fix(intelligence): _parse_importance_response second-level fallback blindly grabs first number

[knqiufan]

Problem

In src/powermem/intelligence/importance_evaluator.py, the _parse_importance_response() method has a three-level fallback for parsing LLM responses:

1. L1: Extract JSON substring via find("{") / rfind("}") and look for importance_score
2. L2: re.findall(r'\d+\.?\d*', response) — grab the first number from the entire response text
3. L3: Return fixed 0.5

The second-level fallback is semantically unverifiable — it grabs any number without context, which can produce incorrect importance scores that cascade into downstream systems.

Impact

importance_score directly affects:

• _classify_memory_type() — determines working / short_term / long_term classification
• initial_retention = initial_retention * importance_score — initial retention rate in Ebbinghaus algorithm
• adjusted_interval = interval * (1 - importance_score * 0.3) — review schedule intervals

Concrete failure scenarios

Scenario	Number grabbed	Actual meaning	Consequence
JSON parse fails but text contains criteria scores	First dimension score (e.g. 0.9)	A single criterion, not overall	Wrong classification
Reasoning mentions "6 dimensions"	6	Text description	min(1.0, 6) = 1.0 → forced long_term
"Importance 85%"	85	Percentage expression	clamp → 1.0
Text mentions threshold "0.6 short_term"	0.6	Config explanation	Unrelated to content

Inconsistency with other modules

Other parsers in the same codebase (parse_fact_extraction_json, parse_memory_actions_json, skill_manager._parse_skills) return empty structures on parse failure and never guess numbers. The importance evaluator is an outlier with higher downstream cost.

Asymmetry with exception fallback

• LLM call throws exception → falls back to _rule_based_evaluation(content, metadata, context) (content-aware)
• LLM call succeeds but parse is ambiguous → may return a random number via L2, never triggers rule engine

L2 "pretends to succeed" rather than "failing safely."

Suggested fix direction

1. L1: Use parse_json_from_text() (existing utility); support both importance_score and overall_score field names; when only criteria_scores is present, compute weighted sum using criteria_weights
2. L2: Only extract numbers in field-name context (e.g. importance_score"\s*:\s*); reject values outside [0.0, 1.0] instead of clamping
3. L3: Return None → caller falls back to _rule_based_evaluation (consistent with exception path)

Related files

• src/powermem/intelligence/importance_evaluator.py — _parse_importance_response, _llm_based_evaluation
• src/powermem/prompts/importance_evaluation.py — Prompt JSON structure with six criteria
• src/powermem/intelligence/intelligent_memory_manager.py — consumes importance_score
• src/powermem/intelligence/ebbinghaus_algorithm.py — thresholds, retention, review scheduling
• src/powermem/utils/utils.py — parse_json_from_text, extract_json (reusable utilities)