fix(injector): read from all devnodes sharing an origin_hash

[nickbock]

Fixes #1309

Problem

On a device that exposes multiple /dev/input/event* nodes, the injector could grab and start its read loop on the wrong node, so the mapped button never produced any output. Observed with a SteelSeries Aerox 9 Wireless: the mapped button emits on one event node, but the injector read from a different one and nothing was injected. A single-interface device (e.g. a plain keyboard) was unaffected.

Root cause

Device selection in Injector keyed everything on get_device_hash(), which is md5(str(capabilities) + name). Two issues:

1. Hash collisions. That hash is not unique per devnode. Some hardware (the Aerox 9 among them) exposes several event nodes with byte-identical capabilities and name, so they produce the same origin_hash. _find_input_device built devices_by_hash = {hash: device} and _grab_devices returned {hash: device} — both collapse colliding nodes to a single entry, so only one twin was grabbed/read. If the event is emitted on the other twin, it is silently missed.

2. Over-strict capability gate. _find_input_device additionally required the mapped code to be present in the node's advertised capabilities. Nodes that emit codes outside their static capabilities had their valid origin_hash discarded, and _update_preset then overwrote it via the device-type fallback — pointing the injector at a different node.

Fix

• Replace _find_input_device with _find_input_devices, returning every devnode whose hash matches the origin_hash (and dropping the capability gate — a matching hash already guarantees identical capabilities to recording time, so the recorded node is authoritative).
• Key _grab_devices by path instead of hash, so colliding nodes are all grabbed.
• Start one read loop per devnode in run(). Forwarding devices stay keyed by hash (one per unique hash; colliding nodes are interchangeable for forwarding).

When origin_hash is absent or matches nothing (old presets / configs moved between machines), behavior is unchanged: the existing _find_input_device_fallback still runs.

Tests

Adds two regression tests in tests/unit/test_injector.py:

• test_grabs_all_nodes_sharing_an_origin_hash — two hash-identical nodes are both grabbed (the Aerox 9 case).
• test_reads_from_origin_hash_node_on_multi_interface_device — a node that under-reports the mapped code in its capabilities is still honored, not overwritten by the fallback.

Both fail on main and pass with the fix. test_injector.py, test_event_pipeline, test_groups, test_event_reader and test_context pass (the two pre-existing subprocess-spawning failures in test_injector fail identically on main and are unrelated to this change).

Verified live on the affected SteelSeries Aerox 9 Wireless: the mapped button now works.

🤖 Generated with Claude Code