal runs unit initialisation sections before the main body. So a module-level ** in an imported module executes with PyPowHook = nil regardless of defect 1.

Control, and it is what makes this a second defect rather than a restatement of the first — predicted before running that a ** in the main module would fix the imported one; it did not:

CPython gives ...95. Both rows are from one program with ** in main, so the hook WAS installed; only the ordering separates them.

What the cost is, stated honestly

One ulp on a ** in an imported module. The hookless path is PyMathExp(e * PyMathLn(b)). It is not a crash and not an F-lane ticket by the "rank the mechanism, not the datatype" test — the mechanism is a hook that does not get installed — but the OBSERVABLE is a last digit, so do not rank it as though it were the ValueError that led here.

The regression fixture already exists and pins the FALLBACK, not this

test/test_nilpy_pow_zero_base_in_an_imported_module.npy (+ its aux module test/nilpy_pow0_mod.py) deliberately contains no ** of its own, because a ** in main would retire its four post-start rows. It therefore passes both before and after this ticket is fixed — it measures pypow_cx's arithmetic, which is correct now. Fixing this ticket makes four of its rows stop reaching that code at all; the import-time rows will still reach it until defect 2 is also fixed. Anyone fixing either half should add a row asserting the ...688795 value, which is the only thing that can observe the hook arriving.