debug

Symptom: $ARGUMENTS ## Reproduce before reasoning - A wrong figure: pin the window (`--since`/`--until` ending before today, never `7d` — the corpus grows while you look), build the binary, run against a copy of the store (`XDG_DATA_HOME=$(mktemp -d)`, then `backfill`) and capture the exact output. Compare `analyze --format json` with the dashboard and the report: the same number rendered differently is a renderer bug, a different number is a derivation bug. - A failing test: run it alone with `-run`, read the golden diff in full, and ask which of the two sides is the truth before touching either. Golden files have hidden a 2× parser error for eleven releases (`docs/corrections.md`). - A parser drift: `doctor --strict` names the canary; find the newest transcript that trips it, cut a minimal redacted sample, and add it to `testdata/` as the reproduction first. ## The usual suspects here - counted once per content block instead of once per response; - a unit chosen before rounding; a share totalled in the wrong dimension; a rate above 100%; - a window one day wider than its name; timezone at the day boundary; - a denominator spanning main-loop, sub-agent and SDK sessions at once (`internal/trace`); - a path that breaks on non-ASCII or a worktree collapsing to `..` (`internal/projectid`); - a shipped migration edited instead of a new one — an upgraded store skips it silently; - a restate path that takes `MAX` and cannot lower a figure (`B116`); - a price missi...