[wip]chore: optimize Claude Code config for context efficiency

[RealiCZ]

Summary

Optimize Claude Code configuration for context efficiency, reducing always-loaded token overhead by ~60%.

Layer	Before	After
AGENTS.md	~2900 tokens, no compression guide, constraints buried in Caveats	~1160 tokens, NEVER/ALWAYS/Verification/Compact Instructions
Rules	None	New rust.md (coding conventions) and architecture.md (domain knowledge), loaded on demand
Skills	Verbose description (~50 tokens)	Trigger-focused description (~18 tokens)
Hooks	None	PostToolUse (auto cargo check on .rs edits) + Notification (desktop alert)

AGENTS.md changes

• Extracted domain knowledge (dual gas model, resource limits, gas detention, system contracts, external
  environment) to .claude/rules/architecture.md.
• Extracted Rust coding conventions (no_std, cargo sort, default-features) to .claude/rules/rust.md.
• Replaced free-form Caveats with structured NEVER / ALWAYS / Verification / Compact Instructions sections.

Context overhead comparison

Item	Before	After
AGENTS.md	~2900 tokens	~1160 tokens
Skill descriptors	~95 tokens	~30 tokens
Rules (always loaded)	0	0 (on demand)
Hooks context cost	0	0 (not in context)
Total always-loaded	~2995 tokens	~1190 tokens

[Troublor]

osascript? only applicable on MacOS. I don't think such command should be pushed to git.

[Troublor]

Sometimes the agent is in the middle of a large refactor/implementation and one edit only modifies parts of the code. At this point, the project may be not compilable. I think this hook is unnecessary. We can just ask the agent to always check compilation after edit (the logical multi-file edits, not a single edit) in the AGENTS.md or similar system prompts.