Precision: nonlocal write-back and a broader regression corpus #71

Improve precision and broaden the corpus that guards it.

• nonlocal writes flow back to the enclosing function (the last documented approximation).
• Add public repositories of other kinds to tests/regression/repositories.toml (async web, workers, data pipelines, larger CLIs) and review every new finding against the code.
• Fix the false positives and false negatives the new repositories reveal.

Gaps found while measuring on large projects (#69), to close here:

• Frontend: NamedExpr (walrus), yield from and starred assignment targets reject the whole file as a syntax-error (17 files in healthchecks, 22 in rich, 57 in wagtail); support them so those files are analysed.
• Secrets: pure-hex hashes under non-credential keys in JSON data files (commit ids, checksums) flood high-entropy-string (5 500 findings in rich's benchmark results, 1 900 in wagtail); skip hex digests unless the name is credential-like.
• Secrets: hardcoded-credential in test fixtures (37 of 39 in healthchecks, most of wagtail's 148) should be reported at low confidence or skipped under test directories.
• Django: redirect("url-name", tainted_arg) resolves a URL pattern and is not an open redirect; only the first argument is a redirect target (77 findings in healthchecks, 31 in wagtail).

Gaps found by the broadened corpus (dvpwa, pygoat, full-stack-fastapi-template, httpie, luigi):

• Frontend: class keyword arguments (class Task(Base, metaclass=Register)) and with targets other than a single name reject whole files (12 files in luigi, 2 in httpie, 1 in the FastAPI template).
• Sources: environment and process-output are operator-controlled, not attacker-controlled; like argv they should not carry COMMAND and PATH (httpie daemons, luigi contrib and tests).
• aiohttp models are missing, so dvpwa's documented SQL injections and XSS are silent (moved to Detection coverage: syntactic rules and models for aiohttp, Tornado, Bottle and database drivers #70).