llama-github v0.4.5

Adds opt-in per-file and cumulative source limits for memory-bounded PR ingestion while preserving legacy defaults. Oversized or unavailable full source now falls back to the GitHub changed-file patch, with content-free retrieval telemetry and pre-decode/streaming memory guards.