Function looks_like_source_code 

pub fn looks_like_source_code(content: &str) -> bool

Heuristic fallback: does this text look like source code (vs command output)?

Deliberately conservative — it only returns true when code signals clearly dominate and shell/log signals are essentially absent, so genuine logs and build output are still compressed. Used only when the tool name is unknown.

GH #628: the previous version under-counted real source — a decorative separator comment (// ————, // ====) and the call-shaped scaffolding of a test file (describe(…) {, });) scored as non-code, so a genuine source read routed through an unrecognized tool was lossy-compressed and silently lost those separator lines, breaking the model’s subsequent exact-match edit. The fix: comment lines are neutral (never dilute the ratio) and top-level call/closer shapes count as code even at column 0. The shell-signal veto is untouched, so genuine logs and build output are still compressed.