llama.cpp 1M-token brute force falters

This is a retrieval problem, not a raw context problem. `--temp` can change how random or conservative the answer feels, but it will not fix the model's inability to reliably surface the right evidence from a huge, noisy prompt.

• –Long-context models still suffer from lost-in-the-middle and positional bias, so facts buried deep in the prompt are easy to miss.
• –Semi-structured org-mode data is a bad fit for brute-force prompting because hierarchy, links, and near-duplicate events need explicit indexing.
• –Maildir is better treated as a searchable corpus with metadata, threading, and citations, not as a text blob for attention to sort out on its own.
• –A hybrid pipeline with heading-based chunking, keyword search, embeddings, reranking, and short-context synthesis will be much more reliable.
• –If the goal is transparency and operability, add search and retrieval tools around the archive instead of trying to make the model remember everything.