fix(patterns): one generic-element filter for all three click detectors

[Sarcastic-Soul]

Fixes #52

Important

Stacked on #56. Both PRs touch daily_habits, and building this one on main produced a real merge conflict — so it is branched off #56 instead. Merge #56 first; this PR's own change is the second commit (fix(patterns): one generic-element filter…), which is the one to review here. Once #56 lands, this becomes a single-commit diff automatically. Happy to re-cut it against main in either order if you'd rather.

The bug

patterns.py opens with what looks like the canonical list of elements the recorder couldn't meaningfully name:

_GENERIC_ELEMENTS = {"scroll area", "group", "cell", "text", "text field"}

Only one of the three click detectors actually used it. The other two re-implemented the filter in SQL, and none of the three agreed:

Detector	Filter	Case-sensitive?
action_sequences	the Python set	yes
repeated_clicks	all five re-typed in SQL NOT IN	yes
daily_habits	only ('scroll area','group')	yes

What that produced

One click stream — cell x5 on each of two days, plus Scroll Area x5 (recorder capitalization varies):

Detector	before	after
repeated_clicks	Clicked 'Scroll Area' (AXScrollArea) 5x	(dropped)
daily_habits	'cell' on 2 days, 10x total	(dropped)
action_sequences	Scroll Area -> Scroll Area -> …	[AXScrollArea] -> [AXScrollArea] -> …

'Scroll Area' slipped through because SQL NOT IN is case-sensitive and the list is lowercase. 'cell' was reported as a habit because daily_habits' shorter list omitted it — while repeated_clicks, reading the same rows, filtered that exact name out.

Both results are noise in get_patterns, which is served to agents as automation suggestions. The maintenance trap is the quieter half: the constant sits at the top of the file looking authoritative, so editing it silently fixes one detector out of three.

The change

_GENERIC_ELEMENTS becomes the real single source of truth, with two derived forms:

def _is_generic(name: str | None) -> bool:
    return (name or "").strip().lower() in _GENERIC_ELEMENTS

_GENERIC_SQL = ",".join("?" * len(_GENERIC_ELEMENTS))
_GENERIC_PARAMS = tuple(sorted(_GENERIC_ELEMENTS))

• repeated_clicks keeps filtering in SQL — deliberately. It has ORDER BY cnt DESC LIMIT 20, so generic rows filtered late would consume slots that real elements should hold. The NOT IN list is now generated from the set and matched via lower(trim(element_name)), so it can't drift and it's case-insensitive.
• daily_habits drops its SQL list entirely and calls _is_generic in the loop it already runs. No LIMIT to protect, so the simpler form is fine.
• action_sequences swaps its inline in _GENERIC_ELEMENTS for _is_generic, picking up strip + case-folding. Behaviour is otherwise unchanged: a generic element still keeps its slot in the sequence and stands in as [AXRole], because the click did happen — it just doesn't have a name worth printing.

Tests

Five tests added to tests/test_patterns.py:

• test_is_generic_ignores_case_and_padding — every name in the set, upper/title/padded
• test_repeated_clicks_drops_generic_elements_whatever_the_case — the 'Scroll Area' leak, plus a real element alongside it to confirm nothing over-filters
• test_daily_habits_drops_every_generic_element — the 'cell' habit
• test_action_sequences_stands_in_for_generic_elements — sequence still mined, name replaced by role
• test_all_three_detectors_agree_on_what_is_generic — the invariant itself: one capture containing every generic name, and no detector prints any of them. This is the property the three filters used to disagree on, so it's the one guarding against future drift.

Full suite: 117 passed.

Scope

Filtering only. No threshold, ranking, or label format changes; url_patterns, app_switching, and repeated_text are untouched.