[2G-Bug-03]: Credit habit_completions rows toward evaluate_frequency_step_down "Already done" rate on expired nudges

[WilliM233]

Verification

• ruff check . — ✅ Pass (All checks passed!)
• pytest -v — ✅ Pass (1299 passed, 4 skipped)
• Migration applied locally on brain3-dev: N/A (no schema change)
• Postgres-backed test confirmed: N/A (pure ORM logic; no Postgres-specific behavior)

Ran locally against develop HEAD at 2d0a13a immediately before opening this PR.

Summary

evaluate_frequency_step_down now credits a same-date habit_completions row when an expired nudge in the LIMIT-14 window has no response. This mirrors the D5 fix from #184 on the step-down surface so users whose habits complete via routine cascade (or any non-nudge path) can actually step down from daily nudges — the ADHD quality lens again: multiple low-friction completion paths all feel like "I did it", and scaffolding should fade regardless of which channel recorded the completion.

Explicit responses still win. "Already done" counts on its own (back-compat preserved). An explicit non-"Already done" response (e.g., "Skip today") is never overridden by a completion row.

The completion range is bounded by the scheduled_at span of the 14 most recent notifications, so it adapts naturally to the habit's current frequency — a daily habit looks at ~2 weeks of completions; a weekly habit looks at ~14 weeks. FrequencyStepResult shape unchanged; the denominator stays at nudge count (completions do not inflate it).

_assess_habit_stability's first OR'd criterion calls evaluate_frequency_step_down, so stacking stability transitively inherits this rate improvement for routine-cascade-completed habits. No local change at that call site; an integration test verifies the transitive effect.

Changes

• app/services/graduation.py — evaluate_frequency_step_down now queries habit_completions over the scheduled_at span of the LIMIT-14 window. The already_done_count derivation iterates notifications: "Already done" responses count as before; expired notifications with response IS NULL count if a same-date completion exists; all other explicit responses are untouched. FrequencyStepResult shape unchanged.
• tests/test_graduation.py — new TestEvaluateFrequencyStepDownHybridCredit class with 7 cases covering: routine-cascade credit, no-credit-without-row, mixed explicit+completion 100%, same-day de-dup, explicit-negative override guard, denominator-invariant guard, and the transitive stacking-stability integration test.

How to Verify

1. git checkout feature/2G-Bug-03-step-down-hybrid-credit
2. ruff check . — passes clean
3. pytest -v tests/test_graduation.py — all 167 graduation tests pass, including the 7 new hybrid-credit tests
4. pytest -v — full suite: 1299 passed, 4 skipped

Deviations

None. Implementation matches the GitHub issue acceptance criteria and [2G-03] v2 Amendment 15.

Per Wave 2 brief guidance (D7, Wave 3 #186): the hybrid-credit logic here is deliberately inline and near-identical to the #184 implementation on evaluate_graduation. The duplication is expected and retained — the shared read_completion_history helper refactor lands in Wave 3 as #186.

Test Results

```
$ ruff check .
All checks passed!

$ pytest -v
====================== 1299 passed, 4 skipped in 22.91s =======================
```

New tests in TestEvaluateFrequencyStepDownHybridCredit:

• test_routine_cascade_completion_credited_toward_step_down
• test_no_credit_when_no_completion_row_step_down
• test_mixed_response_and_completion_step_down
• test_same_day_nudge_and_completion_not_double_counted
• test_explicit_negative_response_not_overridden_by_completion_step_down
• test_extra_completions_outside_nudge_range_do_not_inflate_denominator
• test_stacking_stability_inherits_step_down_rate_change

All existing step-down and stacking tests continue to pass (AC #8).

Acceptance Checklist

• evaluate_frequency_step_down reads both notification_queue.response='Already done' and habit_completions rows dated to expired-nudge dates
• Completion-credit applies only to expired notifications within the LIMIT-14 window (AC TICKET-02: FastAPI project scaffolding #2)
• Explicit "Already done" continues to count regardless of completion-row presence (AC TICKET-03: Alembic setup, ORM models, and initial migration #3, back-compat)
• Explicit negative responses (e.g., "Not yet", "Skip today") are NOT overridden by completion rows (AC TICKET-04: CRUD endpoints for Domains and Goals #4)
• FrequencyStepResult response shape unchanged; only current_rate / recommend_step_down differ (AC TICKET-05: CRUD endpoints for Projects and Tasks #5)
• Stacking stability's first OR'd criterion transitively benefits — verified via integration test (AC TICKET-06: CRUD endpoints for Tags and task-tag attachment #6)
• Cooldown, minimum-data (≥5), and "already at weekly" gates unchanged (AC TICKET-07: CRUD endpoints for Routines and Routine Schedules #7)
• All existing step-down tests continue to pass (AC TICKET-08: CRUD endpoints for State Check-ins #8)
• New integration tests added (AC TICKET-09: Activity logging endpoints #9)

Closes #185