[professor-moody]

is_likely_safe() previously based its answer on the first pickle stream. This change checks stacked streams incrementally and returns False as soon as a stream has an invalid opcode or is not rated LIKELY_SAFE. It does not parse trailing streams after that decision or collect the streams in a list.

This also addresses the eager parsing noted in the review of this PR. The regression test uses harmless serialized dictionaries and a controlled unsafe verdict, then asserts that the parser was called exactly once with either complete or truncated trailing streams. Additional tests cover an invalid first stream and the empty-file exception. The existing real-analysis unsafe-second-stream test and benign single/stacked-stream controls are retained.

Validation on Python 3.12.15: the focused file passes all 5 tests, and the full suite passes 157 tests plus 151 subtests with no skips. Formatting and Ruff lint checks pass. The advisory type check reports the same 62 diagnostics before and after the change. The focused tests execute every statement and both reported branches of the updated helper.

Files previously accepted from their first stream alone can now be rejected by a later stream. In particular, legacy PyTorch files contain multiple streams, so later streams are now subject to the same safety ratings. This retains the intended behavior change of the original PR.

AI assistance was used in drafting and validating this change; the patch and tests were reviewed and verified by the author.

[dguido]

[P1] fickling/analysis.py:690 loads every stacked stream before all() can reject an unsafe first stream. In a local check, 100,000 tiny trailing streams (about 200 KB) increased memory use by roughly 47 MB. Please parse and check streams incrementally, returning on the first unsafe result.

[professor-moody]

Thanks @dguido! Updated is_likely_safe() in 31f1f27b10ef to parse and inspect stacked pickle streams incrementally so it returns immediately on the first unsafe stream without preloading trailing streams.