---
phase: 01-foundation
plan: 04
type: execute
wave: 2
depends_on: [01-02]
files_modified:
  - pkg/types/chunk.go
  - pkg/engine/finding.go
  - pkg/engine/entropy.go
  - pkg/engine/filter.go
  - pkg/engine/detector.go
  - pkg/engine/engine.go
  - pkg/engine/sources/source.go
  - pkg/engine/sources/file.go
  - pkg/engine/scanner_test.go
autonomous: true
requirements: [CORE-01, CORE-04, CORE-05, CORE-06]

must_haves:
  truths:
    - "Shannon entropy function returns expected values for known inputs"
    - "Aho-Corasick pre-filter passes chunks containing provider keywords and drops those without"
    - "Detector correctly identifies OpenAI and Anthropic key patterns in test fixtures via regex"
    - "Full scan pipeline: scan testdata/samples/openai_key.txt → Finding with ProviderName==openai"
    - "Full scan pipeline: scan testdata/samples/no_keys.txt → zero findings"
    - "Worker pool uses ants v2 with configurable worker count"
  artifacts:
    - path: "pkg/types/chunk.go"
      provides: "Chunk struct (Data []byte, Source string, Offset int64) — shared by engine and sources packages"
      exports: ["Chunk"]
    - path: "pkg/engine/finding.go"
      provides: "Finding struct (provider, key value, masked, confidence, source, line)"
      exports: ["Finding", "MaskKey"]
    - path: "pkg/engine/entropy.go"
      provides: "Shannon(s string) float64 — ~10 line stdlib math implementation"
      exports: ["Shannon"]
    - path: "pkg/engine/filter.go"
      provides: "KeywordFilter stage — runs Aho-Corasick and passes/drops chunks"
      exports: ["KeywordFilter"]
    - path: "pkg/engine/detector.go"
      provides: "Detector stage — applies provider regexps and entropy check to chunks"
      exports: ["Detect"]
    - path: "pkg/engine/engine.go"
      provides: "Engine struct with Scan(ctx, src, cfg) <-chan Finding"
      exports: ["Engine", "NewEngine", "ScanConfig"]
    - path: "pkg/engine/sources/source.go"
      provides: "Source interface with Chunks(ctx, chan<- types.Chunk) error"
      exports: ["Source"]
    - path: "pkg/engine/sources/file.go"
      provides: "FileSource implementing Source for single-file scanning"
      exports: ["FileSource", "NewFileSource"]
  key_links:
    - from: "pkg/engine/engine.go"
      to: "pkg/providers/registry.go"
      via: "Engine holds *providers.Registry, uses Registry.AC() for pre-filter"
      pattern: "providers\\.Registry"
    - from: "pkg/engine/filter.go"
      to: "github.com/petar-dambovaliev/aho-corasick"
      via: "AC.FindAll() on each chunk"
      pattern: "FindAll"
    - from: "pkg/engine/detector.go"
      to: "pkg/engine/entropy.go"
      via: "Shannon() called when EntropyMin > 0 in pattern"
      pattern: "Shannon"
    - from: "pkg/engine/engine.go"
      to: "github.com/panjf2000/ants/v2"
      via: "ants.NewPool for detector workers"
      pattern: "ants\\.NewPool"
    - from: "pkg/engine/sources/source.go"
      to: "pkg/types/chunk.go"
      via: "Source interface uses types.Chunk — avoids circular import with pkg/engine"
      pattern: "types\\.Chunk"
---

<objective>
Build the three-stage scanning engine pipeline: Aho-Corasick keyword pre-filter, regex + entropy detector workers using ants goroutine pool, and a FileSource adapter. Wire them together in an Engine that emits Findings on a channel.

Purpose: The scan engine is the core differentiator. Plans 02 and 03 provide its dependencies (Registry for patterns + keywords, storage types for Finding). The CLI (Plan 05) calls Engine.Scan() to implement `keyhunter scan`.
Output: pkg/types/chunk.go, pkg/engine/{finding,entropy,filter,detector,engine}.go and sources/{source,file}.go. scanner_test.go stubs filled.

NOTE on CORE-07 (mmap large file reading): FileSource uses os.ReadFile() in Phase 1, which is sufficient for the test fixtures. mmap-based reading for files > 10MB is deferred to Phase 4 (Input Sources) where it belongs architecturally alongside all other source adapter work.
</objective>

<execution_context>
@$HOME/.claude/get-shit-done/workflows/execute-plan.md
@$HOME/.claude/get-shit-done/templates/summary.md
</execution_context>

<context>
@.planning/phases/01-foundation/01-RESEARCH.md
@.planning/phases/01-foundation/01-02-SUMMARY.md

<interfaces>
<!-- IMPORTANT: Circular import prevention -->
The sources sub-package (pkg/engine/sources) needs the Chunk type.
If Chunk were defined in pkg/engine, then sources would import engine, and engine imports
sources (for the Source interface) — a circular import. Go will refuse to compile.

Resolution: Define Chunk in pkg/types (a shared, import-free package):
  pkg/types/chunk.go  — defines types.Chunk
  pkg/engine/sources  — imports pkg/types (no circular dep)
  pkg/engine          — imports pkg/types and pkg/engine/sources (no circular dep)

<!-- Provider Registry types (from Plan 02) -->
package providers

type Provider struct {
    Name     string
    Keywords []string
    Patterns []Pattern
    Tier     int
}

type Pattern struct {
    Regex      string
    EntropyMin float64
    Confidence string
}

type Registry struct { ... }
func (r *Registry) List() []Provider
func (r *Registry) AC() ahocorasick.AhoCorasick  // pre-built Aho-Corasick

<!-- Three-stage pipeline pattern from RESEARCH.md Pattern 2 -->
chunksChan      chan types.Chunk  (buffer: 1000)
detectableChan  chan types.Chunk  (buffer: 500)
resultsChan     chan Finding      (buffer: 100)

Stage 1: Source.Chunks() → chunksChan (goroutine, closes chan on done)
Stage 2: KeywordFilter(chunksChan) → detectableChan (goroutine, AC.FindAll)
Stage 3: N detector workers (ants pool) → resultsChan

<!-- ScanConfig -->
type ScanConfig struct {
    Workers  int     // default: runtime.NumCPU() * 8
    Verify   bool    // Phase 5 — always false in Phase 1
    Unmask   bool    // for output layer
}

<!-- Source interface -->
type Source interface {
    Chunks(ctx context.Context, out chan<- types.Chunk) error
}

<!-- FileSource -->
type FileSource struct {
    Path      string
    ChunkSize int // bytes per chunk, default 4096
}

Chunking strategy: read file in chunks of ChunkSize bytes with overlap of max(256, maxPatternLen)
to avoid splitting a key across chunk boundaries.

<!-- Aho-Corasick import -->
import ahocorasick "github.com/petar-dambovaliev/aho-corasick"
// ac.FindAll(s string) []ahocorasick.Match — returns match positions

<!-- ants import -->
import "github.com/panjf2000/ants/v2"
// pool, _ := ants.NewPool(workers, ants.WithOptions(...))
// pool.Submit(func() { ... })
// pool.ReleaseWithTimeout(timeout)
</interfaces>
</context>

<tasks>

<task type="auto" tdd="true">
  <name>Task 1: Shared types package, Finding, and Shannon entropy function</name>
  <files>pkg/types/chunk.go, pkg/engine/finding.go, pkg/engine/entropy.go</files>
  <read_first>
    - /home/salva/Documents/apikey/.planning/phases/01-foundation/01-RESEARCH.md (CORE-04 row: Shannon entropy, ~10-line stdlib function, threshold 3.5 bits/char)
    - /home/salva/Documents/apikey/pkg/storage/findings.go (Finding and MaskKey defined there — engine.Finding is a separate type for the pipeline)
  </read_first>
  <behavior>
    - Test 1: Shannon("aaaaaaa") → value near 0.0 (all same characters, no entropy)
    - Test 2: Shannon("abcdefgh") → value near 3.0 (8 distinct chars)
    - Test 3: Shannon("sk-proj-ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqr") → >= 3.5 (real key entropy)
    - Test 4: Shannon("") → 0.0 (empty string)
    - Test 5: MaskKey("sk-proj-abc1234") → "sk-proj-...1234" (first 8 + last 4)
    - Test 6: MaskKey("abc") → "****" (too short to mask)
  </behavior>
  <action>
Create **pkg/types/chunk.go** — the shared type that breaks the circular import:
```go
package types

// Chunk is a segment of file content passed through the scanning pipeline.
// Defined in pkg/types (not pkg/engine) so that pkg/engine/sources can use it
// without creating a circular import with pkg/engine.
type Chunk struct {
    Data   []byte // raw bytes
    Source string // file path, URL, or description
    Offset int64  // byte offset of this chunk within the source
}
```

Create **pkg/engine/finding.go**:
```go
package engine

import "time"

// Finding represents a detected API key from the scanning pipeline.
// KeyValue holds the plaintext key — the storage layer encrypts it before persisting.
type Finding struct {
    ProviderName string
    KeyValue     string // full plaintext key
    KeyMasked    string // first8...last4
    Confidence   string // "high", "medium", "low"
    Source       string // file path or description
    SourceType   string // "file", "dir", "git", "stdin", "url"
    LineNumber   int
    Offset       int64
    DetectedAt   time.Time
}

// MaskKey returns a masked representation: first 8 chars + "..." + last 4 chars.
// Returns "****" if the key is shorter than 12 characters.
func MaskKey(key string) string {
    if len(key) < 12 {
        return "****"
    }
    return key[:8] + "..." + key[len(key)-4:]
}
```

Create **pkg/engine/entropy.go**:
```go
package engine

import "math"

// Shannon computes the Shannon entropy of a string in bits per character.
// Returns 0.0 for empty strings.
// A value >= 3.5 indicates high randomness, consistent with real API keys.
func Shannon(s string) float64 {
    if len(s) == 0 {
        return 0.0
    }
    freq := make(map[rune]float64)
    for _, c := range s {
        freq[c]++
    }
    n := float64(len([]rune(s)))
    var entropy float64
    for _, count := range freq {
        p := count / n
        entropy -= p * math.Log2(p)
    }
    return entropy
}
```
  </action>
  <verify>
    <automated>cd /home/salva/Documents/apikey && go build ./pkg/types/... && go build ./pkg/engine/... && echo "BUILD OK"</automated>
  </verify>
  <acceptance_criteria>
    - `go build ./pkg/types/...` exits 0
    - `go build ./pkg/engine/...` exits 0
    - pkg/types/chunk.go exports Chunk with fields Data, Source, Offset
    - pkg/engine/finding.go exports Finding and MaskKey
    - pkg/engine/entropy.go exports Shannon using math.Log2
    - `grep -q 'math\.Log2' pkg/engine/entropy.go` exits 0
    - MaskKey("sk-proj-abc1234") produces "sk-proj-...1234"
  </acceptance_criteria>
  <done>pkg/types/Chunk exists (no imports, no circular dependency risk), Finding, MaskKey, and Shannon exist and compile.</done>
</task>

<task type="auto" tdd="true">
  <name>Task 2: Pipeline stages, engine orchestration, FileSource, and filled test stubs</name>
  <files>
    pkg/engine/filter.go,
    pkg/engine/detector.go,
    pkg/engine/engine.go,
    pkg/engine/sources/source.go,
    pkg/engine/sources/file.go,
    pkg/engine/scanner_test.go
  </files>
  <read_first>
    - /home/salva/Documents/apikey/.planning/phases/01-foundation/01-RESEARCH.md (Pattern 2: Three-Stage Scanning Pipeline — exact channel-based code example)
    - /home/salva/Documents/apikey/pkg/types/chunk.go
    - /home/salva/Documents/apikey/pkg/engine/chunk.go (if exists — use pkg/types/chunk.go instead)
    - /home/salva/Documents/apikey/pkg/engine/finding.go
    - /home/salva/Documents/apikey/pkg/engine/entropy.go
    - /home/salva/Documents/apikey/pkg/providers/registry.go (Registry.AC() and Registry.List() signatures)
  </read_first>
  <behavior>
    - Test 1: Scan testdata/samples/openai_key.txt → 1 finding, ProviderName=="openai", KeyValue contains "sk-proj-"
    - Test 2: Scan testdata/samples/anthropic_key.txt → 1 finding, ProviderName=="anthropic"
    - Test 3: Scan testdata/samples/no_keys.txt → 0 findings
    - Test 4: Scan testdata/samples/multiple_keys.txt → 2 findings (openai + anthropic)
    - Test 5: Shannon("sk-proj-ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqr") >= 3.5 (entropy check)
    - Test 6: KeywordFilter drops a chunk with text "hello world" (no provider keywords)
  </behavior>
  <action>
Create **pkg/engine/sources/source.go**:
```go
package sources

import (
    "context"

    "github.com/salvacybersec/keyhunter/pkg/types"
)

// Source is the interface all input adapters must implement.
// Chunks writes content segments to the out channel until the source is exhausted or ctx is cancelled.
// NOTE: Source is defined in the sources sub-package (not pkg/engine) and uses pkg/types.Chunk
// to avoid a circular import: engine → sources → engine.
type Source interface {
    Chunks(ctx context.Context, out chan<- types.Chunk) error
}
```

Create **pkg/engine/sources/file.go**:
```go
package sources

import (
    "context"
    "os"

    "github.com/salvacybersec/keyhunter/pkg/types"
)

const defaultChunkSize = 4096
const chunkOverlap = 256 // overlap between chunks to avoid splitting keys at boundaries

// FileSource reads a single file and emits overlapping chunks.
type FileSource struct {
    Path      string
    ChunkSize int
}

// NewFileSource creates a FileSource for the given path with the default chunk size.
func NewFileSource(path string) *FileSource {
    return &FileSource{Path: path, ChunkSize: defaultChunkSize}
}

// Chunks reads the file in overlapping segments and sends each chunk to out.
// Uses os.ReadFile for simplicity in Phase 1. mmap for files > 10MB is implemented
// in Phase 4 (Input Sources) alongside all other source adapter enhancements.
func (f *FileSource) Chunks(ctx context.Context, out chan<- types.Chunk) error {
    data, err := os.ReadFile(f.Path)
    if err != nil {
        return err
    }
    size := f.ChunkSize
    if size <= 0 {
        size = defaultChunkSize
    }
    if len(data) <= size {
        // File fits in one chunk
        select {
        case <-ctx.Done():
            return ctx.Err()
        case out <- types.Chunk{Data: data, Source: f.Path, Offset: 0}:
        }
        return nil
    }
    // Emit overlapping chunks
    var offset int64
    for start := 0; start < len(data); start += size - chunkOverlap {
        end := start + size
        if end > len(data) {
            end = len(data)
        }
        chunk := types.Chunk{
            Data:   data[start:end],
            Source: f.Path,
            Offset: offset,
        }
        select {
        case <-ctx.Done():
            return ctx.Err()
        case out <- chunk:
        }
        offset += int64(end - start)
        if end == len(data) {
            break
        }
    }
    return nil
}
```

Create **pkg/engine/filter.go**:
```go
package engine

import (
    ahocorasick "github.com/petar-dambovaliev/aho-corasick"
    "github.com/salvacybersec/keyhunter/pkg/types"
)

// KeywordFilter filters a stream of chunks using an Aho-Corasick automaton.
// Only chunks that contain at least one provider keyword are sent to out.
// This is Stage 2 of the pipeline (runs after Source, before Detector).
func KeywordFilter(ac ahocorasick.AhoCorasick, in <-chan types.Chunk, out chan<- types.Chunk) {
    for chunk := range in {
        if len(ac.FindAll(string(chunk.Data))) > 0 {
            out <- chunk
        }
    }
}
```

Create **pkg/engine/detector.go**:
```go
package engine

import (
    "regexp"
    "strings"
    "time"

    "github.com/salvacybersec/keyhunter/pkg/providers"
    "github.com/salvacybersec/keyhunter/pkg/types"
)

// Detect applies provider regex patterns and optional entropy checks to a chunk.
// It returns all findings from the chunk.
func Detect(chunk types.Chunk, providerList []providers.Provider) []Finding {
    var findings []Finding
    content := string(chunk.Data)

    for _, p := range providerList {
        for _, pat := range p.Patterns {
            re, err := regexp.Compile(pat.Regex)
            if err != nil {
                continue // invalid regex — skip silently
            }
            matches := re.FindAllString(content, -1)
            for _, match := range matches {
                // Apply entropy check if threshold is set
                if pat.EntropyMin > 0 && Shannon(match) < pat.EntropyMin {
                    continue // too low entropy — likely a placeholder
                }
                line := lineNumber(content, match)
                findings = append(findings, Finding{
                    ProviderName: p.Name,
                    KeyValue:     match,
                    KeyMasked:    MaskKey(match),
                    Confidence:   pat.Confidence,
                    Source:       chunk.Source,
                    SourceType:   "file",
                    LineNumber:   line,
                    Offset:       chunk.Offset,
                    DetectedAt:   time.Now(),
                })
            }
        }
    }
    return findings
}

// lineNumber returns the 1-based line number where match first appears in content.
func lineNumber(content, match string) int {
    idx := strings.Index(content, match)
    if idx < 0 {
        return 0
    }
    return strings.Count(content[:idx], "\n") + 1
}
```

Create **pkg/engine/engine.go**:
```go
package engine

import (
    "context"
    "runtime"
    "sync"
    "time"

    "github.com/panjf2000/ants/v2"
    "github.com/salvacybersec/keyhunter/pkg/engine/sources"
    "github.com/salvacybersec/keyhunter/pkg/providers"
    "github.com/salvacybersec/keyhunter/pkg/types"
)

// ScanConfig controls scan execution parameters.
type ScanConfig struct {
    Workers int  // number of detector goroutines; defaults to runtime.NumCPU() * 8
    Verify  bool // opt-in active verification (Phase 5)
    Unmask  bool // include full key in Finding.KeyValue
}

// Engine orchestrates the three-stage scanning pipeline.
type Engine struct {
    registry *providers.Registry
}

// NewEngine creates an Engine backed by the given provider registry.
func NewEngine(registry *providers.Registry) *Engine {
    return &Engine{registry: registry}
}

// Scan runs the three-stage pipeline against src and returns a channel of Findings.
// The channel is closed when all chunks have been processed.
// The caller must drain the channel fully or cancel ctx to avoid goroutine leaks.
func (e *Engine) Scan(ctx context.Context, src sources.Source, cfg ScanConfig) (<-chan Finding, error) {
    workers := cfg.Workers
    if workers <= 0 {
        workers = runtime.NumCPU() * 8
    }

    // Declare channels on separate lines to ensure correct Go syntax.
    chunksChan := make(chan types.Chunk, 1000)
    detectableChan := make(chan types.Chunk, 500)
    resultsChan := make(chan Finding, 100)

    // Stage 1: source → chunksChan
    go func() {
        defer close(chunksChan)
        _ = src.Chunks(ctx, chunksChan)
    }()

    // Stage 2: keyword pre-filter → detectableChan
    go func() {
        defer close(detectableChan)
        KeywordFilter(e.registry.AC(), chunksChan, detectableChan)
    }()

    // Stage 3: detector workers → resultsChan
    pool, err := ants.NewPool(workers)
    if err != nil {
        close(resultsChan)
        return nil, err
    }
    providerList := e.registry.List()

    var wg sync.WaitGroup
    var mu sync.Mutex

    go func() {
        defer func() {
            wg.Wait()
            close(resultsChan)
            pool.ReleaseWithTimeout(5 * time.Second)
        }()

        for chunk := range detectableChan {
            c := chunk // capture loop variable
            wg.Add(1)
            _ = pool.Submit(func() {
                defer wg.Done()
                found := Detect(c, providerList)
                mu.Lock()
                for _, f := range found {
                    select {
                    case resultsChan <- f:
                    case <-ctx.Done():
                    }
                }
                mu.Unlock()
            })
        }
    }()

    return resultsChan, nil
}
```

Fill **pkg/engine/scanner_test.go** (replacing stubs from Plan 01):
```go
package engine_test

import (
    "context"
    "testing"

    "github.com/salvacybersec/keyhunter/pkg/engine"
    "github.com/salvacybersec/keyhunter/pkg/engine/sources"
    "github.com/salvacybersec/keyhunter/pkg/providers"
    "github.com/stretchr/testify/assert"
    "github.com/stretchr/testify/require"
)

func newTestRegistry(t *testing.T) *providers.Registry {
    t.Helper()
    reg, err := providers.NewRegistry()
    require.NoError(t, err)
    return reg
}

func TestShannonEntropy(t *testing.T) {
    assert.InDelta(t, 0.0, engine.Shannon("aaaaaaa"), 0.01)
    assert.Greater(t, engine.Shannon("sk-proj-ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqr"), 3.5)
    assert.Equal(t, 0.0, engine.Shannon(""))
}

func TestKeywordPreFilter(t *testing.T) {
    reg := newTestRegistry(t)
    ac := reg.AC()

    // Chunk with OpenAI keyword should pass
    matches := ac.FindAll("export OPENAI_API_KEY=sk-proj-test")
    assert.NotEmpty(t, matches)

    // Chunk with no keywords should be dropped
    noMatches := ac.FindAll("hello world no secrets here")
    assert.Empty(t, noMatches)
}

func TestScannerPipelineOpenAI(t *testing.T) {
    reg := newTestRegistry(t)
    eng := engine.NewEngine(reg)
    src := sources.NewFileSource("../../testdata/samples/openai_key.txt")
    cfg := engine.ScanConfig{Workers: 2}

    ch, err := eng.Scan(context.Background(), src, cfg)
    require.NoError(t, err)

    var findings []engine.Finding
    for f := range ch {
        findings = append(findings, f)
    }

    require.Len(t, findings, 1, "expected exactly 1 finding in openai_key.txt")
    assert.Equal(t, "openai", findings[0].ProviderName)
    assert.Contains(t, findings[0].KeyValue, "sk-proj-")
}

func TestScannerPipelineNoKeys(t *testing.T) {
    reg := newTestRegistry(t)
    eng := engine.NewEngine(reg)
    src := sources.NewFileSource("../../testdata/samples/no_keys.txt")
    cfg := engine.ScanConfig{Workers: 2}

    ch, err := eng.Scan(context.Background(), src, cfg)
    require.NoError(t, err)

    var findings []engine.Finding
    for f := range ch {
        findings = append(findings, f)
    }

    assert.Empty(t, findings, "expected zero findings in no_keys.txt")
}

func TestScannerPipelineMultipleKeys(t *testing.T) {
    reg := newTestRegistry(t)
    eng := engine.NewEngine(reg)
    src := sources.NewFileSource("../../testdata/samples/multiple_keys.txt")
    cfg := engine.ScanConfig{Workers: 2}

    ch, err := eng.Scan(context.Background(), src, cfg)
    require.NoError(t, err)

    var findings []engine.Finding
    for f := range ch {
        findings = append(findings, f)
    }

    assert.GreaterOrEqual(t, len(findings), 2, "expected at least 2 findings in multiple_keys.txt")

    var names []string
    for _, f := range findings {
        names = append(names, f.ProviderName)
    }
    assert.Contains(t, names, "openai")
    assert.Contains(t, names, "anthropic")
}
```
  </action>
  <verify>
    <automated>cd /home/salva/Documents/apikey && go test ./pkg/engine/... -v -count=1 2>&1 | tail -30</automated>
  </verify>
  <acceptance_criteria>
    - `go test ./pkg/engine/... -v -count=1` exits 0 with all tests PASS (no SKIP)
    - `go build ./...` exits 0 with no circular import errors
    - TestShannonEntropy passes — 0.0 for "aaaaaaa", >= 3.5 for real key pattern
    - TestKeywordPreFilter passes — AC matches sk-proj-, empty for "hello world"
    - TestScannerPipelineOpenAI passes — 1 finding with ProviderName=="openai"
    - TestScannerPipelineNoKeys passes — 0 findings
    - TestScannerPipelineMultipleKeys passes — >= 2 findings with both provider names
    - `grep -q 'ants\.NewPool' pkg/engine/engine.go` exits 0
    - `grep -q 'KeywordFilter' pkg/engine/engine.go` exits 0
    - pkg/types/chunk.go exists and pkg/engine/sources imports pkg/types (not pkg/engine)
  </acceptance_criteria>
  <done>Three-stage scanning pipeline works end-to-end: FileSource → KeywordFilter (AC) → Detect (regex + entropy) → Finding channel. Circular import resolved via pkg/types. All engine tests pass.</done>
</task>

</tasks>

<verification>
After both tasks:
- `go build ./...` exits 0 with zero circular import errors
- `go test ./pkg/engine/... -v -count=1` exits 0 with all tests PASS
- `grep -q 'ants\.NewPool' pkg/engine/engine.go` exits 0
- `grep -q 'math\.Log2' pkg/engine/entropy.go` exits 0
- `grep -rq 'pkg/types' pkg/engine/sources/source.go` exits 0 (sources imports types, not engine)
- Scanning testdata/samples/openai_key.txt returns 1 finding with provider "openai"
- Scanning testdata/samples/no_keys.txt returns 0 findings
</verification>

<success_criteria>
- Three-stage pipeline: AC pre-filter → regex + entropy detector → results channel (CORE-01, CORE-06)
- Shannon entropy function using stdlib math (CORE-04)
- ants v2 goroutine pool with configurable worker count (CORE-05)
- FileSource adapter reading files in overlapping chunks using os.ReadFile (mmap deferred to Phase 4)
- pkg/types/Chunk breaks the engine↔sources circular import
- All engine tests pass against real testdata fixtures
</success_criteria>

<output>
After completion, create `.planning/phases/01-foundation/01-04-SUMMARY.md` following the summary template.
</output>