Usage after local midnight vanishes when a turn spans days (turn-level range filter drops post-midnight calls)

[KENSHI601]

Summary: In parseProviderSources (src/parser.ts), turns are range-filtered by turn.calls[0]?.timestamp. For long-running sessions whose turn spans midnight (common with autonomous Codex runs lasting many hours), all calls after midnight are dropped from the new day — today shows nothing until the turn ends, and daily history attributes the whole turn's cost to the turn-start day.

Repro (synthetic rollout):

{"timestamp":"2026-07-28T15:55:00.000Z","type":"session_meta","payload":{"session_id":"x","originator":"codex_cli_rs","cwd":"/tmp/t","model":"gpt-5.5"}}
{"timestamp":"2026-07-28T15:57:00.000Z","type":"response_item","payload":{"type":"message","role":"user","content":[{"type":"input_text","text":"work through midnight"}]}}
{"timestamp":"2026-07-28T15:58:00.000Z","type":"event_msg","payload":{"type":"token_count","info":{"total_token_usage":{"input_tokens":100000,"cached_input_tokens":50000,"cache_write_input_tokens":0,"output_tokens":2000,"reasoning_output_tokens":0,"total_tokens":102000},"last_token_usage":{"input_tokens":100000,"cached_input_tokens":50000,"cache_write_input_tokens":0,"output_tokens":2000,"reasoning_output_tokens":0,"total_tokens":102000}}}}
{"timestamp":"2026-07-28T16:10:00.000Z","type":"event_msg","payload":{"type":"token_count","info":{"total_token_usage":{"input_tokens":200000,"cached_input_tokens":100000,"cache_write_input_tokens":0,"output_tokens":4000,"reasoning_output_tokens":0,"total_tokens":204000},"last_token_usage":{"input_tokens":100000,"cached_input_tokens":50000,"cache_write_input_tokens":0,"output_tokens":2000,"reasoning_output_tokens":0,"total_tokens":102000}}}}

In timezone UTC+8: at local 2026-07-29 00:30, codeburn today reports "No usage data found", although the second token_count (2026-07-29 00:10 local) is squarely inside today. export --from 2026-07-29 likewise returns 0 calls. Same-day control case (user message repeated before the second event) works fine, which isolates the cause to turn-level filtering.

Real-world case: a Codex session running autonomously for 30+ hours (one turn, last user message at 23:57 local) — all post-midnight usage (~150 calls) vanished from today.

Suggested fix: filter calls inside the turn instead of dropping the whole turn:

if (dateRange) {
  const inRangeCalls = turn.calls.filter((c) => {
    const ts = new Date(c.timestamp);
    return !Number.isNaN(ts.getTime()) && ts >= dateRange.start && ts <= dateRange.end;
  });
  if (inRangeCalls.length === 0) continue;
  turn = { ...turn, calls: inRangeCalls };
}

(applies to both the servedSources loop and the durableSources loop in parseProviderSources)

Optionally, in aggregateProjectsIntoDays, attribute per-call cost/token columns to dateKey(call.timestamp) rather than the turn day, so multi-day turns distribute cost across actual days in daily history. Turn-level stats (categories, one-shot) can stay turn-based.

I've verified both changes locally against the repro and a real 51MB rollout: post-midnight usage shows up in today correctly, monthly totals unchanged.

Happy to open a PR with this if you prefer.