Sweep test/** under --strict-uses and classify the failures

Track A built the enforcement ([[bug-pascal-uses-is-transitive]]) and measured it over examples/**: 55 programs, 41 compile today, and all 41 compile IDENTICALLY with --strict-uses on. The remaining unknown is the test/ corpus — 923 .pas plus the .npy suite — and that is a full-corpus run, which is why it is filed here rather than run in a dev loop.

What to run

For each file in test/** (.pas, .npy, and the C corpus if cheap), compile twice and compare only the pass/fail verdict:

pxx <file> /tmp/out                 # baseline
pxx --strict-uses <file> /tmp/out   # enforced

Nothing is executed and no expectation is checked. A file that already fails the baseline is not a finding — 113 of the 934 are helper *_unit.pas files and {%FAIL} cases that never compiled standalone. The finding is the set that compiles clean today and fails only with the flag.

How to classify each new failure

Three shapes account for every one Track A hit, and each has a different owner:

1. A compiler-INJECTED unit that is not marked ambient. The compiler injects builtin/builtinheap, pylib/pyeval (every .npy), and conditionally textfile/math/promocore/softfloat/pxxcio/palpthread. Tell: the missing name is a runtime helper the source never mentions — pyfloormod_i, pystr_of. Fix is one call site in the compiler (ParseUsesUnitAmbient).
2. Two filters that fail to compose. Tell: "no overload of X matches these arguments" with an EMPTY candidate list, rather than an undefined identifier. An out-of-scope candidate was still suppressing an in-scope one. Fix is in the compiler.
3. A genuine missing uses. Tell: the unit really does call another unit's routine without naming it. Fix is one clause in the library — or, as often, naming the RIGHT unit (pylib's FloatToStr came from builtin, not the sysutils it was accidentally reaching).

Report the count per shape, and for shape 3 the (importer, provider, name) triples. That list is the actual remaining work, and it is expected to be short.

Do NOT size this with --warn-uses-leak

The warn pass is noise-dominated: it fires on speculative lookups (FindUClass/FindProc called to ask "is this identifier a class?", answered yes, answer discarded), so its top rows are things like platform -> pylib [class] TPyDict at 2948 hits in a unit that never mentions TPyDict. Two earlier sizings of the parent ticket were wrong for exactly this reason. --strict-uses changes resolution, so it errors only where a name really binds — count failures, not warnings.

Why it is worth the run

--strict-uses becomes the default when this comes back short, which deletes the whole class of bug behind decide-class-namespace-scoping, the tkinter / reportlab Canvas collision, and every future "two libraries export the same name". The parent ticket is parked on this result.

RESULT 2026-08-14 — 1660 sources, ONE strict-only failure

Compile-only sweep, every test/*.pas and test/*.npy, twice each (baseline and --strict-uses), comparing only the pass/fail verdict. Nothing executed, no expectation checked.

1660 sources (1038 .pas + 622 .npy).

Shape 3 — genuine missing uses: NONE

The list this ticket asked for — (importer, provider, name) triples — is empty. No library needs a uses clause added. Combined with Track A's examples/** result (41 of 41 identical), the corpus says the enforcement is essentially already satisfied.

The one finding, and it is not cleanly any of the three shapes

test/test_nilpy_dotted_package_import.npy:

pascal26:3928: error: __pxx_pipe2 needs the thread-safe runtime: rebuild with
--threadsafe (<pthread.h> lowers onto the pxx thread PAL, which that flag
selects; without it this would import a pxx-internal symbol from libc and fail
at load)
  near:       >>> unsigned int sleep

Closest to shape 1 (a conditionally-injected ambient unit): palpthread is marked ambient only under --threadsafe (compiler/cparser.inc:8611), so under strict the pthread path is entered where the baseline never needed it. But it surfaces as a capability diagnostic, not an undefined name, so it does not match shape 1's stated tell. Recording that rather than forcing it into a box.

Under --strict-uses this file has no working configuration, and the reason is the second finding below rather than anything about uses:

So the remedy the diagnostic names is itself broken — see [[bug-a-threadsafe-segfaults-on-every-nilpy-program]]. That is the actual blocker, not a uses problem. Once --threadsafe works for NilPy, this file should be re-checked; it may well pass, or need only the ambient marking.

Recommendation for [[bug-pascal-uses-is-transitive]]

The sweep came back short, as that ticket hoped. One file, no shape-3 work, no library changes. The remaining question is whether that one file's shape-1 marking is worth doing before flipping the default, or whether the flip should wait on the --threadsafe fix so the file can be re-measured honestly. That is Track A's call — T's job was the number, and the number is 1.

Method note

Run detached under systemd-run --user, writing each verdict to a TSV as it lands and skipping what the TSV already holds on restart. The first attempt buffered everything to the end and was killed by its own 30-minute timeout, losing ~1700 completed compiles for nothing.

Log

• 2026-08-14 — resolved, commit da498b969.