Release SkillSpector 2.13.2 - #922

Prepares 2.13.2 from the published v2.13.1 maintenance baseline, preserving all its history and existing fixes, third-party dependency lock, and provider defaults.

Publication approved by the release owner on 28 October 2026 after reviewing the completed validation, including both large-skill CAD timeout failures and the deferred issues below. The tested candidate remains unchanged.

Included backports:

• fix(analyzer): treat a command wrapper with no command as complete #779: stop reporting incomplete analysis for command-wrapper references with no wrapped command at a proven boundary.
• fix: correct wrapper parsing and TP4 batch limits #917: parse sudo -n and descriptor redirections correctly; raise the MCP TP4 batch cap from 64 to the existing 128-file limit while retaining the other resource limits.
• Fix false parser limits in Node assertion templates #921: recognize direct diagnostic templates on verified Node assertion imports, and classify .mjs/.cjs helpers consistently, including archive members. Keep security findings and ambiguous or effectful cases visible.

The three patches apply without implementation changes and match their original merged PRs. The release contains every v2.13.1 commit and deletes no v2.13.1 path. Release-only changes update package metadata and versioned notes. This PR targets the maintenance branch so published fixes remain included; the original fix PRs remain merged into main.

Validation for candidate 03b2a5edb58b844842ee9b3721dfac00858ba01f:

• Current-head GitHub CI: all six jobs passed. The full unit job reports 12,395 passed, 23 skipped, 137 deselected, and four existing expected failures, with 90% coverage. Lint, Docker smoke, TypeScript, change detection, and DCO checks also passed.
• Fresh isolated 2.13.2 verification: 1,912 focused test passes plus nine semantic/runtime/degraded-LLM checks; package builds, wheel/sdist contents, CLI/import versions, lint/format, dependency lock, and release-helper dry run passed. These targeted checks overlap with the full CI suite and are not added to its unique-test total.
• Fresh acceptance: 72 offline CLI scans across 18 complete skill/control directories, with 36 matching source/wheel pairs and no execution or identity failures. Expected parser partials clear while unrelated findings, source locations, missing references, and unresolved-command limits remain intact.
• All 22 additional boundary checks passed in source and installed-wheel modes: JavaScript token, nesting and character limits, and TP4's 63-batch boundary. Existing targeted tests also cover the 64/65 and 128/129 batch boundaries.
• Live clean/positive controls: all 12 TP4 cases passed across v2.13.1 and this candidate in source and installed-wheel modes.
• Live full CLI comparisons: all 12 cases passed for the three reported shell/Python/JavaScript parser reproductions across both versions and both installation modes. The expected parser limitations clear in the candidate; unrelated static findings and missing references remain. Model-generated findings and confidence can vary, including between runs on the same code. One baseline run recovered from a real structured-output schema validation failure through its existing retry path; this is not proof of a new general malformed-response fix.
• Live large-skill TP4 acceptance failed in both candidate modes. Both baseline modes complete the old 64-batch allowance and reproduce its output limit. The candidate plans all 118 batches, but source completes only 65 and the installed wheel only 66 before the unchanged 600-second workflow budget expires. The remaining 53 and 52 ranges respectively are explicitly retained as runtime_limit. This verifies removal of the old batch cap for this input, but does not establish complete live inspection of the large skill. All 28 planned live acceptance cases are finished: 26 pass and these two fail. No longer-budget reruns were used to replace those failures.
• Live coverage uses the configured NVIDIA gateway and one configured model. The large-skill case exercises preparation plus TP4, not the full all-analyzer CLI workflow. Other providers and agent CLI runtimes are not qualified by these runs. The first baseline full CLI trial failed after connection retries; its original results and unknown failed-call costs remain preserved. An isolated diagnostic and the subsequent full CLI run succeeded, but the initial connection failure's cause remains unproven.
• Earlier identical-code verification and failed infrastructure trials are retained as historical evidence; they are not relabeled as fresh 2.13.2 or live-provider passes.

The JavaScript fix is bounded to verified assertion-message ownership. General JavaScript parsing and shell commands reconstructed from piped standard input remain outside these backports. The new TP4 limit is 128 batches: a separately reported 145-batch skill still exceeds it. Persistent obfuscated-instruction limitations, evaluator scope-admission and deduplication policy, and general malformed-provider-response failures are not resolved by these backports. Existing independent inspection limits remain in place. GitHub release v2.13.2 is published with wheel and source-distribution assets. Both downloaded assets are byte-identical to the tested local builds. The tag points to merge commit abd82935b67f765a9541013c0866c6349251dcdf, whose tree exactly matches the tested candidate. This workflow does not publish to PyPI.