Strengthen verification tests and add the ASD test-audit workflow

[david-hoffman]

Verification tests previously fixed independent checks to an incidental total order, repeated fixture-authored assertions, and let some negative cases fail before reaching their claimed guard. This change checks operation counts and required dependency edges, exercises lint through the public command, and isolates inventory, dynamic-prefix, and shard-identity rejection paths. It removes redundant assertions and unused support while retaining real installation, receipt, hook, subprocess, and coverage proof.

The supporting test-audit skill adapts OpenClaw's workflow to agentic software delivery (ASD), with evidence-based discovery and separate repair and review. It preserves blind specialist restrictions and requires separate authorization for production repairs. Executable code changes are confined to four verification test files; supporting changes are instructions, documentation, and a lesson. No non-test code, tooling, workflow, or configuration changes are included. Scope and audit evidence are recorded in TEST-AUDIT-002.

Validation: canonical PR verification selected Tier 5/full and passed all 25 steps: 1,940 tests passed (185 warnings; no failures or skips), with 3,421/3,421 statements and 902/902 branches across all 21 owned runtime files. Global and every package coverage are exactly 100%; no statements or branches are excluded or missing. The renewed focused run passed 391 tests in 437.94 s on Python 3.12.14, macOS 27.0 arm64. Targeted Black and git diff --check both returned 0. The clean locked environment passed all 8 preflight steps, a receipt-reuse smoke test, and a compiled SciPy probe. The initial attempt retained 295 passes, 23 failures, and 73 errors; all 96 unsuccessful cases were classified as environment failures. No tests or gates were weakened to resolve them.

A confirmed baseline product issue is deferred outside this test-only repair scope: the public metrics command rejects complete passed pr, library, doctor, and release receipts with Unknown receipt mode. A separate probe of this candidate’s real completed pr receipt reproduced exit 1. Controlled orchestration probes confirmed the other three modes; those fixtures establish compatibility only, not domain coverage. The independent reviewer classified this unchanged issue as outside the candidate and selected gate; it remains explicitly disclosed for a separate runtime repair.

Current state

• Authorization and route: owner-authorized routine audit, repair, and PR. No merge or publication authorization.
• Candidate: 6ca7c16844ef0ebcc53bdaee7eb6db474ab39b75 on codex/test-audit-verification; unchanged base b995f5843edbb9247cdb3d9c0b1b08c97010e02a; target codex-main. Established contract; no specialist checkpoint.
• Required verification: canonical python tools/verification.py pr --base origin/codex-main selected Tier 5/global full scope and returned 0. All 25 steps passed; 1,940 tests passed; exact global/per-package 100% statement and branch coverage. Receipt digest: 5b0c683169b02c88344ea17b4f878c007c2a5839e69c2d78f1f104dcc7172faa. Host: Python 3.12.14, macOS 27.0 arm64; all four domains measured freshly.
• Independent review: fresh reviewer accepted the exact committed candidate after verifying content hashes, base, focused/full evidence, receipt seals/logs, source inputs, and retained artifact hashes. No blocking candidate findings.
• Hosted Linux/macOS/Windows verification: pending. Local evidence establishes this host only. Coverage.py cannot measure shell-hook statements or branches.
• Blockers and next action: no local submission blockers. The candidate is ready for PR review. Keep the metrics repair deferred; hosted verification and verified required-check protection remain merge prerequisites.
• Metrics: 0 new contract scenarios; 6 agent launches (2 discovery, 1 routine author, 1 fresh reviewer, 1 environment setup, 1 tooling preparation), plus coordinator ASD instruction edits; no specialist A/B rounds or C repairs. Focused attempts: 1 environment-failed, 1 renewed passing. No owner spending cap; measured tokens and dollars unavailable.