fix: add explicit UTF-8 encoding to all text file I/O on Windows

[PLP-N8n]

Summary

On Windows, Python's open() defaults to CP1252 encoding, causing UnicodeDecodeError when reading UTF-8 content (diaries, configs, transcripts, knowledge graph files). Path.read_text() has the same problem.

This patch adds encoding="utf-8" to all text-mode open() calls and read_text() calls across 12 source files. Read operations also get errors="replace" to gracefully handle undecodable bytes rather than crashing.

Changes

File	Pattern	Count
cli.py	read_text() + open("a")	2
config.py	open("r") + open("w")	4
dialect.py	open("r") + open("w")	11
diary_ingest.py	read_text()	1
hooks_cli.py	open("a") + read_text()	6
layers.py	open("r")	1
miner.py	open() + read_text()	2
palace.py	open("w")	1
repair.py	open("w") + open()	2
room_detector_local.py	open("w")	1
spellcheck.py	open()	1
split_mega_files.py	read_text(errors="replace")	2

Total: 34 call sites hardened

Binary-mode open("rb") and os.open() calls are left unchanged — they don't need encoding.

Test plan

• All 2140 tests pass, 46 skipped, 0 failures
• python -c "import mempalace" imports cleanly
• No bare open() or read_text() calls remain (verified via grep)
• Syntax check passes on all 12 modified files

🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request significantly enhances MemPalace by adding a model evaluation harness, support for office document mining, and within-wing entity connectivity. It also hardens the system through UTF-8 I/O enforcement, improved database segment validation, and environment leak protection. Feedback focuses on improving the robustness of venv detection and tool diagnostics, refactoring redundant logic, and ensuring consistent encoding in hook scripts.

[gemini-code-assist]

The write_text call is missing an explicit encoding="utf-8". On Windows, this will default to the system codepage (e.g., CP1252), which contradicts the primary objective of this PR to harden all text I/O with UTF-8 encoding.

Suggested change

	pid_file.write_text(f"{proc.pid} {int(time.time())}")
	pid_file.write_text(f"{proc.pid} {int(time.time())}", encoding="utf-8")

[gemini-code-assist]

The use of a hardcoded index parents[3] to locate the virtual environment root is fragile and platform-dependent. While this may work on Windows, standard POSIX (Linux/macOS) venv layouts include an additional pythonX.Y directory, meaning the venv root would be at parents[4]. Additionally, if the file is at a shallow path, parents[3] might resolve to the filesystem root (e.g., /), leading to incorrect path resolution. Consider searching upwards for a directory containing bin/python or Scripts/python.exe for a more robust solution.

[gemini-code-assist]

The logic for generating a wing slug (.lower().replace(" ", "_").replace("-", "_")) is repeated multiple times within this function (lines 813, 831, and 841). This should be refactored into a local helper function to improve maintainability and ensure consistent slug generation across all branches.

[gemini-code-assist]

Relying on regex matching against TypeError messages is fragile, as the exact format of these messages is an internal implementation detail of the Python interpreter and can change between versions (or even between different implementations like PyPy). While this provides better error reporting for the AI client, it should be treated as a best-effort enhancement with a robust fallback.