Model fails to read/reference CLAUDE.md contents despite being loaded in context

Bug Description

When CLAUDE.md is loaded into the context window (as intended), the model fails to actually reference or process its contents when asked direct questions. The information is in context but effectively ignored.

Reproduction

1. Set up a project with a CLAUDE.md containing metadata and instructions
2. The CLAUDE.md is loaded into the system prompt as project instructions
3. Ask the model a direct question whose answer is in the CLAUDE.md (e.g., "What is the tool working folder?")
4. The model will guess, ask the user to clarify, or search externally — instead of reading the CLAUDE.md that is already in its context

Specific Example

My CLAUDE.md has this on line 1:

<!-- crucible-sync: install=/Users/.../Crucible-stable, hash=..., synced=... -->

When asked "What is the tool working folder?", the model:

• Guessed multiple wrong answers
• Asked the user to clarify 5 times
• Searched external files, MCP tools, and ran bash commands
• Never referenced line 1 of its own CLAUDE.md

The user had to explicitly say "So you don't have an md file?" and "You have no system prompt?" before the model finally read its own instructions.

Expected Behavior

The model should treat CLAUDE.md as its primary reference and consult it first when asked questions about its operating context, tool paths, workspace configuration, or role.

Impact

• Wastes significant user time
• Undermines trust in the CLAUDE.md mechanism
• Makes the entire project instructions feature unreliable if the model won't read what's loaded

Environment

• Model: claude-opus-4-6
• Tool: Claude Code CLI
• CLAUDE.md was loaded via standard project instructions mechanism and visible in system prompt