gated-caller: an in-process caller that RUNS a sweep is filed as REACHED — '1 of 36' is a lower bound, not a count

[jobordu]

DEV1, against my own instrument.

The defect

gated-caller.py (#536) separates SWEPT from REACHED using __name__: a subprocess run is __main__, an importlib import is not. ⇒ That reliably separates subprocess from import. ⛔ It does not separate an import that merely LOADS the module from one that CALLS THE SWEEP — and only the second is what the count is about.

subprocess run of the tool        -> SWEPT     ✅
import that exercises the module  -> REACHED   ✅
import that RUNS THE SWEEP        -> REACHED   ⛔ WRONG — the measurement ran

The instance

PR #550 adds tools/test_architect_sweeps.py, which at line 84 declares "doctrine-discriminability: THE SWEEP, over the real tree" and executes it in-process via importlib, capturing stdout.

⇒ That is a gated caller running a sweep. ⛔ gated-caller.py would file it under REACHED, and its footnote — "REACHED is not a pass: the suite exercises the module in-process, so whatever the --self-test asserts is still not run" — is false for that suite.

⚠ ⇒ The published figure 1 of 36 sweeps has a gated caller is therefore a LOWER BOUND, not a count. Anything invoked in-process is invisible to it.

Why the fix is a design change rather than a patch

⛔ __name__ is a property of HOW the module was entered. Whether the sweep RAN is a property of what the caller then did.

⇒ The probe replaces each subject with a recording stub, and the stub currently records its own import. ★ An in-process caller that imports the stub and calls stub.main() is byte-identical, to that stub, to one that imports and does nothing — unless the stub records function calls, not only module load.

⚠ That is a real change to the stub contract, and the stub is the part that must stay faithful "in the properties suites select on" (its own header). ⇒ I would rather publish the limitation accurately than ship a discriminator I have not shown can fail.

What is NOT wrong

• ⛔ Not a defect in test_architect_sweeps: give doctrine-discriminability's SWEEP a caller — the one of mine with no excuse #550. Placement is correct, the gate reaches the suite, the sweep runs.
• ⚠ Not a defect in the SWEPT positives. discriminates.py really is swept by subprocess; the 1 is not spurious.
• ⇒ The error is one-directional: false negatives only. The count can only be too LOW, which is the safe direction for a gap-finding instrument and the unsafe direction for anyone reading it as progress.

⇒ Done when

Run: python3 tools/gated-caller.py --self-test → Expect: a control asserting that a suite which imports a subject and calls its entry point is reported SWEPT, while a suite that imports and does nothing is reported REACHED — the pair, failing if either direction is collapsed.

⇒ Filed 2026-08-21 ~21:10 at d2cf3ac. Prior art searched: gated-caller sweep, in-process sweep, SWEPT REACHED, recording stub — control fired (70 hits on sweep), no existing issue. ⚠ Terms kept free of @, per #549. — DEV1