feat(patterns): add temporal_rhythm detector (seventh deterministic detector) (#20) - #39

Closes #10. Re-submits #20 on fresh main base with review feedback addressed.

Summary

Adds temporal_rhythms() as the seventh deterministic detector in src/activity_frames/patterns.py.

What it does

Clusters focused app frames by local 30-minute hour bins across calendar days. Emits WorkPattern(kind="temporal_rhythm") when:

• A bin/span is hit on $\ge 3$ distinct days (MIN_DAYS_FOR_RHYTHM = 3)
• Regularity $\ge 0.60$ (fraction of active days in the window that hit the bin)

Adjacent qualifying bins per app are merged into a single contiguous span before the global [:12] cut, preventing an always-focused app from occupying all output slots.

Example label:
Cursor active 09:00-10:00 on 5/5 days (regularity 1.00)

Review Feedback Addressed (from #20)

• Adjacent Bin Merging: Merges contiguous qualifying 30-minute bins per app into one WorkPattern block before the [:12] cut.
• SQL Query Guard: Added LIMIT 50000 to the SQL query.
• Docstring Formula: Aligned docstring formula to distinct_days_hit / total_active_days_in_window.
• Scope: Hour-of-day clustering only (no inter-occurrence intervals).

Tier Contract (SPEC §7)

The label carries observed time windows and a regularity score only — no interpretation or intent labels (e.g., "morning routine"). The detector remains strictly Tier-1 (measured, deterministic).

Checklist

• Output stays deterministic (no clock, network, randomness, or model calls)
• New behavior has a test; pytest -q passes
• New schema fields, if any, are measured - or live in the tier-2 inferred block with a confidence tag
• No typed text or page content is emitted by default