refactor(core): overhaul file filtering, fix set -e traps, and add extraction summaries

[JacksonFergusonDev]

Description

This PR addresses several edge cases in the context extraction pipeline related to file classification, terminal feedback, and Bash strict mode constraints.

Previously, the file --mime-encoding check was too slow, misclassified empty files (like __init__.py) as binary, and allowed noisy text-based assets (like .svg or minified CSS) to leak into the LLM context. Furthermore, the focal files command lacked visibility into skipped or truncated files and contained a set -e trap during file counting.

Changes

• Global Noise Filtering: Implemented a centralized FOCAL_NOISE_EXTS array in lib/core.sh (mirroring the Python backend). These extensions are compiled into regex for the formatter and injected into fd flags to keep the fzf UI clean of noisy assets.
• Empty File Preservation: Added a [ ! -s "$file" ] intercept in format_file_for_llm to ensure empty structural files (e.g., __init__.py) are captured rather than discarded as binaries.
• Monolithic Text Failsafe: Replaced data-specific truncation with a global 1500-line limit for all readable text files to prevent context window blowouts from massive log files or database dumps.
• Bash Strict Mode Fix: Replaced post-increment operators (((var++))) in libexec/files with safe arithmetic assignments (var=$((var + 1))) to prevent premature script termination when initialized at zero under set -e.
• Extraction Summary: focal files now tallies the status of all processed files and prints a clean, macroscopic summary to stdout upon completion.