What a dummy argument may say: local-parameter bounds, lower bounds, generics, where masks, elementwise logicals

[chenyueqi]

Stacked on #22. Closes #17, #18, #19, #24. Found by the CLUBB extension's tier-0 units.

• The flat adapter declares dummy dims with the unit's own module parameters without importing them #18 A dummy's bound naming a local integer parameter of the subprogram (weights_zt2zm(ngrdcol, nzm, t_above:t_below)) is folded to the value in the interface record, recorded under folded_bounds. The wrapper compiled with an undeclared name before; the sampler had no table for it.
• A dummy with a non-unit lower bound gets the upper bound as its extent, on both sides of the gate #24 An axis declared lb:ub keeps its extent on both sides: the flat adapter and the plain wrapper declare lb:ub, the bit-exact gate sizes ub - lb + 1. lhs(-2:2, ngrdcol, ndim) has five rows, not two.
• A public generic interface whose specifics are private leaves nothing to gate #17 A specific of a public generic is public through it (public_via); the flat adapter uses, calls and re-exports the generic, and the adapter module's own wrappers call it through the generic too, the only name the module lets out.
• static.rwset counts the emitter's where-mask temporaries as target-only names #19 _wn and _we<d>_<n> join _wm on the read/write check's discard list.
• Logical operators on array operands are elementwise (~, &, |) outside a WHERE too: any( .not. l_ok ) over a logical array was emitted as Python's scalar not.

Full suite green. CLUBB tier 0 goes from 8/14 to 12/14 bit-exact on this branch (mean_adv, tridiag_lu_solvers, penta_lu_solvers, bicgstab_solvers; t_in_k_module gains its three generic specifics, new_pdf eight of nine subprograms). The two left are #20, complex values.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Cne4hrGgYqhTMzVgzJtXYd

[chenyueqi]

Review of #25 against two questions: generic or CLUBB-specific, and any change to an existing rule.

Genericity: pass. All five items are Fortran-language or engine-internal rules. _wm/_wn/_we<d>_<n> are the numpy emitter's own WHERE-mask temporaries (statements.py, already on main); the discard regex simply lagged the emitter, and the new test checks the negatives. CLUBB appears only in docstrings. (The KINDS_64 domain-name list this branch carries comes from #22; commented there.)

Existing rules: no test modified or deleted, all six test files are additions. lb:ub sizing agrees with the translation's existing lower-bound handling (rules/indexing.py _origin, subprograms.py local sizing, all untouched). public_via flows to every consumer of public (modules.py, numpy_anchor.py, flatten.py, bitexact.py), which is the right semantics but wider than the description says. Two gaps inline: the bound fold is partial for symmetric -nd:nd bounds and nothing refuses the leftover, and _array_valued falls back to the scalar spelling when rank cannot be settled (main's behavior, but the description does not say the new rule is not fail-closed). folded_bounds is recorded in the facts JSON only; a line in reading-the-evidence.md would make it findable.

[chenyueqi]

Only a bare name folds, so dimension(-nd:nd, ...) becomes {"lb": "- nd", "ub": "3"}; the new test's lhs assertion documents exactly this and its comment says the leftover is "the wrapper's to refuse", but nothing refuses it: f2py.py::_extent spells - nd:3 into the wrapper verbatim (gfortran undeclared-symbol build failure, not an engine refusal) and bitexact.py harvests nd from the lb text as a free extent to redraw. -n:n is common Fortran (stencils, band matrices). Not a regression against main, but the description's "folded to the value" is half true for symmetric bounds. Suggest folding a negated name too, or refusing when one bound of an axis folded and the other did not.

[chenyueqi]

Swallowing Unanalyzable here emits Python's scalar not/and/or for an operand whose rank the engine could not settle, and semantics.rank returns 0 for an undeclared Name without raising, so a whole-array logical use-imported from another module never reaches this branch at all and is spelled not x. At runtime that raises for len>1 arrays (loud) but not np.array([True]) is silently False. This is main's behavior, so the PR strictly improves coverage, but the rule is not fail-closed and the description does not say so; the docstring's "as before" is the only mention.

[chenyueqi]

The places where this branch turns a refusal on main into a silent fallback are recorded apart from the review, in #32 (rows 5-6), so each is ruled on rather than merged by omission. A row there closes when the rule refuses again, records what it assumed in the evidence, or takes the assumption from the extension's configuration.