petal/internal/docs/passport.go

package docs

// Writing passport: a standalone, printable report showing *how* a document was
// written — when each snapshot landed, how the word count grew, how the work
// broke into sessions. It exists because automated "AI detector" verdicts are
// unreliable and skew against non-native English writers, so the useful thing to
// hand someone who doubts your authorship is not a score but a record.
//
// The report is deliberately modest about what it proves (see passportLimits):
// it evidences a plausible writing process, it does not certify one.

import (
	"crypto/sha256"
	"database/sql"
	"encoding/hex"
	"errors"
	"fmt"
	"net/http"
	"time"

	"github.com/go-chi/chi/v5"

	"gitea.parodia.dev/drwily/petal/internal/db"
	"gitea.parodia.dev/drwily/petal/internal/httputil"
)

// Passport tuning.
const (
	// sessionGap is the idle time that separates one writing session from the
	// next. Auto-snapshots fire at most every 3 minutes while typing, so any
	// gap far above that means the writer stepped away. 45 minutes keeps a
	// coffee break inside one session but splits morning from evening work.
	sessionGap = 45 * time.Minute

	// jumpNoteThreshold is the share of the final word count a single
	// snapshot-to-snapshot increase must exceed before the report calls it out.
	// A large jump is the first thing a skeptical reader will ask about, so the
	// report raises it rather than leaving it to be discovered.
	jumpNoteThreshold = 0.25
)

// chainHash links a snapshot to its predecessor. Covering prev_hash makes each
// hash depend on the entire history before it, so altering any earlier snapshot
// invalidates every later one; covering created_at means a row cannot be
// silently backdated.
//
// This detects tampering with the local database. It is not third-party
// attestation — someone with the database and this function could regenerate a
// consistent chain from scratch.
func chainHash(prevHash, docID string, createdAt time.Time, wordCount int, text string) string {
	h := sha256.New()
	fmt.Fprintf(h, "%s\x00%s\x00%d\x00%d\x00%s",
		prevHash, docID, createdAt.UTC().UnixNano(), wordCount, text)
	return hex.EncodeToString(h.Sum(nil))
}

// --- report model -----------------------------------------------------------

// passportSession is one continuous stretch of work — snapshots with no
// sessionGap-sized pause between them.
type passportSession struct {
	Start, End time.Time
	Snapshots  int
	WordsAdded int // net change across the session; negative when trimming
}

// Duration is the observed length of the session: first snapshot to last. A
// single-snapshot session reports zero, which is why total active time is
// described as a lower bound.
func (s passportSession) Duration() time.Duration { return s.End.Sub(s.Start) }

// chain verification outcomes, in the order the report prefers to report them.
const (
	chainVerified     = "verified"     // every hash recomputes and every link holds
	chainGaps         = "gaps"         // hashes valid, links broken — consistent with pruning
	chainPartial      = "partial"      // some snapshots predate the hash chain
	chainUnverifiable = "unverifiable" // no snapshot carries a hash
	chainBroken       = "broken"       // a hash does not match its own contents
)

// passportData is everything the template renders.
type passportData struct {
	Doc      db.Document
	Versions []db.DocumentVersion // ascending by time

	Sessions   []passportSession
	FirstAt    time.Time
	LastAt     time.Time
	Span       time.Duration // wall-clock first snapshot → last
	ActiveTime time.Duration // summed session durations; a lower bound

	LargestJump    int // biggest single snapshot-to-snapshot word increase
	LargestJumpAt  time.Time
	LargestJumpIdx int  // index into Versions, so the chart can mark it
	NoteJump       bool // jump is large enough to be worth pre-empting

	ChainStatus   string
	UnhashedCount int
	GeneratedAt   time.Time
}

// buildPassport derives the report from a document and its snapshots, which must
// be ordered oldest-first. It assumes nothing about snapshot spacing.
func buildPassport(doc db.Document, versions []db.DocumentVersion) passportData {
	d := passportData{
		Doc:         doc,
		Versions:    versions,
		GeneratedAt: time.Now(),
	}
	if len(versions) == 0 {
		d.ChainStatus = chainUnverifiable
		return d
	}

	d.FirstAt = versions[0].CreatedAt
	d.LastAt = versions[len(versions)-1].CreatedAt
	d.Span = d.LastAt.Sub(d.FirstAt)

	cur := passportSession{Start: versions[0].CreatedAt, End: versions[0].CreatedAt, Snapshots: 1}

	// Baseline for the running session's net-words figure. Later sessions
	// measure from the *previous* session's final count, not from their own
	// first snapshot, because that first snapshot already contains the few
	// minutes of typing that preceded it — measuring from it would drop that
	// work. The first session is the exception: it measures from its own first
	// snapshot rather than from zero, so a history whose early snapshots were
	// pruned understates session one instead of reporting the words it never
	// saw as a sudden addition.
	startWords := versions[0].WordCount

	for i := 1; i < len(versions); i++ {
		v, prev := versions[i], versions[i-1]

		if delta := v.WordCount - prev.WordCount; delta > d.LargestJump {
			d.LargestJump, d.LargestJumpAt, d.LargestJumpIdx = delta, v.CreatedAt, i
		}

		if v.CreatedAt.Sub(prev.CreatedAt) > sessionGap {
			cur.WordsAdded = prev.WordCount - startWords
			d.Sessions = append(d.Sessions, cur)
			cur = passportSession{Start: v.CreatedAt, End: v.CreatedAt, Snapshots: 1}
			startWords = prev.WordCount
			continue
		}
		cur.End = v.CreatedAt
		cur.Snapshots++
	}
	cur.WordsAdded = versions[len(versions)-1].WordCount - startWords
	d.Sessions = append(d.Sessions, cur)

	for _, s := range d.Sessions {
		d.ActiveTime += s.Duration()
	}

	final := versions[len(versions)-1].WordCount
	d.NoteJump = final > 0 && float64(d.LargestJump)/float64(final) > jumpNoteThreshold

	d.ChainStatus, d.UnhashedCount = verifyChain(doc, versions)
	return d
}

// verifyChain recomputes every snapshot's hash and checks that each links to the
// one before it. Returns the outcome and how many snapshots predate the chain.
//
// Broken *links* are not evidence of tampering on their own: auto-snapshot
// pruning legitimately removes rows from the middle of the history, which severs
// the links across the hole. So a link break is reported as a gap unless the
// document is in preserve-history mode, where nothing should ever be removed. A
// hash that fails to match its *own* contents is unambiguous, and always broken.
func verifyChain(doc db.Document, versions []db.DocumentVersion) (status string, unhashed int) {
	var (
		hashed    int
		linkBreak bool
		prevHash  string
		havePrev  bool
	)

	for _, v := range versions {
		if v.ContentHash == "" {
			unhashed++
			havePrev = false // can't vouch for what follows an unhashed row
			continue
		}
		hashed++

		want := chainHash(v.PrevHash, v.DocID, v.CreatedAt, v.WordCount, v.ContentText)
		if want != v.ContentHash {
			return chainBroken, unhashed
		}
		if havePrev && v.PrevHash != prevHash {
			linkBreak = true
		}
		prevHash, havePrev = v.ContentHash, true
	}

	switch {
	case hashed == 0:
		return chainUnverifiable, unhashed
	case linkBreak && doc.PreserveHistory:
		// Nothing should have been removed from a preserved history.
		return chainBroken, unhashed
	case linkBreak:
		return chainGaps, unhashed
	case unhashed > 0:
		return chainPartial, unhashed
	default:
		return chainVerified, unhashed
	}
}

// --- HTTP -------------------------------------------------------------------

// passport renders the report for one document as a standalone HTML download.
// HTML rather than PDF for the same reason as the other exports: a CJK-safe PDF
// needs an embedded Unicode font or a headless browser. The page is styled for
// printing, so "Save as PDF" in the browser produces the handoff artifact.
func (h *Handler) passport(w http.ResponseWriter, r *http.Request) {
	docID := chi.URLParam(r, "id")

	doc, err := h.fetch(docID)
	if errors.Is(err, sql.ErrNoRows) {
		notFound(w)
		return
	}
	if err != nil {
		httputil.ServerError(w, err)
		return
	}

	versions, err := h.passportVersions(docID)
	if err != nil {
		httputil.ServerError(w, err)
		return
	}

	body := renderPassport(buildPassport(doc, versions))

	filename := sanitizeFilename(doc.Title)
	if filename == "" {
		filename = "untitled"
	}
	filename += " - writing passport.html"

	w.Header().Set("Content-Type", "text/html; charset=utf-8")
	w.Header().Set("Content-Disposition",
		fmt.Sprintf("attachment; filename*=UTF-8''%s", urlEscapeFilename(filename)))
	w.Header().Set("Content-Length", fmt.Sprintf("%d", len(body)))
	_, _ = w.Write(body)
}

// passportVersions loads every snapshot oldest-first with the fields the report
// and the chain check need — including content_text, which the list endpoint
// omits as too heavy but verification cannot do without.
func (h *Handler) passportVersions(docID string) ([]db.DocumentVersion, error) {
	rows, err := h.DB.Query(
		`SELECT v.id, v.doc_id, v.title, v.content_text, v.word_count, v.kind,
		        v.created_at, v.content_hash, v.prev_hash
		   FROM document_versions v
		   JOIN documents d ON d.id = v.doc_id
		  WHERE v.doc_id = ? AND d.user_id = ?
		  ORDER BY v.created_at ASC, v.rowid ASC`,
		docID, db.LocalUserID,
	)
	if err != nil {
		return nil, err
	}
	defer rows.Close()

	var out []db.DocumentVersion
	for rows.Next() {
		var v db.DocumentVersion
		if err := rows.Scan(
			&v.ID, &v.DocID, &v.Title, &v.ContentText, &v.WordCount, &v.Kind,
			&v.CreatedAt, &v.ContentHash, &v.PrevHash,
		); err != nil {
			return nil, err
		}
		out = append(out, v)
	}
	return out, rows.Err()
}