perf(aw-transform): cache categorize/tag results per event data to fix month-view timeout

[TimeToBuildBob]

Companion to ActivityWatch/aw-server-rust#657 — same fix requested by @ErikBjare.

Root cause

categorize() and tag() in aw_transform/classify.py were re-evaluating every regex rule against every event individually. For a typical month with 50 000+ events from aw-watcher-window and 20 category rules, this meant ~1 000 000 regex evaluations per query. The vast majority of those events share identical app+title data — heartbeat-based watchers emit the same payload many times per second.

Fix

Added a function-local dict cache inside both categorize() and tag(), keyed on json.dumps(e.data, sort_keys=True). Only the first occurrence of each distinct data fingerprint is matched against the rule set; subsequent events with identical data reuse the cached result in O(1). Each event receives its own list copy to prevent mutation aliasing across events.

json.dumps(..., sort_keys=True) guarantees a consistent key regardless of dict insertion order.

Expected impact

For a month's worth of data with 50 000 events but only ~200 distinct app/title pairs:

• Before: 50 000 × 20 = 1 000 000 regex evaluations
• After: 200 × 20 = 4 000 regex evaluations (~250× less)

Changes

• aw_transform/classify.py: cache added to categorize() and tag(); _categorize_one() and _tag_one() kept as internal helpers (unchanged)
• New test test_categorize_cache_correctness: 101 events (50 Firefox + 1 vim + 50 Firefox), verifies correct category per data shape and that mutating one event's category list does not affect others (list-copy correctness)

[greptile-apps]

Greptile Summary

The PR memoizes category and tag rule results by serialized event data to reduce repeated regex evaluation, while copying cached lists onto each event.

• Adds function-local caches to categorize and tag.
• Adds a categorization test covering repeated data fingerprints and list-copy isolation.
• The new serialization-based key can reject transformed event data that classification previously accepted.

Confidence Score: 4/5

The PR should not merge until classification cache keys can handle transformed event data without aborting composed queries.

The new unconditional JSON serialization fails when a supported upstream transformation stores Event objects in the data payload; the missing profiling evidence is additionally actionable but non-blocking.

Files Needing Attention: aw_transform/classify.py

Important Files Changed

Filename	Overview
aw_transform/classify.py	Adds classification caches, but the JSON cache key can abort composable query pipelines containing non-serializable transformed event data.
tests/test_transforms.py	Covers category-cache reuse and list isolation, but not tagging or classification after transforms that embed Event objects.

_{Reviews (1): Last reviewed commit: "perf(aw-transform): cache categorize/tag..." | Re-trigger Greptile}

[greptile-apps]

Serialization breaks transformed events

When chunk_events_by_key precedes categorize, data["subevents"] contains Event objects, so serializing the complete payload for the cache key raises TypeError and aborts the query; tag has the same issue at line 68.

Knowledge Base Used: Event grouping and normalization

[greptile-apps]

Optimization lacks profiling measurements

These caches implement a performance optimization without profiling or benchmark measurements, so reviewers cannot validate the claimed bottleneck, representative improvement, or the cost of serializing every event payload.

Rule Used: Before implementing performance optimizations, mea... (source)

Learned From
gptme/gptme#707

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

[TimeToBuildBob]

CI fixed (pushed d41a406).

Root cause: json.dumps(e.data, sort_keys=True) raised TypeError when event data contained non-JSON-serializable values — specifically, after chunk_events_by_key embeds nested Event objects in e.data["subevents"]. aw_query/query2.py catches all TypeError exceptions on function calls and re-raises them as "invalid amount of arguments", which made the failure look like a signature mismatch.

Fix: Wrap the json.dumps call in a try/except in both categorize() and tag(), falling back to str(id(e.data)) for non-serializable data. That path gets no cache benefit but is correct — no false cache hits.

All 31 tests in test_transforms.py and test_query2.py::test_query2_query_functions pass locally.