Fix audit follow-ups for usage evidence, teardown preservation, and status checks

[pacphi]

Usage could report incomplete OpenCode costs as complete totals, combine unrelated context measurements into a percentage, and allocate unbounded input buffers. Router teardown could overwrite post-setup user edits. Codex plugin checks also incorrectly rejected bundled capitalized skill names that Codex itself loads successfully.

This PR addresses those agentic-kit behaviors in eight focused commits:

• Preserve observed and estimated cost portions, paired context evidence, and explicit acquisition coverage; bound JSONL and selected SQLite-session reads.
• Preserve edited/unknown/interrupted router and environment projections during teardown, with exact receipts and atomic restoration.
• Correct Codex diagnostic metadata/pricing/privacy and explicitly document its independent-comparison limits.
• Enforce the actual unflagged SQLite Node runtime range and add minimum-runtime CI coverage.
• Accept Codex-supported skill display names; report memory filenames without inferring backend/writer identity from file presence.
• Clarify usage scope, cache/provenance labels, and locale formatting; record reproducible audit evidence and troubleshooting guidance.

Ruflo's CLI/MCP memory path divergence remains an upstream defect. This PR does not merge, delete, migrate, or globally redirect existing memory stores, and does not patch external plugins/packages. Fresh isolated confirmation is posted at ruvnet/ruflo#3196 (comment); follow-up integration is tracked in #213, which must remain open after this PR.

Validation:

• pnpm test: 3,908 passed, six platform skips, zero failures; subsequent legacy suites passed. Coverage: 92.02% lines / 80.89% branches / 91.58% functions.
• Actual Node 22.13.0: 85 integrated tests passed.
• pnpm run test:ui: 491 browser assertions and six suites passed before the final backend status/plugin changes; focused status/plugin regressions pass afterward.
• Typecheck, lint, hard complexity ceiling, Markdown lint, build, and whitespace checks passed. Existing advisory complexity warnings remain.
• Ownership helper: 100% focused line/branch/function coverage.
• All audit synthetic reproductions pass with matching source digests. Separate isolated Ruflo reproduction still demonstrates the upstream split.

Evidence: docs/audits/211-remediation.md, docs/audits/211-remediation-results.json, and docs/audits/plugin-memory-status-followup.md.

Refs #211, #212. Upstream integration follow-up: #213.