Dungeon generation: performance profiling + tuning legibility

[mparrett]

Split out from the dungeon-generation investigation in #87, which has served its purpose: the dev viewer (#88) and the connectivity/floating-door fixes (#89, #91, #92 — both metrics to 0) landed. What is left from that proposal is performance and tuning work, tracked here as its own scope.

Scope

Performance (goal 3 in #87)

• Profile generation cost per level across depth and seed; find the hot paths and edges.
• Evaluate whether the bounding optimization is still carrying its weight now that connectivity is guaranteed by union-find (Guarantee level connectivity with union-find #89), or whether native xxh3 + the cleanup passes let us relax it.
• Offer targeted optimizations where the profile justifies them.

Tuning for game design (goal 4 in #87, feeds #40)

• Make generation parameters legible and adjustable so future game-design choices can tune layout without code surgery.
• The map viewer (Add a standalone dungeon map viewer + audit (tools/mapgen.lg) #88) is the qualitative + statistical harness for this.

Not in scope here

Alternate generation algorithms raised in the #87 thread — BSP, lock-and-key progression, cellular-automata organic levels — are their own design exploration. File separately if we pursue them; this issue is profiling + tuning of the current generator.

Context

• Parent investigation: Proposal: Dungeon generation dev tool + optimizations #87
• Baseline data: Proposal: Dungeon generation dev tool + optimizations #87 comment thread (Marcin-original vs current, floating-door/connectivity sweeps).
• Downstream game-design consumer: Build a Coherent World, Bestiary, and Depth Progression Framework #40.