fix: correct wrapper parsing and TP4 batch limits - #917

sudo -n ... 2>&1 can be marked partially inspected because the parser consumes -n as an operand option and splits descriptor redirections incorrectly. Separately, TP4 stops larger skills at 64 batches despite allowing 128 files.

Treat sudo's -n as a flag, parse descriptor redirections atomically, and align TP4's batch allowance with the existing 128-file limit. Dynamic or malformed commands remain incomplete; byte, token, runtime, file and finding limits remain enforced. Added regression coverage includes Python wrapper names, shell options and redirections, 64/65/128/129 batch boundaries, late provider/size/deadline failures, source ranges, and the finding cap.

This PR targets main. Main already contains the wrapper-boundary fix from #779, so that backport is no longer part of this diff. For a 2.13.2 release based on current 2.13.1, cherry-pick #779 (623284fa9e275673ba69d3e82b793bf1aaf15057) first if it is absent, then this PR. Preserve the release branch's existing Python ownership guards when resolving backport conflicts.

Validation for d988f33d13b82fb41d696c2cc5a52b106e75e78b, using fresh base/HEAD wheels and main's locked dependencies in isolated containers:

• 56 offline CLI scans of 14 full skills across base/HEAD and source/wheel; all 28 source/wheel pairs match. UART improves from 33/34 to 34/34 files inspected, retaining all 68 independent findings. The other 13 skill reports are unchanged.
• 1,784 additional edge tests passed (1,675 parser/reporting and 109 provider/lifecycle), extending the earlier 330 focused checks on the identical source tree. Coverage includes malformed and dynamic commands, limits, finding preservation, schema failures, retries, deadlines, concurrency and provider-safe input handling. Repository Ruff lint/format checks also pass.
• Live LLM verification using configured Claude Opus 5 through an OpenAI-compatible gateway: both affected full-skill scans inspect all files (6/6 and 34/34), retaining static findings. Six clean/Python-malicious/Markdown-malicious TP4 control runs across base and HEAD meet their expected outcomes.
• The live large-skill TP4 comparison reproduces the baseline's 64-batch output limit at the reported file. HEAD completes 118/118 batches over 73 paths, with no TP4 partial work, in 463.5 seconds under the unchanged 600-second workflow deadline and concurrency of 2. Both revisions use identical provider-safe input, and completed source ranges match the planned ranges.

Two schema-invalid responses during the live UART scan recovered through existing retries. This PR does not change malformed-response handling; persistent failures remain incomplete. Existing missing-reference warnings still leave the affected full-skill reports partial. Live coverage is limited to this configured model/endpoint, and the CAD acceptance covers TP4. Full-suite CI is separate from these local checks.

The maximum TP4 batch allowance doubles; byte, token, runtime, file and finding limits remain enforced. Workbook scope-exclusion acceptance and evaluator chunk limits require separate changes.