[scks1]

Disclosure: this contribution was written by Árbol, an AI assistant working under the direction of the account holder, who set the acceptance criteria and verified the results before submission.

Problem

A comment that appears inside a variation before its first move is
collected by GameBuilder as a pending starting comment. If the
variation contains no move at all — an empty variation such as
({lonely}), or a variation whose moves are all skipped after an
illegal move — the pending comments are never consumed, and they leak
onto the next move parsed afterwards, wherever that is:

import io
import chess.pgn

pgn = "1. e4 e5 ({lonely}) 2. Nf3 Nc6 *"
game = chess.pgn.read_game(io.StringIO(pgn))

nf3 = game.variations[0].variations[0].variations[0]
nf3.starts_variation()    # False
nf3.starting_comments     # ['lonely']  <- leaked

That violates the documented contract of
GameNode.starting_comments ("Only nodes that actually start a
variation can have a starting comment"). It also breaks export round
trips: the exporter writes the starting comment before 2. Nf3, at
which position it re-parses as a post-move comment on e5, so the
comment is silently re-attached to a different node:

parse  -> Nf3.starting_comments == ['lonely'], e5.comments == []
export -> 1. e4 e5 { lonely } 2. Nf3 Nc6 *
parse  -> e5.comments == ['lonely'], Nf3.starting_comments == []

Nested variations are affected the same way
((2. Bc4 ({lonely}) Bc5 ...) lands lonely in
Bc5.starting_comments). Related symptoms of the same leak: a
comment in an empty variation in front of a sibling variation becomes
that sibling's starting comment, and a comment in an empty variation
at the end of a game is dropped entirely.

The defect is also present in the released 1.11.2 GameBuilder
(single-string comment API); this patch is against master.

Fix

Scope the pending starting-comment buffer to the variation it was
collected in:

• begin_variation() saves the enclosing buffer and starts a fresh
  one (so nested variations cannot mix buffers).
• end_variation() attaches any still-pending comments to the
  comments of the node the variation branched from — the position
  the annotation belongs to — and restores the saved buffer.

With this, the example above parses with e5.comments == ['lonely']
and Nf3.starting_comments == [], and export round trips are the
identity. Legitimate starting comments (variations that do contain
moves) are unaffected.

Tests

Added regression tests covering: the mainline leak, round-trip
identity, the nested empty variation, end-of-game preservation,
sibling non-contamination, nested buffer scoping, an invariant sweep
(no node carries starting comments without starting a variation),
and legitimate starting comments. Full test suite with this change: 297 tests, OK
(22 skipped, external engine/tablebase binaries unavailable).
Unpatched master runs 296 tests, OK, and the new regression test
fails there. mypy --strict chess passes.