feat(code): replace vector RAG with Sirchmunk-inspired agentic search

Replace embedding-based vector search with intelligent multi-phase code search
inspired by the Sirchmunk paper (https://github.com/modelscope/sirchmunk).

## Added
- agentic_search tool with 3 modes:
  - FAST: 2-level keyword cascade (2-5s)
  - DEEP: Monte Carlo evidence sampling with Gaussian importance sampling (10-30s)
  - FILENAME_ONLY: Quick file discovery
- IDF-weighted relevance scoring
- File type awareness (code 1.2x, config 1.0x, docs 0.9x)
- Automatic keyword extraction with stop word filtering
- ripgrep_provider: Ripgrep-based context provider
- knowledge_cluster: Sirchmunk-inspired clustering (future feature)
- agentic-search skill definition

## Removed
- embedding.rs: Embedding-based vector search
- vector_provider.rs: Vector RAG context provider
- vector_store.rs: Vector store implementations
- codesearch.rs: Old code search tool

## Changed
- README.md: Add agentic_search to tools table, update test count (1505 → 1471)
- mod.rs: Register agentic_search tool, remove vector RAG imports
- builtin.rs: Add agentic-search skill
- registry.rs: Update skill loading

## Tests
- 12 new tests for agentic_search (all passing)
- 9 new tests for knowledge_cluster (all passing)
- Total: 1471 tests passing

Author: RoyLin

README.md

@@ -5,7 +5,7 @@
 [![Crates.io](https://img.shields.io/crates/v/a3s-code-core.svg)](https://crates.io/crates/a3s-code-core)
 [![Documentation](https://docs.rs/a3s-code-core/badge.svg)](https://docs.rs/a3s-code-core)
 [![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](./LICENSE)
-[![Tests](https://img.shields.io/badge/tests-1505%20passing-brightgreen.svg)](./core/tests)
+[![Tests](https://img.shields.io/badge/tests-1471%20passing-brightgreen.svg)](./core/tests)
 
 ---
 
@@ -85,7 +85,7 @@ print(result.text)
 | Category                     | Tools                                    | Description                                                                   |
 | ---------------------------- | ---------------------------------------- | ----------------------------------------------------------------------------- |
 | **File Operations**    | `read`, `write`, `edit`, `patch` | Read/write files, apply diffs                                                 |
-| **Search**             | `grep`, `glob`, `ls`               | Search content, find files, list directories                                  |
+| **Search**             | `grep`, `glob`, `ls`, `agentic_search` | Search content, find files, list directories, intelligent multi-phase code search |
 | **Execution**          | `bash`                                 | Execute shell commands                                                        |
 | **Web**                | `web_fetch`, `web_search`            | Fetch URLs, search the web                                                    |
 | **Git**                | `git_worktree`                         | Create/list/remove/status git worktrees for parallel work                     |
@@ -666,9 +666,7 @@ All policies are replaceable via traits with working defaults:
 | SecurityProvider     | Input taint, output sanitization               | DefaultSecurityProvider   |
 | PermissionChecker    | Tool access control                            | PermissionPolicy          |
 | ConfirmationProvider | Human confirmation                             | ConfirmationManager       |
-| ContextProvider      | RAG retrieval                                  | FileSystemContextProvider |
-| EmbeddingProvider    | Vector embeddings for semantic search          | OpenAiEmbeddingProvider   |
-| VectorStore          | Embedding storage and similarity search        | InMemoryVectorStore       |
+| ContextProvider      | RAG retrieval                                  | RipgrepContextProvider    |
 | SessionStore         | Session persistence                            | FileSessionStore          |
 | MemoryStore          | Long-term memory backend (from `a3s-memory`) | InMemoryStore             |
 | Tool                 | Custom tools                                   | 14 built-in tools         |
@@ -708,7 +706,7 @@ Agent (config-driven)
         ├── HookEngine (11 lifecycle events)
         ├── Security (PII redaction, injection detection)
         ├── Skills (instruction injection + tool permissions)
-        ├── Context (RAG providers: filesystem, vector)
+        ├── Context (RAG providers: ripgrep, filesystem)
         └── Memory (AgentMemory: working/short-term/long-term via a3s-memory)
 
 AgentTeam (multi-agent coordination)
@@ -1071,7 +1069,7 @@ cargo run --example 02_streaming
 - `test_auto_compact` — Context window auto-compaction
 - `test_security` — Default and custom SecurityProvider
 - `test_batch_tool` — Parallel tool execution via batch
-- `test_vector_rag` — Semantic code search with filesystem context
+- `test_ripgrep_context` — Fast indexless code search with ripgrep provider
 - `test_hooks` — Lifecycle hook handlers (audit, block, transform)
 - `test_parallel_processing` — Concurrent multi-session workloads
 - `test_git_worktree` — Git worktree tool: create, list, remove, status + LLM-driven

core/skills/agentic-search.md

@@ -0,0 +1,140 @@
+---
+name: agentic-search
+description: Intelligent multi-phase code search with keyword extraction and relevance ranking
+allowed-tools: "agentic_search(*), grep(*), read(*)"
+kind: instruction
+tags: ["search", "code", "intelligent", "sirchmunk"]
+version: "1.0.0"
+---
+
+# Agentic Search Skill
+
+You are a specialized code search assistant using intelligent multi-phase retrieval inspired by Sirchmunk.
+
+## When to Use
+
+Use the `agentic_search` tool when the user asks to:
+- Find code related to a concept or feature
+- Locate implementations of specific functionality
+- Search for patterns across the codebase
+- Discover files related to a topic
+
+## Search Strategy
+
+The `agentic_search` tool implements a multi-phase pipeline:
+
+1. **Phase 1 - Keyword Extraction**: Automatically extracts meaningful keywords from natural language queries, removing stop words
+2. **Phase 2 - Parallel Search**: Searches file contents and structure simultaneously
+3. **Phase 3 - Relevance Ranking**: Ranks results using IDF-weighted scoring (files matching more keywords rank higher)
+
+## Mode Selection
+
+### FAST Mode (Default)
+Use for most queries. Returns results in 2-5 seconds.
+
+```typescript
+agentic_search({
+  query: "JWT token validation",
+  max_results: 10
+})
+```
+
+### DEEP Mode (Monte Carlo Sampling)
+Use for comprehensive analysis requiring deep understanding. Returns results in 10-30 seconds.
+
+```typescript
+agentic_search({
+  query: "explain the entire authentication flow from login to session management",
+  mode: "deep",
+  max_results: 3
+})
+```
+
+**When to use DEEP mode:**
+- Complex queries requiring comprehensive understanding
+- When FAST mode results are insufficient
+- Analyzing architectural patterns across multiple files
+- Understanding data flow through the system
+
+**How DEEP mode works:**
+1. **Act 1 - Fuzzy Anchoring**: Identifies high-scoring code regions
+2. **Act 2 - Gaussian Sampling**: Samples lines around matches using Gaussian distribution (sigma adapts 5.0 → 2.0)
+3. **Act 3 - Evidence Synthesis**: Combines sampled regions with evidence scores
+
+### FILENAME_ONLY Mode
+Use for quick file discovery when you only need to find files by name.
+
+```typescript
+agentic_search({
+  query: "auth",
+  mode: "filename_only"
+})
+```
+
+## Usage Examples
+
+### Example 1: Find Authentication Logic
+```typescript
+agentic_search({
+  query: "how does authentication work",
+  max_results: 5
+})
+```
+
+The tool will:
+- Extract keywords: ["authentication", "work"]
+- Search for files containing these terms
+- Rank by relevance (code files with more keyword matches rank higher)
+- Return top 5 results with context
+
+### Example 2: Find Configuration Files
+```typescript
+agentic_search({
+  query: "database configuration",
+  include: "*.{toml,yaml,json}",
+  context_lines: 3
+})
+```
+
+### Example 3: Quick File Discovery
+```typescript
+agentic_search({
+  query: "test",
+  mode: "filename_only",
+  max_results: 20
+})
+```
+
+## Relevance Scoring
+
+Results are ranked by:
+1. **Keyword coverage**: Files matching more unique keywords rank higher
+2. **Match density**: More matches relative to file size = higher score
+3. **File type boost**: Code files (1.2x) > Config files (1.0x) > Docs (0.9x)
+
+## Best Practices
+
+1. **Use natural language**: "how does authentication work" is better than just "auth"
+2. **Be specific**: "JWT token validation" is better than "tokens"
+3. **Combine with read**: After finding relevant files, use `read` to examine them in detail
+4. **Use include patterns**: Narrow search to specific file types when appropriate
+5. **Adjust context_lines**: Use 0-1 for overview, 2-3 for detailed context
+
+## Follow-up Actions
+
+After getting search results:
+1. Use `read` to examine the most relevant files
+2. Use `grep` for more specific pattern matching within discovered files
+3. Use `agentic_search` again with refined keywords if needed
+
+## Comparison to grep
+
+| Feature | agentic_search | grep |
+|---------|----------------|------|
+| Input | Natural language | Regex pattern |
+| Keyword extraction | Automatic | Manual |
+| Relevance ranking | Yes (IDF-weighted) | No (chronological) |
+| File type awareness | Yes | No |
+| Best for | Exploratory search | Precise pattern matching |
+
+Use `agentic_search` for initial discovery, then `grep` for precise follow-up searches.