[Model behavior] Does not follow the user's written working standards consistently

The user keeps written standards in always-loaded files (CLAUDE.md plus memory): review procedure, evidence order for verifying API behavior, brevity rules, no unrequested actions, no comments beyond one line, one PR per message. The model applied them unevenly: correct in some messages, ignored in the next.

Examples from one session:

• Multiple PR verdicts delivered in one message, against a "one PR per message" rule.
• Comment threads read in truncated form, against a rule to read them whole.
• A verdict reversed on a question alone, against "never flip a verdict without new verified evidence".
• Long answers where the standard says one to three lines for a clean result.

Repro steps

1. Write explicit working standards into CLAUDE.md and memory files that load every session.
2. Work through a long session with reviews, fixes and questions.
3. Compare each response against the written standards; observe violations of rules that are present in context.

Expected

The written standards applied to every message.

Actual

Standards followed intermittently; the user has to point out each violation.

Environment

Claude Code CLI, macOS, Opus model, long sessions with context compaction.

Related

#101574, #101576, #101577