[flyinprogrammer]

Closes #305

Summary

Each extra chunk from an already-listed file is scaled by _FILE_SATURATION_DECAY. At 0.5, a strong second chunk of the top file still pushes other files out of the top 10; 0.25 lets another file through:

 # semble/ranking/penalties.py
-_FILE_SATURATION_DECAY = 0.5
+_FILE_SATURATION_DECAY = 0.25

rerank_topk({parser.py#1: 1.0, parser.py#2: 0.9, loader.py: 0.4})
  0.5 : parser.py#1, parser.py#2 (0.45), loader.py (0.40)
  0.25: parser.py#1, loader.py (0.40), parser.py#2 (0.225)

Evidence

The new test_rerank_topk_prefers_another_file_over_a_second_chunk asserts the 0.25 order above.

• Before (main's penalties.py): 1 failed.
• After: 1 passed.
• Full suite: make test 392 passed; make lint and make typecheck clean.

NDCG@10, compared with main (95% CIs resample repos):

Query set	Δ	95% CI	Better / worse queries
Benchmark annotations	+0.09	[+0.04, +0.15]	—
Commit-message queries (#301)	+0.08	[+0.04, +0.13]	41 / 0

run_benchmark: 0.852 → 0.853. Decay 0.1 and 0.0 change nothing further; 0.75 and 1.0 are worse.

Merge Danger

Door: two-way

A one-constant ranking change with no index or cache change; reverting restores the old ranking exactly.

Blast Radius: ranking

Only result order changes, and only when a file has several chunks among the candidates; the result count is unchanged.