[Bug] Claude Code violates architectural and testing rules under task pressure; inlines SQL, skips HTTP/JWT tests, ignores migration init patterns

Bug Description
The user's CLAUDE.md and persistent memory state several rules as highest-priority. Claude violated
them during a 5-phase implementation task:

1. Response style (flagged in CLAUDE.md as the highest-priority rule, with prior incidents in memory).

Claude wrote walls of text, multi-paragraph end-of-turn summaries, and unnecessary code blocks
throughout. The rule is "1–3 line summary, terse, one question max, details to a doc not pasted
inline."

1. "Tests must verify the actual feature." Backend tests hit the DB layer directly; no test went

through the HTTP/JWT path. No browser walkthrough of Brownfield toggle / dismiss / search mode. Same
shape as the prior "32 passing tests, missed a missing field" incident already in memory.

1. Service → repository layering. Four spots inlined SQL/SQLAlchemy where a repository method should

exist:
inline."

1. "Tests must verify the actual feature." Backend tests hit the DB layer directly; no test went

through the HTTP/JWT path. No browser walkthrough of Brownfield toggle / dismiss / search mode. Same
shape as the prior "32 passing tests, missed a missing field" incident already in memory.

1. Service → repository layering. Four spots inlined SQL/SQLAlchemy where a repository method should

exist:

• product_api.py /products/search endpoint built SELECT/JOIN inline.
• portfolio_section_service.py did select(Product) and update(Portfolio).values(build_type=...)

directly.

• portfolio_cascade_service.py did select(ProductType) directly.

Stated reason on self-audit: "the query felt simple, I was pushing through phases." Rationalization,
not principle.

1. Docker init wiring gap. Added migration 025 and a new master-data loader; did not surface that EI

migrations have no automated init runner until the user asked. Memory rule: "always find and use
proper migration/seed/init scripts."

1. Memory hygiene. Did not write a memory entry for the discovered fact that EI has no automated

migration runner — recurring pitfall.

Pattern across violations: task-size pressure overrode standing rules. Every violation was a "this
one's small enough to skip" rationalization. CLAUDE.md exists precisely to prevent that compounding.

Environment Info

• Platform: darwin
• Terminal: vscode
• Version: 2.1.165
• Feedback ID: 7304799e-6fbf-4439-8b31-9a271e7fa0e7

Errors

[]