[yashrajp22]

Inspect supported shell programs reconstructed from uploaded sources, generated helpers and archives. Recovered findings retain the consuming command's location and source provenance. Unknown identity, source mutation, unsupported execution and resource limits remain explicitly incomplete.

The change tracks bounded source/artifact state and assignment-time values, audits helper effects before reusing bytes, and applies existing normalized and declared-marker security checks within the parent's deadline and finding budget. Child failures retain the first unresolved cause at the enclosing consumer without losing earlier findings.

Compatible correction: 40e8543260f3f1cb645d365c46267bb0515727f7. This branch incorporates #973, #997 and #999; those prerequisites should merge first. Independent retry and TAR/GZIP work remains in #998 and #1000.

The latest correction carries #973's bounded constant-field proof and import-evidence guards through the typed/generated-source integration. Opaque imports and known omitted inventory paths cannot establish optional caller proof. The original shell-security assertion remains intact.

Current follow-up evidence:

• Standalone fix: preserve language boundaries during static inspection #973 carrier 1aa76e62 passes 597 focused checks from source and 597 from its fresh wheel. Full unchanged CAD static inspection reaches 109/109 in both modes, with security findings and reference caveats retained.
• Combined a8b67911 passes 3,384/3,384 focused checks in each mode, across 13 modules. All 148 fresh-wheel package files match source. Its 50 offline scans have exact source/wheel parity in all 25 pairs and satisfy 46/46 controls. Only CAD changes from 6d90c033: 108/109 becomes 109/109 without finding changes; all other normalized findings and inspection accounting match.
• A separate 12/12 full-skill edge matrix passes with six matching source/wheel pairs. The opaque-import caller correctly remains incomplete without losing prior findings, while the destructive neighbor retains its same-file TM1 finding. The added coverage warning is not a confirmed vulnerability.
• Final combined live validation completed 22 scans across 11 directories, passing 18/18 static expectations with matching static accounting in all 11 pairs. CAD reaches 109/109 in both modes. All six customer runs (CAD, SONiC and ArchCM, source and wheel) completed planned AI work; 16/22 runs have complete AI accounting overall. Six negative-control runs retained 11 final provider refusals. The record preserves 39 response-validation error attempts, two HTTP 503 responses, four connection errors and six missing-usage requests. No completed scan was repeated; this is not an error-free or universal-AI-pass claim.

Prior evidence remains bound to earlier heads: 8ad763cb passed 2,256/2,256 installed-package checks; combined 6d90c033 completed 100 offline and 50 live scans with 44/44 static controls in each, exact offline source/wheel parity and 52/52 live typed-message checks. Its 3,283 focused parser and 97 batch-security checks passed. Seven live runs retained AI-review limitations. Earlier 740d14a8 evidence comprises 180 offline scans with 146 control expectations and 44 live scans with 38 targeted static expectations; its AI degradation remains recorded.

DriveOS remains 9/10 in the new offline comparison. Its first gap is the quoted multiline printf substitution at tests/test_workflow.sh:409–410 (unsupported_syntax/printf_invocation): safe reconstruction is unproved in its sourced/handler-capable shell context. That diagnostic does not establish invalid Bash or an outside-bundle dependency. CAD's earlier 108/109 result is historical; the standalone and combined follow-up both restore 109/109 offline.

This PR remains draft: current-head CI has five successful jobs and its unit job is still running. Greptile reviewed the three-file #973 correction with no security finding and one nonblocking inventory-indexing suggestion, which was evaluated and deferred; that scope does not cover this entire stacked integration. CodeRabbit failed before analysis. Coverage, vulnerabilities and provider failures are separate conclusions; incomplete inspection alone does not prove a vulnerability.