red fives (akadora) incorrectly discarded issue

[mzpkaka]

When reviewing the game logs, I noticed that red fives (akadora) are being incorrectly discarded in various situations，not just during riichi
https://riichi.dev/games/a00cb631-360e-42fb-aa00-647836d85f65
E4 Player UnluckyJJ

[smly]

@mzpkaka Thanks for the report. After investigating, this isn't a server/viewer bug. The server logs the mjai JSON as received (with the correct 5mr/5pr/5sr) and the viewer just renders it. The undesired discards are produced by riichienv itself, for two combined reasons:

1. The action space cannot distinguish red fives. As noted in Question: how to distinguish between a tsumogiri action and a non-tsumogiri action #188 (comment), Action.encode() collapses the 136-ID tile space into 34 types via tile / 4, so red and normal 5 share the same discard ID and the model cannot express "drop normal 5, keep red 5".

2. find_action always resolves the ambiguity in favour of the red five. Observation::find_action returns the first legal action whose encoded ID matches, with no tie-break. Red fives have the lowest tile IDs within their rank (16/52/88), so they sit at the front of the candidate list and always win. This affects every inference path.

The mjai-replay path is unaffected, since select_action_from_mjai matches by tid_to_mjai and distinguishes 5m from 5mr.

Plan

Two options on the table:

• A. Extend the action space to add red-five discards (3 new IDs) and update the model side accordingly.
• B. Keep the action space as-is and fix it on the Rust side by preferring non-red candidates in find_action.

I'll evaluate both and decide based on the verification results. (I currently lean toward plan B, since situations where preferring to discard a red five are extremely rare.)

[smly]

Picked plan B. Merged to main in #209.