Add cross-platform WGO support and reproducible audit cost controls

[Patrick-Loves-Espresso]

What changed

• Add first-class Claude and OpenCode installation, command, and frontmatter support while retaining one shared WGO skill source.
• Strengthen reviewer selection, prioritization, evidence coverage, and bounded worker contracts.
• Add wgo:cost and automatic post-synthesis cost closeout, with an operationalization refresh.
• Freeze a provenance-derived session manifest and calculate request-level, model-by-model API-equivalent costs from accessible Codex JSONL records.
• Require exact child metadata and bounded task lifecycle intervals so inherited history, unchanged usage echoes, and later session reuse are not double counted.
• Block reviewer-wave and synthesis transitions until every recursively spawned task has exactly one terminal outcome.
• Add a dated rate card, reusable cost report template, realistic nested-session fixtures, and executable contract tests.

Why

WGO needs portable first-class workflows across supported agent platforms and reproducible audit economics. The lifecycle and parsing changes prevent incomplete worker runs from entering synthesis and prevent forked Codex history or repeated token-state events from inflating cost estimates.

Impact

The implementation remains workflow-, skill-, and data-only. It adds no application runtime, persistent helper, third-party dependency, ccusage integration, or OS-specific binary. Cost results are explicitly API-equivalent estimates rather than Codex invoices, and unknown rates remain unpriced rather than silently becoming zero.

Validation

• python3 -m unittest tests.test_skill_contract -v: 48 passed, 0 failed, 0 errors, 0 skipped
• Both audit-only and operationalized fixtures reconcile exactly
• Nested child requests are counted once; inherited history, repeated echoes, post-terminal reuse, and an unrelated same-workspace session are excluded
• git diff --check passed
• Both frozen manifests passed JSON validation

The optional standalone skill validator could not start because PyYAML is unavailable in the local and bundled runtimes; no dependency was installed.