fix: Synth event recorder logs settings changes twice (near-duplicate events)

[thorwhalen]

Closes #5

Root cause

One logical parameter change flowed through two code paths that each log:

1. Synth.update() records the full update dict (the intended chokepoint).
2. When the update touches any settings parameter, update() calls _rebuild_graph(rebuild_updates), which recorded the settings subset again after rebuilding — a few milliseconds later (the rebuild time).

That exactly matches the issue's log: freq-only (dial) changes recorded once, while every change touching waveform/attack (settings) appeared twice, ~0.5–3 ms apart, the second event holding only the settings subset (identical when the whole update was settings).

_rebuild_graph has exactly one caller (update()), so its record call was pure double-logging — not two legitimate writes, hence no dedup/coalesce window.

Fix

Remove the _record_update call from _rebuild_graph; update() is the single logging chokepoint and its event already contains the rebuild updates. Recorded-event schema (relative_time, updates_dict) is unchanged — replay/render consumers are unaffected.

Evidence (issue repro, run headless with audio='dummy')

Before: 9 events, 3 near-duplicate pairs (matching the issue's listing).
After: 6 events, 0 near-duplicate pairs — one event per logical change:

(0, {'freq': {'value': 440, 'time': 0.025, 'mul': 1, 'add': 0}, 'attack': 0.01, 'waveform': 'sine'})
(1.010, {'freq': 660.0})
(2.013, {'freq': 440, 'waveform': 'triangle'})
(2.516, {'waveform': 'square', 'attack': 0.5})
(4.526, {'waveform': 'sine'})
(5.537, {})

Tests

New hum/tests/test_event_recording.py — 4 headless regression tests (fake synth output + fake dials, no pyo server; a minimal fake pyo module is injected when pyo isn't installed, cleaned up so the import-safety tests still see the real environment):

• test_dial_only_change_records_exactly_one_event (guards against overcorrection)
• test_settings_only_change_records_exactly_one_event
• test_dial_plus_setting_change_records_exactly_one_event (the issue's signature case)
• test_sequence_of_changes_records_one_event_each (the issue's full sequence + schema/ordering checks)

Mutation-tested: re-introducing the removed record call turns 3 of the 4 tests red in both a with-pyo and a no-pyo environment (the dial-only test correctly stays green — dial changes were never duplicated); restoring the fix turns them green.

Also added hum/tests to pytest testpaths — CI invokes pytest with no path arguments, so hum/tests/ (including the existing import-safety tests) was not being collected in CI at all. Verified by simulating the exact CI pytest invocation (--doctest-modules + ignore list) in a no-pyo environment: 9 passed, 2 skipped.

Full local suite (pytest tests/ hum/tests/ -v, pyo + portaudio available): 10 passed, including the end-to-end test_synth_frequency_sequence audio test.

https://claude.ai/code/session_01GUjT9e7sAyckG1SR9czApr