fix(transcript): sum usage.iterations to count fallback attempts

[k-berger]

Problem

Token usage reported by the plugin can undercount what Claude Code itself reports (e.g. via /usage). One confirmed cause: when a request is retried on a fallback model (observed in a real transcript as a Fable attempt followed by an Opus fallback under a single requestId), the transcript entry's top-level message.usage only mirrors the final iteration. Every billed attempt is listed in usage.iterations[].

Real example: top-level showed input=5607, output=698, while the summed iterations are input=11214, output=700 — a 50% input undercount for that call. Cache tokens are similarly undercounted.

Fix

ReadTurnUsage now decodes usage.iterations and sums the token fields across iterations when there is more than one. With zero or one iteration (the normal case — the single iteration duplicates the top-level fields), behavior is byte-identical to before. The streaming keep-last-per-requestId dedup is unchanged; streamed duplicates repeat the same iterations array, so each API call is still counted once.

Testing

• New unit tests: fallback summing (modeled on the real payload), single-iteration parity, streaming dedup of fallback entries, empty-iterations fallback to top-level.
• Spot-checked against the real fallback transcript (now reports 11214/700) and a fresh single-call conversation (unchanged: 12587/979).
• go test ./... and go vet clean.