pre-index match data to eliminate redundant loops in assembly

[agentksimha]

find_best_assembly_match previously used 4 nested loops, reconstructing
a string pair_key and re-iterating all color buckets on every call
during the Step 4 assembly loop.

Changes:

• Add _build_fragment_match_index() — called once at end of Step 3,
  flattens color buckets and indexes matches by fragment ID in both
  directions for O(1) neighbour lookup
• Refactor find_best_assembly_match() to use the index — replaces the
  two inner loops with a single max() over a pre-flattened list
• Add early exit when similarity >= 1.0
• Add defaultdict/chain imports

Time complexity of the assembly loop drops from O(N² · A · R · C · M)
to O(P·M) for the one-time index build + O(N · K · M) for all assembly
calls combined, where K is actual neighbours (sparse) vs R (full set)