[MHohlios]

Why

In an area with two look-alike rooms side by side (two "Southern Meadow"s), the map once mistook one for the other and gave the west meadow a south path that doesn't exist. note_exits refreshed a room's exit list on every visit but never removed links, so the bogus path outlived every visit.

What changed

• Layout::drop_unlisted_links (map/layout.rs): on a sure arrival (the tracker is convinced, not adrift, not dark), the current room drops each link in a direction it no longer lists, and the far room drops its return link when that link points back.
• A doorway (an exit once seen with a door) keeps its link on both ends, so shut or hidden doors still work.
• Maze rooms are skipped on both ends: a maze stands for many squares, but its exits are only the last square's.
• Doc comments in map/room.rs and persist/schema.rs updated to the new rule.

Tests

• New a_sure_revisit_drops_a_path_the_room_no_longer_lists: reaches the bad state through a plausible walk (a crooked path fools the map into the wrong meadow), revisits surely, and checks the link is gone after reopening the database. It fails without the fix.
• New a_maze_keeps_its_links_when_its_last_square_lists_fewer_exits.
• links_stay_when_a_room_later_lists_fewer_exits is now a_doorway_keeps_its_link_when_the_room_later_lists_fewer_exits: it sees a closed door first, since only a doorway's link now survives a missing listing.
• cargo fmt --check, cargo clippy --workspace --all-targets --release -- -D warnings and cargo test -p mudling-core pass.

Follow-ups

• A wrong guess the map is convinced of now also drops the real paths of the room it guessed. They're relearned on the next walk through those exits; consider skipping the drop when doubt ended only by the long-doubt fallback or a look-alike link guess (crates/mudling-core/src/map/tracker.rs:382).

Reviewed with Wiffletree: 1 full review and 1 re-check, 2 rounds (Tests, Claude, Style)