Skip nil input and memoize plconvert

[tilthouse]

Summary

plconvert (in lib/utils.rb) is called twice per reminder by Reminder#initialize — once for color, once for messaging. Each call forks /usr/bin/plutil. On a typical 4000-reminder store that's ~8000 fork+exec invocations, which dominates cold-start runtime for the tasks command.

Two observations on real data make the work redundant:

• messaging (zContactHandles) is null for the vast majority of reminders. Every plconvert(nil) call still forks plutil, sends an empty stdin, and comes back with nothing.
• color is a list-level attribute that the SQL JOIN duplicates onto every reminder row. ~12 unique color blobs across ~40 lists end up getting re-parsed thousands of times.

Empirical impact (3853-reminder store):

• 100 sequential plutil invocations: ~0.43 s.
• 7700 sequential invocations (the worst case here): ~33 s.
• After this fix, ~12 invocations: ~0.05 s.

This is one of two upstream fixes that take this user's icalpal tasks cold start from ~100 s to ~0.6 s. The other is filed separately as an issue (the ZMembershipsOfRemindersInSectionsAsData blob join in Reminder::QUERY — happy to file that as a PR too once we agree on the right SQL approach).

The fix

Two changes in lib/utils.rb:

def plconvert(obj)
  return nil if obj.nil? || (obj.respond_to?(:empty?) && obj.empty?)

  cached = PLCONVERT_CACHE[obj]
  return cached if cached || PLCONVERT_CACHE.key?(obj)

  # ... existing fork+parse logic ...

  PLCONVERT_CACHE[obj] = result
end

Plus a top-level PLCONVERT_CACHE = {} constant for the cache itself.

The cached || PLCONVERT_CACHE.key?(obj) check makes nil results cacheable too — a blob that fails to parse (Plist::UnimplementedElementError returns nil) doesn't get re-tried per row.

Behavior preservation

• Parsed objects for the same input are byte-identical to the previous behavior (we cache plutil's own output).
• nil inputs previously went through plutil → the rescue-or-empty-result path → effectively nil. The shim returns nil directly, same observable result.
• No public API changes. Return shape (Array of plist objects, or nil) is unchanged.
• Process-scoped cache; not thread-safe (icalPal is single-threaded).

Tests

Adds test/plconvert_test.rb — 5 cases using stdlib minitest:

• nil input returns nil without spawning a subprocess
• empty string returns nil without spawning a subprocess
• repeated calls with the same input only fork plutil once
• distinct inputs each spawn their own subprocess
• a blob that produces nil is still cached (no re-fork on subsequent calls)

The Open3 stub is implemented via singleton-method aliasing rather than Minitest::Mock because the latter was removed in minitest 6.x — the test file should run cleanly on either minitest version.

Run with:

ruby test/plconvert_test.rb

I confirmed the suite catches the regression — reverting the patch causes all 5 tests to fail.

Notes

• No changes to runtime dependencies, gemspec, or executable surface.
• Rubocop clean against the project's existing style.
• Happy to adjust naming (PLCONVERT_CACHE could be private inside a module if you'd prefer), wrap the cache in something LRU-bounded, or split the changes if you want them in smaller pieces.