fix(api): daily summary uses DueDate-authoritative due-today (no stale pending habits)

[thomasluizon]

The bug

A fresh AI daily summary (fromCache:false) listed habits as still-pending (e.g. "café da manhã", "pesar na balança") even though the user had logged them and the Today screen showed everything done.

Root cause: the summary decided "due today" with a frequency-pattern check (HabitScheduleService.GetScheduledDates(...).Count > 0), while the Today list is DueDate-authoritative. After logging a good recurring habit, LogHabitCommand advances its DueDate to the next occurrence. The Today list hides it, but the summary's pattern still fired — and, worse, children of a due routine were force-included regardless of their own schedule and labeled pending unless each child had its own log in today's range. A sub-task resolved on its real occurrence (e.g. a weekly weigh-in whose DueDate has advanced past today) therefore reappeared as pending under the routine.

The fix

• Thread the authoritative userToday (already computed in GetDailySummaryQueryHandler from the user's timezone) into ISummaryService.GenerateSummaryAsync + AiSummaryService.

• Replace the pattern check with the same gate the Today list uses (GetHabitScheduleQuery.DetermineOverdueStatus / MapChildren):

  IsHabitDueOnDate(h, userToday) || HasMissedPastOccurrence(h, userToday) || IsDoneInRange(h, dateFrom, dateTo)
  

  The IsDoneInRange clause keeps habits completed today in the summary so they're acknowledged as done (never dropped).

• Children are gated by the same predicate on their own merit, instead of being force-included whenever the parent qualified.

• Prompt: added a rule to warmly celebrate a fully-completed day without inventing/suggesting any remaining task, and an instruction to write natural, fluent, grammatically-correct text in the target language.

• The 200-char CapToSentence cap and the model (gpt-4.1-mini) are unchanged.

The habit-selection logic was extracted into a small private SelectScheduledHabits seam so it can be unit-tested directly (mirrors the existing reflection-based test style for BuildHabitSection/BuildSummaryPrompt).

Tests (fail-before / pass-after)

New regression tests in AiSummaryServiceTests. Verified red on the old force-included-children logic, green after:

• SelectScheduledHabits_WeeklyChildResolvedOnItsOccurrence_NotSelected — a weekly child whose DueDate advanced past today is no longer selected.
• SelectScheduledHabits_RoutineWithNonDueChild_DoesNotSurfaceChildAsPending — a monthly child not due today is never listed as a pending sub-task under a due parent.

Additional behavior guards (green both ways): daily/weekly good habit logged today stays and is labeled done, a routine whose sub-tasks were all logged today surfaces no pending child, and an unresolved overdue habit still appears.

GetDailySummaryQueryHandlerTests updated for the new userToday parameter.

Validation

• dotnet build Orbit.slnx — 0 errors.
• dotnet test tests/Orbit.Infrastructure.Tests — 971/971 passed.
• dotnet test tests/Orbit.Application.Tests — 1742/1742 passed.

Follow-up to the #182/#186 humanize-summary work; no tracked issue.

🤖 Generated with Claude Code

[claude]

Clean, focused bug fix. The root cause was well-diagnosed: the summary's frequency-pattern check and the Today list's DueDate-authoritative gate were two different sources of truth, causing logged recurring habits (and their children) to re-appear as pending after LogHabitCommand advanced their DueDate. Threading the already-timezone-correct userToday into SelectScheduledHabits and replacing the pattern check with IsHabitDueOnDate || HasMissedPastOccurrence || IsDoneInRange correctly unifies them. The child-gating fix (each child now qualifies on its own merit rather than inheriting parent eligibility) is the key correctness improvement. Regression tests are precise — they fail on the old force-include logic and pass after — and the interface/signature change is consistent across implementation, handler, and all test call sites.