golangLAKEHOUSE/internal/catalogd/subject.go

// Subject manifests + per-subject audit-log chain verification.
//
// Specification: /home/profit/lakehouse/docs/specs/SUBJECT_MANIFESTS_ON_CATALOGD.md.
//
// This file is the Go side of Step 8 (cross-runtime parity). The Rust
// side is the writer (crates/catalogd/src/subject_audit.rs); Go is a
// READER + VERIFIER only — both sides serialize / hash with identical
// algorithms so a chain written by Rust verifies under Go.
//
// Algorithm (must match Rust crates/catalogd/src/subject_audit.rs exactly):
//   row_hmac = HMAC-SHA256(key, prev_chain_hash_bytes || canonical_row_bytes)
//
// where:
//   - prev_chain_hash_bytes is the ASCII bytes of the previous row's
//     row_hmac field, OR the literal ASCII string "GENESIS" for the
//     first row (no hex decode!)
//   - canonical_row_bytes is the row JSON with the row_hmac field
//     dropped, with all object keys sorted alphabetically at every
//     nesting depth, with no insignificant whitespace
//   - the final hash is rendered as 64 lowercase hex characters
package catalogd

import (
	"crypto/hmac"
	"crypto/sha256"
	"encoding/hex"
	"encoding/json"
	"fmt"
	"os"
	"sort"
	"strings"
	"time"
)

const GenesisHash = "GENESIS"

// SubjectManifest mirrors crates/shared/src/types.rs::SubjectManifest.
// Keep field tags byte-identical so manifest JSON written by Rust round-trips.
type SubjectManifest struct {
	Schema             string             `json:"schema"`
	CandidateID        string             `json:"candidate_id"`
	CreatedAt          time.Time          `json:"created_at"`
	UpdatedAt          time.Time          `json:"updated_at"`
	Status             string             `json:"status"`
	Vertical           string             `json:"vertical,omitempty"`
	Consent            SubjectConsent     `json:"consent"`
	Retention          SubjectRetention   `json:"retention"`
	Datasets           []SubjectDatasetRef `json:"datasets"`
	SafeViews          []string           `json:"safe_views"`
	AuditLogPath       string             `json:"audit_log_path"`
	AuditLogChainRoot  string             `json:"audit_log_chain_root"`
}

type SubjectConsent struct {
	GeneralPii GeneralPiiConsent `json:"general_pii"`
	Biometric  BiometricConsent  `json:"biometric"`
}

type GeneralPiiConsent struct {
	Status       string     `json:"status"`
	Version      string     `json:"version,omitempty"`
	GivenAt      *time.Time `json:"given_at,omitempty"`
	WithdrawnAt  *time.Time `json:"withdrawn_at,omitempty"`
}

type BiometricConsent struct {
	Status         string     `json:"status"`
	RetentionUntil *time.Time `json:"retention_until,omitempty"`
}

type SubjectRetention struct {
	GeneralPiiUntil time.Time `json:"general_pii_until"`
	Policy          string    `json:"policy,omitempty"`
}

type SubjectDatasetRef struct {
	Name      string `json:"name"`
	KeyColumn string `json:"key_column"`
	KeyValue  string `json:"key_value"`
}

// SubjectAuditRow mirrors crates/shared/src/types.rs::SubjectAuditRow.
// JSON field order does NOT matter (canonicalizer sorts at hash time)
// but field names + types MUST match for round-trip.
type SubjectAuditRow struct {
	Schema         string         `json:"schema"`
	Ts             time.Time      `json:"ts"`
	CandidateID    string         `json:"candidate_id"`
	Accessor       AuditAccessor  `json:"accessor"`
	FieldsAccessed []string       `json:"fields_accessed"`
	Result         string         `json:"result"`
	PrevChainHash  string         `json:"prev_chain_hash"`
	RowHmac        string         `json:"row_hmac"`
}

// AuditAccessor mirrors crates/shared/src/types.rs::AuditAccessor.
//
// IMPORTANT: trace_id has NO omitempty — Rust's serde always writes
// the field (even empty) because `#[serde(default)]` only affects
// READS. Stripping it here produces different canonical bytes than
// the writer used → HMAC mismatch on real production rows. (Caught
// 2026-05-03 by subject_audit_parity.sh on a live WORKER-1 row with
// trace_id="".)
type AuditAccessor struct {
	Kind    string `json:"kind"`
	Daemon  string `json:"daemon"`
	Purpose string `json:"purpose"`
	TraceID string `json:"trace_id"`
}

// LoadSubjectManifest reads + parses a subject manifest JSON file.
func LoadSubjectManifest(path string) (*SubjectManifest, error) {
	bytes, err := os.ReadFile(path)
	if err != nil {
		return nil, fmt.Errorf("read manifest %s: %w", path, err)
	}
	var m SubjectManifest
	if err := json.Unmarshal(bytes, &m); err != nil {
		return nil, fmt.Errorf("parse manifest %s: %w", path, err)
	}
	return &m, nil
}

// AuditLogEntry is one parsed audit row PLUS the raw line bytes from
// the file. The raw bytes are load-bearing: we MUST canonicalize from
// them (not from the re-marshaled struct) to avoid time-precision drift.
//
// Why: Rust's chrono RFC3339-AutoSi serializes nanoseconds with 9 digits
// (e.g. "461439210"). Go's time.RFC3339Nano strips trailing zeros (e.g.
// "46143921"). Round-tripping through time.Time changes the byte sequence
// and breaks the HMAC. Read once, hash from the original bytes.
// (Caught 2026-05-03 by subject_audit_parity.sh on a real WORKER-5 row
// whose nanoseconds happened to end in 0.)
type AuditLogEntry struct {
	Row SubjectAuditRow
	Raw []byte
}

// ReadAuditLog parses an audit JSONL file. Returns one AuditLogEntry
// per non-empty line. Defensive: blank lines are skipped, unparseable
// lines are skipped (matches Rust read_rows_in_range).
func ReadAuditLog(path string) ([]AuditLogEntry, error) {
	bytes, err := os.ReadFile(path)
	if err != nil {
		if os.IsNotExist(err) {
			return nil, nil
		}
		return nil, fmt.Errorf("read audit log %s: %w", path, err)
	}
	var entries []AuditLogEntry
	for _, line := range strings.Split(string(bytes), "\n") {
		trimmed := strings.TrimSpace(line)
		if trimmed == "" {
			continue
		}
		var row SubjectAuditRow
		if err := json.Unmarshal([]byte(trimmed), &row); err != nil {
			continue
		}
		entries = append(entries, AuditLogEntry{Row: row, Raw: []byte(trimmed)})
	}
	return entries, nil
}

// canonicalRowBytesFromRaw is the production canonicalizer used by
// VerifyChain. It accepts the raw line bytes from the audit JSONL,
// drops the row_hmac key, sorts all keys alphabetically at every
// nesting depth, and re-emits compact JSON. Numbers + strings pass
// through verbatim (json.Number preserves number text, strings are
// re-JSON-escaped identically to Rust's serde_json).
//
// CRITICAL: hashing must always go through this function. NEVER hash
// the bytes produced by canonicalRowBytesFromStruct against a real
// production chain — the time-precision drift will give wrong HMACs
// for any row whose nanoseconds end in 0.
func canonicalRowBytesFromRaw(rawLine []byte) ([]byte, error) {
	dec := json.NewDecoder(strings.NewReader(string(rawLine)))
	dec.UseNumber()
	var v any
	if err := dec.Decode(&v); err != nil {
		return nil, fmt.Errorf("decode raw line: %w", err)
	}
	if obj, ok := v.(map[string]any); ok {
		delete(obj, "row_hmac")
	}
	var buf strings.Builder
	if err := writeCanonical(&buf, v); err != nil {
		return nil, err
	}
	return []byte(buf.String()), nil
}

// canonicalRowBytesFromStruct is for the parity probe's known-answer
// vector + tests where we WANT to canonicalize a Go-built row from
// scratch (no original bytes exist yet). Production verification uses
// canonicalRowBytesFromRaw against the file bytes; this function is
// only safe for synthetic inputs whose ts has no trailing-zero nanos.
func canonicalRowBytesFromStruct(row *SubjectAuditRow) ([]byte, error) {
	raw, err := json.Marshal(row)
	if err != nil {
		return nil, fmt.Errorf("marshal row: %w", err)
	}
	return canonicalRowBytesFromRaw(raw)
}

// writeCanonical recursively writes v as canonical JSON: object keys
// sorted alphabetically, no insignificant whitespace. Arrays preserve
// element order (semantically significant per spec §3).
func writeCanonical(buf *strings.Builder, v any) error {
	switch t := v.(type) {
	case map[string]any:
		keys := make([]string, 0, len(t))
		for k := range t {
			keys = append(keys, k)
		}
		sort.Strings(keys)
		buf.WriteByte('{')
		for i, k := range keys {
			if i > 0 {
				buf.WriteByte(',')
			}
			ks, err := json.Marshal(k)
			if err != nil {
				return fmt.Errorf("marshal key: %w", err)
			}
			buf.Write(ks)
			buf.WriteByte(':')
			if err := writeCanonical(buf, t[k]); err != nil {
				return err
			}
		}
		buf.WriteByte('}')
	case []any:
		buf.WriteByte('[')
		for i, elem := range t {
			if i > 0 {
				buf.WriteByte(',')
			}
			if err := writeCanonical(buf, elem); err != nil {
				return err
			}
		}
		buf.WriteByte(']')
	default:
		// json.Number, string, bool, nil — encoding/json renders these
		// the same way Rust's serde_json does (compact, RFC-8259-conformant).
		bs, err := json.Marshal(v)
		if err != nil {
			return fmt.Errorf("marshal scalar: %w", err)
		}
		buf.Write(bs)
	}
	return nil
}

// CanonicalAndHmac is a helper for the parity probe: canonicalizes the
// row from scratch (struct → JSON → canonical), computes the HMAC
// against the given prev_hash, returns both. Used only for the
// known-answer fixture; production verification uses VerifyChain
// (which canonicalizes from the original file bytes to dodge time
// precision drift).
func CanonicalAndHmac(row *SubjectAuditRow, key []byte, prevHash string) (canonical []byte, hmacHex string, err error) {
	canonical, err = canonicalRowBytesFromStruct(row)
	if err != nil {
		return nil, "", err
	}
	hmacHex = computeRowHMAC(key, prevHash, canonical)
	return canonical, hmacHex, nil
}

// computeRowHMAC = HMAC-SHA256(key, prev_hash_ascii_bytes || canonical_row_bytes)
// rendered as 64-char lowercase hex.
func computeRowHMAC(key []byte, prevHash string, canonical []byte) string {
	mac := hmac.New(sha256.New, key)
	mac.Write([]byte(prevHash))
	mac.Write(canonical)
	return hex.EncodeToString(mac.Sum(nil))
}

// VerifyChain replays the chain for a subject's audit log. Returns the
// number of rows verified, the chain tip (last row's row_hmac, or
// GENESIS when log is empty), and an error describing the first chain
// break found.
//
// CRITICAL: canonicalizes from each entry's RAW LINE BYTES, not from
// the parsed struct. Round-tripping through encoding/json + time.Time
// strips trailing zeros from RFC3339 nanos and produces different
// canonical bytes than the Rust writer used (which used chrono's
// AutoSi 9-digit format). Catches: 2026-05-03 WORKER-5 nanos
// 461439210 → Rust emits 9 digits, Go's time.RFC3339Nano emits 8.
//
// Special case: empty/missing log returns (0, GENESIS, nil).
func VerifyChain(entries []AuditLogEntry, key []byte) (count int, chainTip string, err error) {
	prev := GenesisHash
	for i, entry := range entries {
		if entry.Row.PrevChainHash != prev {
			return i, prev, fmt.Errorf("chain break at row %d: prev_chain_hash=%q expected=%q",
				i+1, entry.Row.PrevChainHash, prev)
		}
		claimed := entry.Row.RowHmac
		canonical, cerr := canonicalRowBytesFromRaw(entry.Raw)
		if cerr != nil {
			return i, prev, fmt.Errorf("canonicalize row %d: %w", i+1, cerr)
		}
		recomputed := computeRowHMAC(key, prev, canonical)
		if recomputed != claimed {
			return i, prev, fmt.Errorf("hmac mismatch at row %d: stored=%s recomputed=%s",
				i+1, claimed, recomputed)
		}
		prev = claimed
		count = i + 1
	}
	chainTip = prev
	return count, chainTip, nil
}

// LoadKeyFile reads a signing key file. Refuses keys shorter than 32
// bytes (matches Rust: 2026-05-03 kimi BLOCK fix lifted the minimum
// from 16 to 32 to align with HMAC-SHA256 best practice).
func LoadKeyFile(path string) ([]byte, error) {
	key, err := os.ReadFile(path)
	if err != nil {
		return nil, fmt.Errorf("read key file %s: %w", path, err)
	}
	if len(key) < 32 {
		return nil, fmt.Errorf("signing key is %d bytes; recommend ≥32 bytes for HMAC-SHA256", len(key))
	}
	return key, nil
}