#1565 captured a function passed by name as a call argument
(executor.submit(fn), Thread(target=fn), map(fn, xs)) only when the
callback was defined in the SAME file — it resolved through the in-file
label map. But the dominant real-world shape is cross-module: the callback
is imported (`from .handlers import on_event; pool.submit(on_event)`), so
the same-file map can't see it and the edge was dropped — exactly the
caller a blast-radius query must not miss.
When the argument identifier isn't defined in-file (and isn't shadowed by
a param/local — that guard already ran), emit an `indirect` raw_call and
let the existing cross-file resolution pass handle it, branching to a
distinct indirect_call/INFERRED edge instead of calls. It rides the same
single-definition god-node guard and import-evidence disambiguation as
direct calls (parity: when a direct call resolves, so does the indirect
one; when the name is ambiguous, both bail). Two added safeguards: the
target must be a real callable def (per-file callable_nids unioned across
the corpus) so an imported data constant can never become a dispatch
target, and an existing direct `calls` edge for the pair pre-empts it.
Smoke-verified: imported single-def callback resolves; ambiguous name
bails (same as direct); imported data constant rejected; direct+indirect
to the same target keeps only the direct edge; cross-file keyword
target=cb resolves. Full suite 2710 passed.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
safishamsi
cf747abeaf