fix: avoid impossible ukeire draws in shanten efficiency

[hong0311]

Summary

• skip impossible draw candidates in calculate_best_ukeire() and calculate_best_ukeire_3p() when the post-discard hand already contains four copies of a tile
• add 4p/3p regression tests for encode_shanten_efficiency() on a quad-containing draw state
• add a standalone repro script for the Houou replay report so the bug can be scanned from a local Tenhou JSON paipu

Problem

encode_shanten_efficiency() could panic on a real Tenhou replay because the ukeire search was evaluating illegal five-of-a-kind hypothetical hands. The panic surfaced from the nyanten lookup in riichienv-core/src/shanten.rs.

Repro

Before this fix, the following replay reproduced the panic locally:

• log id: 2024010100gm-00a9-0000-1d4dbec0
• first failing scan on this branch's repro harness: round_index=0, decision_index=90, seat=1

Verification

• uv run maturin develop
• uv run pytest tests/env/test_apply_event.py -k 'encode_shanten_efficiency_handles_quad_draw'
• uv run pytest tests/test_shanten.py tests/env/test_sanma.py -k 'shanten'
• uv run python scripts/repro_encode_shanten_efficiency_houou.py

[smly]

Thanks for the report! I was able to reproduce the issue you pointed out. I'll review the fix diff now.

FYI. I'm planning to spend this weekend setting up a validation pipeline using Tenhou game logs and adding test cases as needed. I'll create subtasks under #116 and start working on them.

Below is the visualization for debugging:

[smly]

A few separate pre-existing issues also became visible while looking into this, but I think we should go ahead and merge this fix for now. This feature, which is part of feat_v2, had been deferred on our side, and we’ll verify that it works correctly for v0.5.0. Thanks for reporting this bug!

[smly]

(note) The new new_hand_counts[tile_type] >= 4 guard makes sense and should prevent the panic.

That said, ukeire += 4 - visible_counts[tile_type as usize] still seems to over-count remaining tiles, because visible_tiles only includes discards / melds / dora indicators and does not include tiles already present in new_hand.

I checked the callers (encode_shanten_into in both observation/encode.rs and observation_3p/encode.rs) and all_visible is built from discards, melds, and dora indicators, so the player's own hand is not included.

For example, if new_hand already has 3 copies and visible_counts is 0, this counts 4 remaining copies even though only 1 can still exist.

This looks like a pre-existing correctness issue rather than something introduced by this PR, so handling it separately would also be reasonable.

[smly]

(nit) new_hand_counts is rebuilt from scratch inside the outer for (idx, _) in hand_tiles.iter().enumerate() loop, even though new_hand differs from hand_tiles by exactly one tile. Building a base count array once before the loop and adjusting it per iteration would be simpler and a bit more efficient:

let mut base_counts = [0u8; TILE_MAX];
for &tile in hand_tiles {
    let tile_type = (tile / 4) as usize;
    if tile_type < TILE_MAX {
        base_counts[tile_type] += 1;
    }
}

for (idx, _) in hand_tiles.iter().enumerate() {
    let removed_type = (hand_tiles[idx] / 4) as usize;
    let mut new_hand_counts = base_counts;
    new_hand_counts[removed_type] -= 1;
    // ...
}

The runtime difference is negligible for a ~14-tile hand, so this is just a cleanup suggestion.

[Copilot]

Pull request overview

Fixes a shanten-efficiency panic by preventing ukeire search from considering illegal “5th copy” draw candidates, and adds regression coverage plus a local repro harness for the reported Tenhou Houou replay.

Changes:

• Skip draw candidates in calculate_best_ukeire() / _3p() when the post-discard hand already contains 4 copies of that tile type.
• Add 4P/3P regression tests ensuring encode_shanten_efficiency() handles a quad-on-draw state without crashing.
• Add a standalone script to reproduce/scan the panic from a local Tenhou JSON paipu via mjai-reviewer.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 1 comment.

File	Description
riichienv-core/src/shanten.rs	Filters out impossible draw tile types during ukeire evaluation to avoid illegal hypothetical hands.
tests/env/test_apply_event.py	Adds 4P/3P regression tests for quad draw states exercising encode_shanten_efficiency().
scripts/repro_encode_shanten_efficiency_houou.py	Adds a local repro/scanning harness for the Tenhou Houou panic scenario.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

You can also share your feedback on Copilot code review. Take the survey.