fix(mcp): 9 quality improvements for search, save, and pipeline

[kckylechen1]

Summary

Reapplies fixes from #42 onto the restructured codebase (db split into modules, hub_ops split into subdirectory, tool_params became a directory).

Of the original 13 fixes, 4 were already applied on main (auto_link scope, enrichment concurrency, search skip response, generate_summary fallback). The remaining 9 are applied here.

Changes

• importance clamp: save_memory now clamps importance to [0.0, 1.0]
• EXTRACTION_PROMPT + fact_to_entry: Prompt requests entities/persons; fact_to_entry parses them from LLM output
• cross-DB score normalization: Global and project results re-normalized by各自 max before merge sort
• hub_feedback false success: hub_record_feedback returns bool; handler returns recorded: false for non-existent IDs; rating clamped [0.0, 5.0]
• strip_code_fence: Uses rfind("```") for last closing fence (handles nested)
• noise filter: Added 6 AI boilerplate patterns (I apologize, As an AI, etc.)
• SKIP_PATTERNS: Long-text check uses named variable references instead of [3..5] slice
• sync_memories: Reads known state from both global and project DBs

Test plan

• cargo check — zero warnings
• cargo test -p memory-core — 39 tests pass

[gemini-code-assist]

Code Review

This pull request introduces several enhancements to the memory system, including updating the feedback mechanism to return a boolean status, clamping rating and importance values, and improving the robustness of code fence stripping in LLM responses. It also extends the memory extraction prompt and logic to support entities and persons, and updates the synchronization process to merge known revisions from project-specific databases. Additionally, a cross-DB score normalization step was added to the search logic to ensure results from global and project scopes are comparable. A review comment suggests optimizing the normalization logic by leveraging the existing sort order to find maximum scores more efficiently.

[gemini-code-assist]

Since combined_results is already sorted by final_score in descending order (lines 241-246), calculating global_max and project_max using fold and f64::max is inefficient as it iterates through the entire list. You can simply find the first occurrence for each scope to get the maximum score.

Suggested change

	let global_max = combined_results
	.iter()
	.filter(|(_, s)| matches!(s, DbScope::Global))
	.map(|(r, _)| r.score.final_score)
	.fold(0.0_f64, f64::max);
	let project_max = combined_results
	.iter()
	.filter(|(_, s)| matches!(s, DbScope::Project))
	.map(|(r, _)| r.score.final_score)
	.fold(0.0_f64, f64::max);
	let global_max = combined_results
	.iter()
	.find(|(_, s)| matches!(s, DbScope::Global))
	.map(|(r, _)| r.score.final_score)
	.unwrap_or(0.0);
	let project_max = combined_results
	.iter()
	.find(|(_, s)| matches!(s, DbScope::Project))
	.map(|(r, _)| r.score.final_score)
	.unwrap_or(0.0);

[kckylechen1]

Removed the per-store re-normalization block instead of micro-optimizing it. That keeps the merged ranking on the raw final scores and avoids flattening relevance gaps between global and project hits.