Fix session token totals and breakdown

[SmartAI]

Changes

• Include separately recorded reasoning tokens in output and total usage without double-counting Anthropic output.
• Show exact token counts so the breakdown adds up to the total card.
• Distinguish unreported cache writes from explicitly reported zero; show reporting coverage for partial data.
• Rebuild stale daily aggregates without replaying logs and normalize reports from older remote backends.
• Document the accounting rules and add provider-to-log-to-analytics regression coverage.

Verification

• Reproduced the reasoning-token undercount in the desktop flow before the fix.
• 82 tests passed across analytics, provider adapters, Codex, web routes, and desktop integration, including the million-event benchmark.
• Three follow-up desktop checks passed, including the analytics flow.
• Ruff, focused mypy checks, and git diff --check passed. Fixed one existing quick-chat type annotation exposed by mypy.

Provider fixtures cover Anthropic, OpenAI, Codex, DeepSeek, and custom/local protocol variants, including reasoning, caching, missing/partial usage, and repeated attempts. These are controlled fixtures, not live billing reconciliation for every model.