Drop the en-passant TODO and lock in current behavior with a regression test

[BrandonZacharie]

Summary

Smaller than I pitched. When I dug into the # TODO: Append annotation for En passant take. marker at game/game.py:438, the work was already effectively done — I'd misread the code. The engine's regular pawn-capture branch fires for en passant because Move.perform() sets taken_piece from taken_cell (the captured pawn's actual square, not the empty destination), so the elog already records exd6 — the standard PGN notation for an en passant capture, no e.p. suffix needed. JSON and PGN save/load both round-trip the move correctly.

So instead of changing behavior, this PR:

1. Replaces the misleading TODO with a comment explaining what the engine does (and why the regular branch handles it).
2. Adds a regression test that exercises the full en passant path end-to-end: elog token, board state (d5 cleared, d6 holds the white pawn), JSON round-trip, and PGN round-trip via PGNFile. That fence prevents future refactors of the move/capture branches from silently breaking en passant.

Why the previous test sweep didn't catch this implicitly

The Carlsen/Nakamura PGN tests pass moves through PGNFile.game() which uses parse_pgn_move() and then game.move() — the en passant detection in Board.move_piece does the right thing for any pawn diagonal-into-empty move, so the existing tests would have flagged a regression in that path. But the fast-suite tests didn't directly exercise en passant against the engine's own move-side flow, so a refactor that broke the elog token (without breaking parse) could have slipped through. The new test plugs that gap.

What I considered and didn't do

• Adding e.p. suffix to the elog — would diverge from Carlsen.pgn / Nakamura.pgn conventions and break the equality assertion in test_load_pgn_file. Standard PGN doesn't require it.
• Adding an ilog event entry (e.g., (d5, "ep")) for cli display — would break the JSON Game.load path, which expects ilog event strings to be piece-promotion letters in PIECE_NAME_TYPE_MAP. Could be done with a load() change but the value/risk ratio is poor.

Test plan

• make test — fast suite (208 passed, 2 deselected) in ~1.4 s.
• coverage report after make test shows TOTAL 100%.
• Manual: play 1.e4 a6 2.e5 d5, capture en passant with e5 → d6. Save as PGN. Confirm movetext ends ...exd6. Reload — board state matches.

https://claude.ai/code/session_01RnQvxtXfn9JFTN8jpNyriL

---

Generated by Claude Code