feat: add core deck views and native child usage

[Trecek]

The offline observability deck now includes spend, efficiency, skill, and role journeys, with separate skill-run and native child-invocation populations.

• Prepare measures and ratios through the shared core helpers, preserving provider/harness boundaries, sample sizes, unknown states, and exact cohort selections.
• Index verified native Claude and Codex child transcripts with unique spawning-session ownership and child-owned model/provider evidence. Canonicalize Claude inclusive input and fold repeated Codex request snapshots.
• Introduce report-index row schema v2 admission and child-evidence fingerprints so incompatible generations and changed child evidence rebuild correctly.
• Preserve bidirectional skill/role links and literal offline definitions even when a selected cohort has no usage. Add per-signal manual review flags with contributor links, inline disabled reasons, and state that resets on rerender.
• Update CLI consumers, regression coverage, architecture inventories, asset checks, and diagnostics documentation.

Validation:

• All pre-commit hooks passed.
• Configured MCP task test-local-gate passed: 32,327 passed, 289 skipped, 15 xfailed.
• Generated a synthetic deck through the actual CLI and inspected it in Chromium from file://: all four views, facet-preserving links in both directions, empty/missing definitions, review toggle/reset, light/dark themes, and print/PDF. No runtime exceptions or remote requests.
• Added an automated packaging regression test that builds a wheel offline in a per-test output directory and verifies all four new view scripts in its packaged asset tree. The source registry and dependency-order checks remain intact.
• Built a wheel and confirmed the shared deck code and all four view scripts match the final source.
• The separately named task test-check command has not been run independently; MCP testing uses the configured task test-local-gate command.

Closes #4651