quantization affects context window degradation threshold - q4 breaks earlier than q5

I tested llama 3.3 70b yesterday with q4_k_m, q5_k_m, and q6_k quantization on same task set - function signature generation with correct default argument values across increasing context windows from 16k to 26k in 1k increments. 200 test cases per context length per quantization level. For q4_k_m: semantic accuracy starts degrading at 17.8k context (91.2% correct), drops to 73.4% at 19k, 58.1% at 21k, 41.7% at 23k. Syntax accuracy remains >94% across all context lengths. For q5_k_m: degradation starts later at 19.4k context (89.7% correct), drops to 76.8% at 21k, 64.2% at 23k, 53.9% at 25k. For q6_k: degradation starts at 20.1k context (88.4% correct), 79.3% at 22k, 71.2% at 24k, 65.7% at 26k. This suggests quantization precision directly affects where rope scaling degradation becomes visible - lower precision moves the threshold earlier in context window. The pattern is consistent: all quants preserve syntax but lose semantics, but q4 loses semantic coherence roughly 2k context earlier than q5, and q5 roughly 1k earlier than q6. For production deployments targeting >20k context, this means q4 is not viable even though vram savings are tempting.