Summary

Fixes #726.

When the Claude OAuth usage response omits the five_hour block (or when
five_hour is present but has no utilization value), mapOAuthUsage
was discarding the entire snapshot with ClaudeUsageError.parseFailed("missing session data"),
even if the response still contained seven_day, seven_day_sonnet,
seven_day_opus, or seven_day_oauth_apps. The user's menu bar therefore
went blank despite the API returning usable data — matching the exact
symptom reported in the issue ("raw response returned, but parser drops
it after a 5h query returns none").

Reproduction

Before this change, on main:

let json = """
{
  "seven_day": { "utilization": 42, "resets_at": "2025-12-29T23:00:00.000Z" },
  "seven_day_sonnet": { "utilization": 17, "resets_at": "2025-12-29T23:00:00.000Z" }
}
"""
try ClaudeUsageFetcher._mapOAuthUsageForTesting(Data(json.utf8))
// throws ClaudeUsageError.parseFailed("missing session data")

After this change the same input returns a ClaudeUsageSnapshot with the
seven-day window promoted to primary (and kept as secondary) and the
Sonnet window as opus.

Approach

mapOAuthUsage now falls back through the available windows in a simple
chain of ?? expressions:

guard let primary = makeWindow(usage.fiveHour, windowMinutes: 5 * 60)
    ?? makeWindow(usage.sevenDay, windowMinutes: 7 * 24 * 60)
    ?? makeWindow(usage.sevenDaySonnet, windowMinutes: 7 * 24 * 60)
    ?? makeWindow(usage.sevenDayOpus, windowMinutes: 7 * 24 * 60)
    ?? makeWindow(usage.sevenDayOAuthApps, windowMinutes: 7 * 24 * 60)
else {
    throw ClaudeUsageError.parseFailed("missing session data")
}

secondary and opus assignments are unchanged, so in the common case
(five-hour present) the snapshot looks identical to today. The error is
only thrown when nothing at all is available.

This mirrors the tolerant decoding pattern added for Codex OAuth in #710.

Tests

Three new regression tests under Tests/CodexBarTests/ClaudeUsageTests.swift:

1. mapOAuthUsage falls back to seven-day window when five_hour is absent
2. mapOAuthUsage falls back when five_hour has no utilization — the exact
   shape from the reporter's debug output
3. mapOAuthUsage throws when no windows are present — makes sure we
   don't synthesise fake data when the response is empty

Build / verification

• swift build → clean on Swift 6.2 (Xcode 16.4, macOS 15.5)
• swiftformat --config .swiftformat <changed files> → clean
• Regression tests are colocated with ClaudeUsageTests and use the
  existing _mapOAuthUsageForTesting debug hook; no new debug surface
  was needed.

Note on swift test: swift test currently fails on this repo from
a clean checkout (both on main and with this change) because the
KeyboardShortcuts dependency's #Preview macros can't resolve
PreviewsMacros under plain SwiftPM. That's pre-existing and unrelated
to this fix — happy to adjust if CI has a different entry point I should
be running.

Scope

Code change is small and localized: 14 lines in ClaudeUsageFetcher.swift,
no public API or behavioural changes when five_hour is present.