fix(usage): count cached tokens in every total — Codebase/Today/Yesterday were ~99% short

[Fahad090NP]

What

The CLI's per-message tokens.input excludes cache reads, and the sums used for the usage rows and charts (input + output + reasoning) dropped cache.read entirely. DeepSeek V4 sessions routinely carry ~700K cached prompt tokens per message, so the displayed totals massively undercounted (verified: codebase showed ~6M tokens vs the real ~954M).

Fix

• HistoryRow now carries a computed tokensTotal = input + output + reasoning + cache.read, verified against the DB's authoritative tokens.total (matches real rows exactly).
• Every counting site uses it: sumDailyUsage (Today/Yesterday), buildUsageSeries (charts + Models tab), codebaseUsage (Codebase row).
• Extension-tracked entries already included cached tokens in promptTokens, so the two sources are now consistent.

Audited and left as-is (already correct)

• Requests — 1 per assistant message + 1 per extension-tracked entry (no overlap).
• Price — CLI rows use the DB's own per-message cost (authoritative); extension-side billing uses billable = prompt − cached.

Verification

New regression test with a cache-heavy DeepSeek row; 277 unit tests, strict ESLint + tsc + prettier + markdownlint green, VSIX packages and installs cleanly.

[Fahad090NP]

Hey @ltmoerdani One thing to flag, I noticed you mention this in your every message that you run tests locally and run linters, etc. Let me explain that the linter that I were implemented never allows developer to commit until the husky runs all linters, tests, and formatters.

[Fahad090NP]

You can simply test the actual changes and do not worry about linters and tests.

[Fahad090NP]

Conflicts resolved — merged the latest main (which now includes #150, #151 and #154) into fix/token-counting. The only conflict was the CHANGELOG (both the token-counting entry and the DeepSeek-thinking entry belong in [Unreleased]); both are kept. Branch is mergeable again — 281 tests, strict lint + tsc green.

[ltmoerdani]

This is a clean fix. I checked the diff, tokensTotal computed once in normalizeHistoryRows, then used across all 3 counting sites. Right approach, not scattered.

One small note: test fixtures now all need tokensTotal manually. Not a blocker, but if there are a lot of new tests down the road, a factory helper might save some boilerplate. Just a heads up, not for this PR.

Conflict resolution is solid both CHANGELOG entries (token-counting + DeepSeek thinking) kept. Checks 2/2 green.