Context continuation should re-read CLAUDE.md and memory files

Problem

When a conversation runs out of context and continues in a new session, Claude Code receives a conversation summary but does NOT re-read project instruction files (CLAUDE.md, memory files, README files referenced in instructions). This causes standing rules that were followed correctly in the original session to be violated in the continuation.

Example

My project's CLAUDE.md explicitly says: "NEVER ask users to deploy raw PHP files, FTP/SFTP to servers, or install as mu-plugin." This rule was followed perfectly for months across many sessions. But when a session ran out of context and continued, the summary didn't include this rule, and the AI immediately told me to "upload polymash-bridge.php to your WordPress site's mu-plugins folder" — directly violating the instruction.

Expected behavior

When resuming from a compacted/continued session, Claude Code should re-read all project instruction files (CLAUDE.md, auto-memory, any referenced docs like README files) BEFORE taking any action. The conversation summary captures task state but not standing rules — both are needed.

Current workaround

Adding redundant copies of rules to CLAUDE.md, memory files, and inline README files. But even with all three present, they're not consulted when context is rebuilt from a summary.

Impact

This undermines the entire value proposition of CLAUDE.md and memory files. Users invest time documenting project rules precisely so they're followed consistently. When context continuation silently drops these rules, it erodes trust and creates frustrating failures that the user has to catch and correct.