feat(detection): close the Python insecure-deserialisation gaps (#87)

[bkd-dotcom]

Closes #87. The gaps were identified by @AdvaitVarhade in #91 — credited in CONTRIBUTORS.md and the changelog.

Extends the existing block, rather than adding a second one

That distinction is the entire reason #91 couldn't merge as written. Findings dedup on (file, line, category), so a parallel rule using a different category name for the same class defeats the dedup and doubles every finding. One rule_id per class, one category.

Measured on the same 8-line fixture from that review:

	findings
main before	2
#91 as written	4 (every one duplicated)
this PR	2 ✅

Now flagged

case	before	after
marshal.loads/load	miss	caught
shelve.open	miss	caught
yaml.unsafe_load	miss	caught
yaml.load(x, Loader=yaml.Loader)	miss	caught

That last one is the best catch of the set. The old condition was:

if target in ("yaml.load",) and not any(kw.arg == "Loader" for kw in node.keywords):

i.e. any Loader= counted as safe — so an explicitly unsafe loader passed silently. The Loader is now resolved and matched, not merely counted.

Also fixes a pre-existing false positive it exposed

Because the old check only inspected keywords, a positional safe loader was flagged:

yaml.load(x, yaml.SafeLoader)   # was a finding; it is safe

The Loader now comes from the keyword or the second positional argument, matched on its last path segment — so yaml.SafeLoader, a bare imported SafeLoader, and CSafeLoader are all recognised.

All five safe forms verified silent: Loader=yaml.SafeLoader · Loader=SafeLoader · positional yaml.SafeLoader · Loader=yaml.CSafeLoader · yaml.safe_load.

Verification

283 tests pass, ruff clean. Three new tests: the four newly-caught cases, the five safe forms, and a scan_repository test asserting exactly two findings on the duplicate fixture with a single category — which is the regression that would catch a future parallel-rule mistake.

On #91

I'll close it pointing here. @AdvaitVarhade found three real gaps that main genuinely missed — that's the hard part — and the review offered to land the patch with credit if they preferred not to rewrite. Credit is in CONTRIBUTORS.md and the changelog.

[github-actions]

Signetry Reviewer — 🟡 Needs human review

A human should decide — the required check is pending.

Deterministic gates (the authority)

Gate	Status
Required status check	⏳ pending
Secret scan	✅ clean
CI permission / OIDC	✅ no forbidden change
Dependency skew	✅ ok
All green	❌

Findings

No issues found by the deterministic scanners.

Merge

A human should review and merge.

This review is advisory. It never merges on its own judgement — the deterministic gates + a human are the authority. Findings can have false negatives; a green bot verdict is not a guarantee.