🐛 fix(user-memory): strengthen anti-bloat rules with HARD RULE blocks and concrete examples

[LoveS0ph1e]

Summary

The previous anti-bloat rule (UPDATE rule 6, added in an earlier server-side patch) was a single prose paragraph — no examples, no decision logic, no self-check. In production with deepseek-v4-flash, 62% of implicit_traits descriptions still contained specific dates after UPDATE cycles, because the model treated the rule as advisory rather than mandatory.

This PR rewrites the anti-bloat rules using techniques already proven effective in the same prompt file: the CRITICAL LANGUAGE RULE achieves near-perfect compliance via a HARD RULE label and bookend repetition (appears at both the top and bottom of the prompt). The anti-bloat rules now use the same pattern.

Changes

Two sections in prompts/en/profile.py:

UPDATE rule 6 (rewritten)

• Upgraded from prose paragraph to **Durable abstraction — HARD RULE (anti-bloat)** block
• IF/THEN logic: the model is told exactly when to rewrite vs append — "IF the new conversation is another instance of a pattern already covered… THEN you MUST re-synthesise"
• Two WRONG/RIGHT example pairs:
  1. Simple appended instance (napping example)
  2. The real-world failure mode: an already-bloated description mixing multiple sub-topics with dates, receiving a third dated instance — the model must rewrite the entire description, not patch the new instance onto the end
• Self-check: "Before finalizing… re-read every description… if any contains a date/weekday/clock-time token, rewrite that description now"

INIT principle 3 (rewritten)

• Same HARD RULE upgrade with IF/THEN structure and WRONG/RIGHT example
• Self-check: "Before finalizing: check every description for date/weekday/clock-time tokens; rewrite any that contain one"

Verification

Tested against a real bloated production profile (26 implicit_traits, 62% dated):

• Baseline (old prose rule): 0/3 clean
• Candidate (this PR): 3/3 clean — zero dates, new information correctly folded into generalizations
• Also verified through the default production path (prompt=None, no test-only override)

Scope

• Only the English prompt file is modified; the Chinese (zh/profile.py) prompts are not imported by the current ProfileExtractor paths
• No schema changes, no code changes — pure prompt text
• Backward compatible: the prompt structure is unchanged; only rule 6 and principle 3 wording differs

[LoveS0ph1e]

Superseded by #4, which consolidates the target-aware and anti-bloat changes into one minimal PR. The anti-bloat HARD RULE from this PR is included verbatim in #4 (now also mirrored in the zh prompt, with a test and CHANGELOG entry). Closing in favor of #4.