Skip to content
go-libfossil Artifact [51dae6e7df]
Login

Artifact [51dae6e7df]

Check-ins Using Download Hex

Artifact 51dae6e7df562df632a2c0f36851a93358f20c240aec62acba2ab08c64d33438:

package sync

import (
	"bytes"
	"context"
	"database/sql"
	"errors"
	"fmt"
	"strconv"
	"strings"

	"github.com/danmestas/go-libfossil/db"
	"github.com/danmestas/go-libfossil/internal/auth"
	"github.com/danmestas/go-libfossil/internal/blob"
	"github.com/danmestas/go-libfossil/internal/content"
	"github.com/danmestas/go-libfossil/internal/hash"
	"github.com/danmestas/go-libfossil/internal/manifest"
	"github.com/danmestas/go-libfossil/internal/repo"
	"github.com/danmestas/go-libfossil/internal/shun"
	"github.com/danmestas/go-libfossil/internal/xfer"
	"github.com/danmestas/go-libfossil/simio"

	libfossil "github.com/danmestas/go-libfossil/internal/fsltype"
)

// DefaultCloneBatchBytes bounds the wire bytes a single clone round emits.
//
// A clone batch is bounded by output size, never by artifact count. A count
// cannot adapt to artifact size: 200 tiny control artifacts and 200 large
// blobs produce wildly different messages, and neither corresponds to what
// the transport can carry. §8.2 leaves the bound entirely to the server
// ("clients MUST NOT infer count or size from range"), so the only real
// constraint is that the response stays a size a server can hold and that a
// repository drains in rounds proportional to its bytes, not its artifacts.
//
// emitCloneBatch now retransmits stored content per §7.2 (issue #98) rather
// than the fully expanded artifacts it used to send, so bytesSent tracks the
// same quantity canonical's 5,000,000 max-download bounds: repository bytes,
// not the 80x-220x-larger expanded form. This constant was chosen before that
// fix, against the expanded-bytes regime documented in issue #88 where a
// smaller value pushed fossil- and sqlite-sized corpora past MaxRounds --
// that constraint is gone now that wire bytes and repository bytes are the
// same quantity. Retuning this value toward canonical's constant is issue
// #109's territory (server pagination/budget accounting), not this fix's; it
// is left at its prior value here to avoid colliding with that work in
// flight. 16,000,000 stays a safe, if generous, upper bound either way: it
// only widens how much stored content one round may carry.
//
// This bounds an ordinary round at the budget plus the artifact that crosses
// it, and emitCloneBatch caps that sum at xfer.MaxDecompressedBytes so the
// client can always decode a round (issue #109). The one exception is a single
// artifact larger than that bound: it is sent alone, in a round of its own, to
// make progress -- so a repository holding one 2 GB artifact still emits a 2 GB
// round, but such an artifact is unclonable regardless. Measured worst case is
// ~25 MB across the three repositories above, since the largest single
// artifacts are 9.4 MB (libfossil) and 17.0 MB (sqlite). Against the count
// bound's unbounded 44.9 MB that is a large reduction, and concurrent clones
// multiply it.
const DefaultCloneBatchBytes = 16_000_000

// A round this server emits must stay inside the bound the client applies when
// it decompresses one, or a clone between two libfossil peers fails on a
// message this same code produced (issue #104). emitCloneBatch enforces that at
// runtime: it flushes a round rather than let an artifact carry it past
// xfer.MaxDecompressedBytes, and sends an over-bound artifact alone (issue
// #109). This compile-time guard makes that enforcement well-conditioned: with
// the bound at least twice the budget, filler accumulates up to one budget and
// any artifact up to a second budget still fits in the same round, so ordinary
// artifacts ride along and only genuinely large ones (above bound minus budget)
// take a round of their own. Raising DefaultCloneBatchBytes past half of
// xfer.MaxDecompressedBytes underflows this unsigned expression and fails the
// build rather than the clone.
const _ = uint(xfer.MaxDecompressedBytes - 2*DefaultCloneBatchBytes)

// cloneCardOverheadBytes over-approximates the non-payload bytes of an
// encoded cfile card -- keyword, hash, delta-source hash, two decimal
// lengths, and newline come to somewhat more than the plain "file" card's
// ~58 -- so that a repository of zero-length artifacts still paginates
// instead of emitting the whole blob table into a single response. The slack
// also absorbs the PrivateCard that may precede a card, which is not charged
// separately. This is charged against data before encodeCFile's zlib
// compression runs, so it is already conservative in the same direction as
// the framing overhead it approximates. Over-charging is the safe direction
// and cannot be defeated: a corpus of nothing but zero-length artifacts caps
// at ~167,000 cards per round, about 9.7 MB of actual output, still inside
// the budget.
const cloneCardOverheadBytes = 96

// cloneCapPrefix tags a CookieCard value that carries the upper rid bound
// of a clone session. The handler captures max(rid) on the first clone round
// and echoes it via CookieCard; the client returns it on every subsequent
// round so the (otherwise stateless) server can scope its blob query and
// converge even if the hub is being written to during the clone session.
const cloneCapPrefix = "lf-clone-cap:"

// parseCloneCap returns the snapshot rid encoded in a CookieCard value,
// or (0, false) if the value isn't a clone-cap cookie. Non-matching
// cookies (e.g. arbitrary sync session cookies) leave the bound at zero,
// which preserves pre-snapshot behavior for forward/back compat.
func parseCloneCap(v string) (int, bool) {
	if !strings.HasPrefix(v, cloneCapPrefix) {
		return 0, false
	}
	rid, err := strconv.Atoi(v[len(cloneCapPrefix):])
	if err != nil || rid < 0 {
		return 0, false
	}
	return rid, true
}

// HandleFunc is the server-side sync handler signature.
// Transport listeners call this with decoded requests and write back the response.
type HandleFunc func(ctx context.Context, r *repo.Repo, req *xfer.Message) (*xfer.Message, error)

// HandleOpts configures optional behavior for HandleSync.
type HandleOpts struct {
	Buggify      BuggifyChecker // nil in production.
	Observer     Observer       // nil defaults to no-op.
	ContentCache *content.Cache // nil = no caching.
	Clock        simio.Clock    // time for the reply's timestamp comment; nil = real time
}

// HandleSync processes an incoming xfer request and produces a response.
// Stateless per-round — the client drives convergence.
func HandleSync(ctx context.Context, r *repo.Repo, req *xfer.Message) (*xfer.Message, error) {
	return HandleSyncWithOpts(ctx, r, req, HandleOpts{})
}

// HandleSyncWithOpts processes an incoming xfer request with optional
// fault injection. Used by DST harness; production callers use HandleSync.
func HandleSyncWithOpts(ctx context.Context, r *repo.Repo, req *xfer.Message, opts HandleOpts) (*xfer.Message, error) {
	if r == nil {
		panic("sync.HandleSync: r must not be nil")
	}
	if req == nil {
		panic("sync.HandleSync: req must not be nil")
	}

	obs := resolveObserver(opts.Observer)
	ctx = obs.HandleStarted(ctx, HandleStart{
		Operation: detectOperation(req),
	})

	h := &handler{repo: r, buggify: opts.Buggify, cache: opts.ContentCache, clock: opts.Clock}
	if h.clock == nil {
		h.clock = simio.RealClock{}
	}
	resp, err := h.process(ctx, req)
	if err == nil && resp == nil {
		panic("sync.HandleSync: resp must not be nil on success")
	}

	obs.HandleCompleted(ctx, HandleEnd{
		CardsProcessed: len(req.Cards),
		FilesSent:      h.filesSent,
		FilesReceived:  h.filesRecvd,
		Err:            err,
	})
	return resp, err
}

// detectOperation checks request cards to determine if this is a clone or sync.
func detectOperation(req *xfer.Message) string {
	for _, c := range req.Cards {
		if _, ok := c.(*xfer.CloneCard); ok {
			return "clone"
		}
	}
	return "sync"
}

// remoteHasEntry records that the client announced a blob via igot.
type remoteHasEntry struct {
	isPrivate bool // IsPrivate flag from the client's igot card
}

// handler holds per-request state while processing cards.
type handler struct {
	repo            *repo.Repo
	buggify         BuggifyChecker
	resp            []xfer.Card
	pushOK          bool                      // client sent a valid push card
	pullOK          bool                      // client sent a valid pull card
	cloneMode       bool                      // client sent a clone card
	fatal           bool                      // a card rule ended the request; resp holds only its error
	broke           bool                      // card loop broken; the post-loop still runs
	errsReset       int                       // counted errors a reset discarded (§5.3 errors N)
	projectCode     string                    // the repository's project-code
	serverCode      string                    // the repository's server-code
	uncompressed    bool                      // reply goes out uncompressed (clone v3)
	cloneDone       bool                      // clone card has no usable cursor: send nothing
	peerVersion     int                       // client's release, from client-version
	peerDate        int                       // client's check-in date, from client-version
	cloneSeq        int                       // rid of the next blob to send, from the client's clone card
	cloneSnapMax    int                       // upper rid bound for this clone session (from cookie); 0 = capture fresh in emitCloneBatch
	uvCatalogSent   bool                      // true after sending UV catalog
	pendingPhantoms []pendingPhantom          // igot names awaiting flushPhantoms
	reqClusters     bool                      // client sent pragma req-clusters
	filesSent       int                       // files sent in response (for observer)
	filesRecvd      int                       // files received from client (for observer)
	syncPrivate     bool                      // true if pragma send-private was accepted
	nextIsPrivate   bool                      // true if a private card precedes the next file/cfile
	syncedTables    map[string]*SyncedTable   // cached table definitions
	xrowsSent       int                       // table sync rows sent
	xrowsRecvd      int                       // table sync rows received
	cache           *content.Cache            // nil = passthrough to content.Expand
	clock           simio.Clock               // time for the reply's timestamp comment
	remoteHas       map[string]remoteHasEntry // UUIDs the client announced via igot (mirrors Fossil's onremote table)
	remoteLacks     map[string]bool           // UUIDs the client asked for (fossil's unk)
	logins          int                       // login cards read so far; only one is allowed

	// Auth state
	user   string // verified username ("nobody" if no login card)
	caps   string // capability string from user table
	authed bool   // whether login card was verified
	// cloneRead is Read derived from Clone before any login, as page_xfer
	// derives it from nobody's capabilities ahead of its card loop.
	cloneRead bool
}

func (h *handler) initAuth() {
	h.user = "nobody"
	h.caps = ""
	h.authed = false
	var caps string
	err := h.repo.DB().QueryRow("SELECT cap FROM user WHERE login='nobody'").Scan(&caps)
	if err == nil {
		h.caps = caps
	}
	h.cloneRead = auth.CanClone(h.caps)
}

// canRead reports the requester's Read permission for this request, the one
// owner of that decision for every read gate the handler applies: the
// unversioned catalog, uvigot and uvgimme answers, xtable-hash and
// read-gated reqconfig.
//
// Clone implies Read for the capabilities a request starts with, nobody's,
// as fossil's page_xfer sets it (`if( g.perm.Clone ) g.perm.Read = 1;`,
// xfer.c:1301) before its card loop and the §7.3 erratum states. A Clone
// that a login card grants does not become Read: check_login only adds the
// user's capabilities, after that point.
//
// Pull is deliberately not one of these gates. fossil trunk's derived Read
// also lets a Clone-only nobody pull; here a pull still needs o
// (auth.CanPull), by the owner's decision.
func (h *handler) canRead() bool {
	return auth.CanRead(h.caps) || h.cloneRead
}

// handleLoginCard verifies a login card and adds the user's capabilities to
// nobody's, as fossil's check_login does through login_set_capabilities: a
// login only grants. It reports the refusal's error text, "" when the login
// is accepted. fossil does not go on serving the request as nobody after a
// refusal, and nor does this: the caller ends card processing there.
func (h *handler) handleLoginCard(c *xfer.LoginCard) string {
	// A login's shared secret is salted with the project code; a repository
	// without one can verify no login, and fossil's page_xfer stops there
	// too ("missing project code"). Refuse it rather than ask VerifyLogin,
	// which takes an empty project code as a programmer error.
	if h.projectCode == "" {
		return "login failed"
	}
	u, err := auth.VerifyLogin(h.repo.DB(), h.projectCode, c)
	if err != nil {
		return "login failed"
	}
	h.user = u.Login
	h.caps += u.Cap
	h.authed = true
	return ""
}

// handleLogin answers a login card at its own position in the request, as
// page_xfer's check_login runs where the card is read: the capabilities it
// grants apply to the cards after it, never to those before. Only one is
// allowed (§6.1, src/xfer.c "error multiple\slogin\scards"). A refusal is
// page_xfer's reset, error and break, so the cards before it keep their
// effect and the end-of-request steps still run (§6.4).
func (h *handler) handleLogin(c *xfer.LoginCard) {
	if c == nil {
		panic("handler.handleLogin: c must not be nil")
	}
	h.logins++
	if h.logins > 1 {
		h.breakLoop("multiple login cards")
		return
	}
	if msg := h.handleLoginCard(c); msg != "" {
		h.breakLoop(msg)
	}
}

// fail ends the request with a single error card, discarding the reply built
// so far, as fossil's cgi_reset_content before an error does.
func (h *handler) fail(msg string) {
	h.resetContent()
	h.resp = []xfer.Card{&xfer.ErrorCard{Message: msg}}
	h.fatal = true
}

// resetContent discards the reply built so far, as fossil's cgi_reset_content
// does before a card-level error. Only the reply is cleared; any blob already
// stored this round stays (fossil keeps its mutations too, §10.1). The errors
// it discards are still counted in the closing comment (§5.3).
func (h *handler) resetContent() {
	h.errsReset += countedErrors(h.resp)
	h.resp = nil
}

// badCommandPrefix begins the error that answers an unknown or malformed card.
const badCommandPrefix = "bad command: "

// badCommand answers a line no card rule accepts. §5.2, src/xfer.c page_xfer's
// final "else": the reply so far is reset, an "error bad\scommand:\s<LINE>"
// card is emitted, and parsing goes on -- so a later card still takes effect.
// line keeps its terminating LF, which the error quotes verbatim.
func (h *handler) badCommand(line string) {
	h.resetContent()
	h.resp = append(h.resp, &xfer.ErrorCard{Message: badCommandPrefix + line})
}

// replyOnlyLine re-encodes a card the server should not have received (error,
// message -- reply-only in fossil's page_xfer, which has no branch for them)
// into the wire line it arrived as, so badCommand can quote it the way fossil
// quotes blob_str(&xfer.line). The LF EncodeCard writes is the one the error
// text carries.
func replyOnlyLine(c xfer.Card) string {
	var buf bytes.Buffer
	if err := xfer.EncodeCard(&buf, c); err != nil {
		// error and message cards always encode; this is a programmer error.
		panic("handler.replyOnlyLine: " + err.Error())
	}
	return buf.String()
}

// breakLoop answers a card with fossil's cgi_reset_content, error and break:
// the reply so far is reset to one error card and no later request card is
// processed, but the post-loop work and the timestamp comment still follow,
// as they do after page_xfer's card loop.
func (h *handler) breakLoop(msg string) {
	h.resetContent()
	h.resp = append(h.resp, &xfer.ErrorCard{Message: msg})
	h.broke = true
	// The reset also discards the cfiles page_xfer's clone card already
	// sent, so no clone batch or clone push card follows the error, and the
	// cluster igots it answered req-clusters with in place.
	h.cloneMode = false
	h.reqClusters = false
}

func (h *handler) process(ctx context.Context, req *xfer.Message) (*xfer.Message, error) {
	// Initialize auth state from nobody user.
	h.initAuth()

	// A login's shared secret and every push and pull card are checked
	// against the project code. A repository without one leaves it empty,
	// which no artifact hash matches.
	if err := h.loadCodes(); err != nil {
		return nil, err
	}

	// Load synced tables.
	if err := h.loadSyncedTables(); err != nil {
		return nil, err
	}

	if h.probeCloneFatal(req.Cards) {
		return h.reply(), nil
	}
	if err := h.processCards(ctx, req.Cards); err != nil {
		return nil, err
	}
	if h.fatal {
		return h.reply(), nil
	}
	// The end-of-request steps run after a broken loop too (§6.4).
	if err := h.emitInventory(); err != nil {
		return nil, err
	}

	if h.cloneMode {
		if err := h.emitClonePush(); err != nil {
			return nil, err
		}
	}

	if err := h.crosslinkReceived(ctx); err != nil {
		return nil, err
	}

	h.resp = append(h.resp, h.timestampComment())
	return h.reply(), nil
}

// processCards handles the request's cards in request order, as page_xfer's
// card loop reads them, until one ends the request (h.fatal) or breaks the
// loop. Each card sees only the state the cards before it established: a
// file card before its push card is refused, and a card before a login card
// has nobody's capabilities (§5.2, §6.4, §6.6).
func (h *handler) processCards(ctx context.Context, cards []xfer.Card) error {
	for _, card := range cards {
		// Phantoms from a run of igot cards are written in one transaction,
		// as fossil writes them inside page_xfer's single write
		// transaction; any other card sees them already stored.
		if _, isIGot := card.(*xfer.IGotCard); !isIGot {
			if err := h.flushPhantoms(); err != nil {
				return err
			}
		}
		if err := h.processCard(ctx, card); err != nil {
			return err
		}
		if h.fatal || h.broke {
			break
		}
	}
	if h.fatal {
		// A fatal rolls the request back (§8.1): nothing pending is kept.
		h.pendingPhantoms = nil
		return nil
	}
	return h.flushPhantoms()
}

// processCard handles one request card.
func (h *handler) processCard(ctx context.Context, card xfer.Card) error {
	if card == nil {
		panic("handler.processCard: card must not be nil")
	}
	switch c := card.(type) {
	case *xfer.LoginCard:
		h.handleLogin(c)
	case *xfer.PragmaCard, *xfer.PushCard, *xfer.PullCard, *xfer.CloneCard,
		*xfer.CloneSeqNoCard, *xfer.CookieCard, *xfer.SchemaCard:
		h.handleControlCard(card)
	case *xfer.BadFileCard:
		// A file/cfile/uvfile line the decoder could not frame: page_xfer
		// resets its reply, emits the handler's error and breaks (§5.2).
		h.breakLoop(c.Message)
	case *xfer.BadCommandCard:
		h.badCommand(c.Line)
	case *xfer.ErrorCard, *xfer.MessageCard:
		// error and message are reply-only: a client that sends one gets a
		// bad command, as fossil's page_xfer has no branch for them (§5.2).
		h.badCommand(replyOnlyLine(c))
	default:
		return h.handleDataCard(ctx, card)
	}
	return nil
}

// probeCloneFatal answers a request that reaches §8.1's invalid clone
// sequence fatal, and reports whether it did. page_xfer answers that fatal
// by rolling back the whole request, so nothing the cards before it stored
// survives. This handler stores as it goes, with no request transaction to
// roll back, so a request holding such a clone card is first walked without
// side effects: only the cards that decide whether the clone card is reached,
// and with which capabilities, are acted on. When the walk ends in the
// fatal, its reply is the answer and no card has written anything.
//
// The walk stores no file, so a file card that fails validation does not
// break it as it breaks the real loop. A request that pushes a bad file
// ahead of an invalid clone card therefore gets the clone error where
// fossil sends the file's; no conforming client sends push and clone
// together (§8.5).
func (h *handler) probeCloneFatal(cards []xfer.Card) bool {
	if !hasInvalidCloneSeq(cards) {
		return false
	}
	probe := &handler{repo: h.repo, clock: h.clock,
		projectCode: h.projectCode, serverCode: h.serverCode}
	probe.initAuth()
	for _, card := range cards {
		probe.probeCard(card)
		if probe.fatal || probe.broke {
			break
		}
	}
	if !probe.fatal {
		return false
	}
	h.resp, h.fatal = probe.resp, true
	h.uncompressed = probe.uncompressed
	return true
}

// hasInvalidCloneSeq reports whether a clone card in cards would meet
// §8.1's fatal if an authorized requester reached it.
func hasInvalidCloneSeq(cards []xfer.Card) bool {
	for _, card := range cards {
		c, ok := card.(*xfer.CloneCard)
		if ok && c.SeqNoIsDecimal && c.Version >= 2 && c.SeqNo <= 0 {
			return true
		}
	}
	return false
}

// probeCard is processCard for probeCloneFatal: the cards that can end the
// card loop or change the capabilities, none of which writes to the
// repository.
func (h *handler) probeCard(card xfer.Card) {
	if card == nil {
		panic("handler.probeCard: card must not be nil")
	}
	switch c := card.(type) {
	case *xfer.LoginCard:
		h.handleLogin(c)
	case *xfer.PushCard:
		h.handlePushCard(c)
	case *xfer.PullCard:
		h.handlePullCard(c)
	case *xfer.CloneCard:
		h.handleCloneCard(c)
	case *xfer.FileCard, *xfer.CFileCard:
		if !h.pushOK {
			h.breakLoop("not authorized to write")
		}
	case *xfer.BadFileCard:
		h.breakLoop(c.Message)
	case *xfer.ConfigCard:
		h.handleConfigPush()
	}
}

// crosslinkReceived crosslinks any newly-received manifests into the
// relational tables (event/leaf/plink/mlink/tagxref). Without this the
// receiving repo stores blobs durably but exposes nothing on the timeline,
// fork detection sees stale leaf state, and the clone protocol cannot
// traverse manifest parents (mlink is empty).
//
// It walks the scanner only when we accepted files this round; pure
// pull/igot rounds add nothing relational to update. It is context-aware
// (#120): the sweep walks the whole repository in one call, so on a large
// repo it is where a request whose client has already gone away would
// otherwise keep working uninterrupted.
func (h *handler) crosslinkReceived(ctx context.Context) error {
	if h.filesRecvd == 0 {
		return nil
	}
	if _, err := manifest.CrosslinkContext(ctx, h.repo); err != nil {
		return fmt.Errorf("HandleSync: crosslink: %w", err)
	}
	return nil
}

// reply is the message the handler answers with.
func (h *handler) reply() *xfer.Message {
	return &xfer.Message{Cards: h.resp, Uncompressed: h.uncompressed}
}

// loadCodes reads the repository's project and server codes. A login's
// shared secret and every push and pull card are checked against the project
// code; a clone's push card carries both. A repository without one leaves it
// empty, which no artifact hash matches.
func (h *handler) loadCodes() error {
	for _, c := range []struct {
		name string
		dst  *string
	}{{"project-code", &h.projectCode}, {"server-code", &h.serverCode}} {
		err := h.repo.DB().QueryRow(
			"SELECT value FROM config WHERE name=?", c.name).Scan(c.dst)
		if err != nil && !errors.Is(err, sql.ErrNoRows) {
			return fmt.Errorf("HandleSync: %s: %w", c.name, err)
		}
	}
	return nil
}

// emitClonePush emits the push card with project-code and server-code so
// the clone client can identify the repo. Only in clone mode — sync clients
// already have codes, and real Fossil treats server-sent "push" as unknown
// during sync.
//
// This must trail the clone batch's clone_seqno card, matching canonical's
// order (fossil-scm xfer.c emits the batch, then clone_seqno at :1571, then
// push at :1577). A real fossil client re-issues `clone 3 SEQNO` every round
// it sees a `push` card while its clone cursor is still positive (xfer.c:2706),
// and it learns the cursor hit zero only from the clone_seqno card. Emitting
// push ahead of clone_seqno makes the client queue one more clone request
// before it sees the terminal clone_seqno 0, so the server serves the whole
// repository a second time — the 2.06x end-to-end blow-up of issue #138.
func (h *handler) emitClonePush() error {
	if !h.cloneMode {
		panic("handler.emitClonePush: not a clone")
	}
	if h.projectCode != "" {
		h.resp = append(h.resp, &xfer.PushCard{
			ProjectCode: h.projectCode,
			ServerCode:  h.serverCode,
		})
	}
	return nil
}

// timestampComment is the comment that ends a completed reply, as fossil's
// page_xfer writes it: the server's UTC time and the request's error count
// (§5.3). A fail-fast reply goes without it.
func (h *handler) timestampComment() *xfer.CommentCard {
	if h.clock == nil {
		panic("handler.timestampComment: nil clock")
	}
	if h.errsReset < 0 {
		panic("handler.timestampComment: negative reset error count")
	}
	errs := h.errsReset + countedErrors(h.resp)
	now := h.clock.Now().UTC().Format("2006-01-02T15:04:05")
	return &xfer.CommentCard{Text: fmt.Sprintf("timestamp %s errors %d", now, errs)}
}

// privateRefusal is the error a requester without the private capability
// gets for pragma send-private or a private card.
const privateRefusal = "not authorized to sync private content"

// countedErrors counts the error cards that §5.3's errors N counts. N is the
// errors raised, as page_xfer's nErr is, and three of them fossil never adds
// to nErr: the unknown-operator error, the private-content refusal and the
// SHA3 version refusal (erratum sync-protocol-01-s5.3-errors-count). The
// count is made from card text because those three texts are fixed by the
// draft and no other error carries them.
func countedErrors(cards []xfer.Card) int {
	n := 0
	for _, c := range cards {
		e, ok := c.(*xfer.ErrorCard)
		if !ok {
			continue
		}
		switch {
		case strings.HasPrefix(e.Message, badCommandPrefix),
			e.Message == privateRefusal,
			e.Message == sha3Refusal().Message:
			continue
		}
		n++
	}
	return n
}

// emitInventory sends what follows the card loop: the cluster igots a
// req-clusters asked for, the igot inventory of a pull, and a clone's batch.
func (h *handler) emitInventory() error {
	// pragma req-clusters is answered with the cluster igots, ahead of the
	// inventory as page_xfer answers it in card order. Like fossil, any
	// requester gets them: they name public artifacts and carry no content,
	// which a gimme still needs Read to fetch.
	if h.reqClusters {
		if err := h.sendAllClusters(); err != nil {
			return err
		}
	}
	// An accepted push asks for the server's phantoms, ahead of the pull's
	// inventory as page_xfer's request_phantoms runs ahead of
	// send_unclustered (§6.4, §8.5).
	if h.pushOK {
		if err := h.requestPhantoms(); err != nil {
			return err
		}
	}
	// If pull was requested, emit igot for all non-phantom blobs.
	if h.pullOK {
		if err := h.emitIGots(); err != nil {
			return err
		}
		if err := h.emitXIGots(); err != nil {
			return err
		}
	}
	// If clone, emit paginated file cards, after the unversioned catalog
	// page_xfer sends with a clone when uv-sync is on (§8.4).
	if h.cloneMode {
		if err := h.sendCloneUVCatalog(); err != nil {
			return err
		}
		if h.cloneDone {
			h.resp = append(h.resp, &xfer.CloneSeqNoCard{SeqNo: 0})
			return nil
		}
		return h.emitCloneBatch()
	}
	return nil
}

func (h *handler) handleControlCard(card xfer.Card) {
	switch c := card.(type) {
	case *xfer.PragmaCard:
		h.handlePragma(c)
	case *xfer.PushCard:
		h.handlePushCard(c)
	case *xfer.PullCard:
		h.handlePullCard(c)
	case *xfer.CloneCard:
		h.handleCloneCard(c)
	case *xfer.CloneSeqNoCard:
		// Deliberately ignored. clone_seqno is a reply-only card, so no
		// conforming client sends one; canonical answers `bad command`. We
		// stay tolerant rather than strict, which is why the model server in
		// TestCloneMultiBatchAgainstCanonicalServer rejects it and we do not.
	case *xfer.CookieCard:
		// Clone clients echo the cap cookie on every round. Sync sessions also
		// use cookies opaquely; non-matching values leave cloneSnapMax at zero.
		if rid, ok := parseCloneCap(c.Value); ok {
			h.cloneSnapMax = rid
		}
	case *xfer.SchemaCard:
		h.handleSchemaCard(c)
	}
}

// handlePushCard accepts a push from a requester with Write, after the
// project check.
func (h *handler) handlePushCard(c *xfer.PushCard) {
	if h.wrongProject(c.ProjectCode) {
		return
	}
	if auth.CanPush(h.caps) {
		h.pushOK = true
		return
	}
	if h.pullOK {
		// A requester who may read but not write is told to pull only,
		// not refused, as fossil's page_xfer answers a push card with
		// "message pull only - not authorized to push" once a pull card
		// has set isPull. The pull still proceeds (§10.2).
		h.resp = append(h.resp, &xfer.MessageCard{
			Message: "pull only - not authorized to push",
		})
		return
	}
	// Otherwise the reply so far is reset and the push refused, but card
	// processing goes on (§6.6): page_xfer has no break here.
	h.resetContent()
	h.resp = append(h.resp, &xfer.ErrorCard{Message: "not authorized to write"})
}

// handlePullCard accepts a pull from a requester with Read, after the project
// check.
func (h *handler) handlePullCard(c *xfer.PullCard) {
	if h.wrongProject(c.ProjectCode) {
		return
	}
	if auth.CanPull(h.caps) {
		h.pullOK = true
		return
	}
	// page_xfer resets, refuses and breaks (§6.6, §10.1).
	h.breakLoop("not authorized to read")
}

// handleCloneCard starts a clone for a requester with Clone, at the rid the
// card's SEQNO names.
func (h *handler) handleCloneCard(c *xfer.CloneCard) {
	if !auth.CanClone(h.caps) {
		h.refuseClone()
		return
	}
	// §4.1, §8.1: page_xfer sends a version 3 clone's reply uncompressed,
	// and sets that before the sequence-number fatal below.
	h.uncompressed = c.Version >= 3
	// §8.1: an otherwise authorized three-token clone with VERSION >= 2
	// and a digit-only SEQNO of zero or less clears accumulated output,
	// rolls back, emits this error, and parses no later request card.
	// probeCloneFatal reaches this ahead of the card loop, before any card
	// has written, which is what stands in for the rollback.
	if c.SeqNoIsDecimal && c.Version >= 2 && c.SeqNo <= 0 {
		h.fail("invalid clone sequence number")
		return
	}
	h.cloneMode = true
	// The pagination cursor rides on the clone card and nowhere else
	// (§8.1; canonical xfer.c:1553 reads token 2 of `clone VERSION
	// SEQNO`). §8.1 withholds the fatal above from a negative SEQNO, which
	// fails digit-only recognition; page_xfer then has no cursor at all, so
	// no content is sent and the reply says the clone is done. A bare
	// `clone` starts from the first rid.
	if c.Version >= 2 && c.SeqNo < 0 {
		h.cloneDone = true
		return
	}
	h.cloneSeq = c.SeqNo
	if h.cloneSeq <= 0 {
		h.cloneSeq = 1
	}
}

// refuseClone answers a clone from a requester without Clone as page_xfer
// does (§8.2, §10.1): the reply so far is reset, the push card naming the
// repository goes ahead of the error, and no later card is processed.
func (h *handler) refuseClone() {
	h.resetContent()
	serverCode := h.serverCode
	if serverCode == "" {
		serverCode = "x" // db_get("server-code", "x")
	}
	h.resp = append(h.resp,
		&xfer.PushCard{ServerCode: serverCode, ProjectCode: h.projectCode},
		&xfer.ErrorCard{Message: "not authorized to clone"})
	h.broke = true
}

// wrongProject refuses a push or pull card naming another project, ahead of
// any capability check, as page_xfer does with "error wrong\sproject" and a
// break (§8.5). The servercode beside it is informational and not checked.
// Cards are handled in request order, so a file card ahead of it has been
// stored, as fossil stores it.
func (h *handler) wrongProject(code string) bool {
	if code == h.projectCode {
		return false
	}
	h.breakLoop("wrong project")
	return true
}

// handlePragma acts on the pragmas the server understands and ignores the
// rest, client-version included.
//
// Pragmas are handled in card order, so a pull card later in the request is
// not yet known here. The catalogs are therefore gated on
// the Read capability itself, as fossil's page_xfer gates uv-hash on
// g.perm.Read, rather than on an accepted pull.
func (h *handler) handlePragma(c *xfer.PragmaCard) {
	if c == nil {
		panic("handler.handlePragma: c must not be nil")
	}
	switch {
	case c.Name == "client-version" && len(c.Values) >= 1:
		// §9.1: the client's release gates SHA3 content (§7.6). A
		// client-version carrying DATE and TIME is answered in place, as
		// page_xfer answers it, so the answer precedes any content.
		h.peerVersion = peerVersion(c)
		h.peerDate = peerDate(c)
		if len(c.Values) >= 3 {
			h.resp = append(h.resp, versionPragma("server-version"))
		}
	case c.Name == "uv-hash" && len(c.Values) >= 1:
		if err := h.handlePragmaUVHash(c.Values[0]); err != nil {
			h.resp = append(h.resp, &xfer.ErrorCard{
				Message: fmt.Sprintf("uv-hash: %v", err),
			})
		}
	case c.Name == "xtable-hash" && len(c.Values) >= 2:
		// The inventory serves both directions of table sync: xgimme
		// pulls rows from it, and a pushing client queues the rows it
		// holds newer (client_tablesync.go handleXIGotResponse). So it
		// needs Read and a direction card, not xgimme's pull alone. A
		// Clone-only requester has Read by derivation (#312) but sends no
		// direction card, so it gets no inventory (#351). Table sync is a
		// go-libfossil extension with no fossil counterpart.
		if h.canRead() && (h.pullOK || h.pushOK) {
			h.handlePragmaXTableHash(c.Values[0], c.Values[1])
		}
	case c.Name == "req-clusters":
		h.reqClusters = true
	case c.Name == "ci-lock" && len(c.Values) >= 2:
		// Fossil takes a lock for any requester. A lock only matters to
		// someone who can commit, and ci-lock-fail names the holder, so
		// this server takes and reports locks for writers alone.
		if !auth.CanPush(h.caps) {
			return
		}
		fail := processCkinLock(h.repo.DB(), c.Values[0], c.Values[1], h.user,
			DefaultCkinLockTimeout)
		if fail != nil {
			h.resp = append(h.resp, &xfer.PragmaCard{
				Name:   "ci-lock-fail",
				Values: []string{fail.HeldBy, fmt.Sprintf("%d", fail.Since.Unix())},
			})
		}
	case c.Name == "ci-unlock" && len(c.Values) == 1 && hash.IsValidHash(c.Values[0]):
		// fossil drops CLIENT-ID's check-in locks for any requester; as with
		// ci-lock, this server lets only writers touch locks.
		if auth.CanPush(h.caps) {
			releaseCkinLocks(h.repo.DB(), c.Values[0])
		}
	case c.Name == "send-private":
		if auth.CanSyncPrivate(h.caps) {
			h.syncPrivate = true
		} else {
			h.resp = append(h.resp, &xfer.ErrorCard{Message: privateRefusal})
		}
	}
}

func (h *handler) handleDataCard(ctx context.Context, card xfer.Card) error {
	switch c := card.(type) {
	case *xfer.IGotCard:
		return h.handleIGot(c)
	case *xfer.GimmeCard:
		return h.handleGimme(c)
	case *xfer.FileCard:
		return h.handleFile(ctx, c.UUID, c.DeltaSrc, c.Content, nil)
	case *xfer.CFileCard:
		return h.handleFile(ctx, c.UUID, c.DeltaSrc, c.Content, c.StoredBlob)
	case *xfer.PrivateCard:
		if !auth.CanSyncPrivate(h.caps) {
			h.resp = append(h.resp, &xfer.ErrorCard{Message: privateRefusal})
			h.nextIsPrivate = false
		} else {
			h.nextIsPrivate = true
		}
		return nil
	case *xfer.ReqConfigCard:
		return h.handleReqConfig(c)
	case *xfer.ConfigCard:
		h.handleConfigPush()
		return nil
	case *xfer.UVIGotCard:
		return h.handleUVIGot(c)
	case *xfer.UVGimmeCard:
		return h.handleUVGimme(c)
	case *xfer.UVFileCard:
		return h.handleUVFile(c)
	case *xfer.XIGotCard:
		return h.handleXIGot(c)
	case *xfer.XGimmeCard:
		return h.handleXGimme(c)
	case *xfer.XRowCard:
		return h.handleXRow(c)
	case *xfer.XDeleteCard:
		return h.handleXDelete(c)
	}
	return nil
}

func (h *handler) handleIGot(c *xfer.IGotCard) error {
	if c == nil {
		panic("handler.handleIGot: c must not be nil")
	}
	// page_xfer reads igot only with an artifact name (blob_is_hname); any
	// other line falls through to its bad command (§5.2). IsValidHash takes
	// lowercase hex only where validate16 also takes A-F: an uppercase name
	// would break the blob table's CHECK constraint, so it is refused here.
	if !hash.IsValidHash(c.UUID) {
		h.badCommand(replyOnlyLine(c))
		return nil
	}
	if _, exists := blob.Exists(h.repo.DB(), c.UUID); exists {
		// Record that the client has this blob so emitIGots can skip it.
		// Mirrors Fossil's remote_has() → onremote table (xfer.c:1471),
		// which fossil fills only under push; it is kept for a pull too, so
		// a pull-only client is not told about what it just announced.
		h.recordRemoteHas(c)
		return nil
	}
	// Only a pushing client can deliver what it announces. Fossil's
	// page_xfer wraps igot handling in `if( isPush )`; asking a pull-only
	// client for artifacts it will never send holds its pull open until
	// MaxRounds (#277).
	if !h.pushOK {
		return nil
	}
	// A private igot makes a phantom only for a requester who may push
	// private content, as page_xfer checks g.perm.Private.
	if c.IsPrivate && !auth.CanSyncPrivate(h.caps) {
		return nil
	}
	// Fossil's igot handling goes through rid_from_uuid → content_new, which
	// returns 0 for a shunned name: no phantom is made, so no gimme follows
	// (src/content.c, #349).
	shunned, err := shun.IsShunned(h.repo.DB(), c.UUID)
	if err != nil {
		return fmt.Errorf("handler: shun check for igot %s: %w", c.UUID, err)
	}
	if shunned {
		return nil
	}
	h.pendingPhantoms = append(h.pendingPhantoms,
		pendingPhantom{uuid: c.UUID, private: c.IsPrivate})
	// page_xfer calls remote_has on the new phantom too, so a file card
	// later in the request that fills it is not announced back.
	h.recordRemoteHas(c)
	return nil
}

// recordRemoteHas notes that the client holds what an igot announced.
func (h *handler) recordRemoteHas(c *xfer.IGotCard) {
	if c == nil {
		panic("handler.recordRemoteHas: c must not be nil")
	}
	if h.remoteHas == nil {
		h.remoteHas = make(map[string]remoteHasEntry)
	}
	h.remoteHas[c.UUID] = remoteHasEntry{isPrivate: c.IsPrivate}
}

// pendingPhantom is an artifact a pushing client announced that
// flushPhantoms has yet to record.
type pendingPhantom struct {
	uuid    string
	private bool
}

// flushPhantoms records the pending igot names as phantoms in one
// transaction, as page_xfer's rid_from_uuid(..., 1, isPriv) does: push
// creates server phantoms from client igot (§3.3, §8.5). The gimme for them
// is sent by requestPhantoms after the card loop, so a file card later in
// the same request fills one without being asked for. One transaction for
// the run avoids a WAL commit per announced artifact.
func (h *handler) flushPhantoms() error {
	if len(h.pendingPhantoms) == 0 {
		return nil
	}
	pending := h.pendingPhantoms
	h.pendingPhantoms = nil
	return h.repo.WithTx(func(tx *db.Tx) error {
		for _, p := range pending {
			rid, err := blob.StorePhantom(tx, p.uuid)
			if err != nil {
				return fmt.Errorf("handler: phantom for igot %s: %w", p.uuid, err)
			}
			if !p.private {
				continue
			}
			if err := content.MakePrivate(tx, int64(rid)); err != nil {
				return err
			}
		}
		return nil
	})
}

// maxPhantomGimme bounds the gimme cards one reply sends for the server's
// phantoms: page_xfer's request_phantoms(&xfer, 500).
const maxPhantomGimme = 500

// requestPhantoms asks a pushing client for the server's phantoms, as
// page_xfer's request_phantoms does after the card loop: every phantom but
// those the client asked for this round (it lacks them), shunned names, and
// private ones outside a private sync (§3.3), at most maxPhantomGimme.
func (h *handler) requestPhantoms() error {
	if !h.pushOK {
		panic("handler.requestPhantoms: push not accepted")
	}
	query := "SELECT uuid FROM phantom CROSS JOIN blob USING(rid)" +
		" WHERE NOT EXISTS(SELECT 1 FROM shun WHERE uuid=blob.uuid)"
	if !h.syncPrivate {
		query += " AND NOT EXISTS(SELECT 1 FROM private WHERE rid=blob.rid)"
	}
	rows, err := h.repo.DB().Query(query + " ORDER BY rid")
	if err != nil {
		return fmt.Errorf("handler: listing phantoms: %w", err)
	}
	defer rows.Close()
	sent := 0
	for sent < maxPhantomGimme && rows.Next() {
		var uuid string
		if err := rows.Scan(&uuid); err != nil {
			return fmt.Errorf("handler: listing phantoms: %w", err)
		}
		if h.remoteLacks[uuid] {
			continue
		}
		h.resp = append(h.resp, &xfer.GimmeCard{UUID: uuid})
		sent++
	}
	return rows.Err()
}

func (h *handler) handleGimme(c *xfer.GimmeCard) error {
	if c == nil {
		panic("handler.handleGimme: c must not be nil")
	}
	// page_xfer reads gimme only with an artifact name (blob_is_hname); any
	// other line falls through to its bad command (§5.2).
	if !hash.IsValidHash(c.UUID) {
		h.badCommand(replyOnlyLine(c))
		return nil
	}
	// The client lacks what it asks for, so the server does not ask for it
	// back (fossil's remote_unk, ahead of the isPull check).
	if h.remoteLacks == nil {
		h.remoteLacks = make(map[string]bool)
	}
	h.remoteLacks[c.UUID] = true
	// Fossil's page_xfer answers gimme only `if( isPull )`: a pull card the
	// requester's Read permission accepted, or a clone, which sets isPull
	// too. Answering any other gimme hands content to whoever names a hash.
	if !h.pullOK && !h.cloneMode {
		return nil
	}
	// BUGGIFY: 5% chance skip sending a file to test client retry.
	if h.buggify != nil && h.buggify.Check("handler.handleGimme.skip", 0.05) {
		return nil
	}
	rid, ok := content.AvailableByUUID(h.repo.DB(), c.UUID)
	if !ok {
		return nil // blob not found — not fatal, skip.
	}
	isPriv := content.IsPrivate(h.repo.DB(), int64(rid))
	if isPriv && !h.syncPrivate {
		// §8.6: outside a private sync, a peer dated 2020-04-13 or later is
		// told the artifact exists and is private; an older or undated one
		// is told nothing, as fossil's send_file decides on remoteDate.
		if h.peerDate >= privateIGotDate {
			h.resp = append(h.resp, &xfer.IGotCard{UUID: c.UUID, IsPrivate: true})
		}
		return nil
	}
	// §7.6: SHA3 content is withheld from a client older than SHA3
	// support, or one that named no release; the refusal is appended, so
	// what the reply already holds stands (§10.1).
	if sha3Barred(c.UUID, h.peerVersion) {
		h.resp = append(h.resp, sha3Refusal())
		return nil
	}
	data, err := h.cache.Expand(h.repo.DB(), rid)
	if err != nil {
		h.resp = append(h.resp, &xfer.ErrorCard{
			Message: fmt.Sprintf("expand %s: %v", c.UUID, err),
		})
		return nil
	}
	if isPriv {
		// BUGGIFY: 5% chance skip the PrivateCard prefix — client should
		// treat the file as public; next sync round corrects the status.
		if h.buggify == nil || !h.buggify.Check("handler.handleGimme.dropPrivateCard", 0.05) {
			h.resp = append(h.resp, &xfer.PrivateCard{})
		}
	}
	h.resp = append(h.resp, &xfer.FileCard{UUID: c.UUID, Content: data})
	h.filesSent++
	return nil
}

func (h *handler) handleFile(ctx context.Context, uuid, deltaSrc string, payload []byte, storedBlob []byte) error {
	if uuid == "" {
		panic("handler.handleFile: uuid must not be empty")
	}
	// §6.6: without an accepted push, the first file or cfile card resets
	// the reply, is refused and ends card processing, as page_xfer's
	// "error not authorized to write" and break do. The end-of-request
	// steps of a pull accepted earlier still run.
	if !h.pushOK {
		h.breakLoop("not authorized to write")
		return nil
	}
	// A shunned artifact is dropped on receipt, with no error, as fossil's
	// xfer_accept_file drops it (§7.7); it is not received.
	shunned, err := shun.IsShunned(h.repo.DB(), uuid)
	if err != nil {
		return err
	}
	if shunned {
		h.nextIsPrivate = false
		return nil
	}
	// BUGGIFY: 3% chance reject a valid file to test client re-push.
	if h.buggify != nil && h.buggify.Check("handler.handleFile.reject", 0.03) {
		h.resp = append(h.resp, &xfer.ErrorCard{
			Message: fmt.Sprintf("buggify: rejected file %s", uuid),
		})
		return nil
	}
	vis := visibility(h.nextIsPrivate)
	h.nextIsPrivate = false
	if err := storeReceivedFile(ctx, h.repo, uuid, deltaSrc, payload, storedBlob, vis, nil); err != nil {
		h.resp = append(h.resp, &xfer.ErrorCard{
			Message: fmt.Sprintf("storing %s: %v", uuid, err),
		})
		return nil
	}
	_, ok := blob.Exists(h.repo.DB(), uuid)
	if h.buggify != nil && h.buggify.Check("handler.handleFile.missingAfterStore", 0.01) {
		ok = false
	}
	if !ok {
		h.resp = append(h.resp, &xfer.ErrorCard{
			Message: fmt.Sprintf("blob %s missing after store", uuid),
		})
		return nil
	}
	h.filesRecvd++
	return nil
}

// handleReqConfig answers reqconfig NAME with the configuration groups NAME
// selects and the requester may read (config_groups.go). Anything else — a
// bare name, an unknown group, a group the requester lacks the capability
// for — gets no cards and no error, as fossil answers it, so the rest of the
// exchange goes on.
func (h *handler) handleReqConfig(c *xfer.ReqConfigCard) error {
	if c == nil {
		panic("handler.handleReqConfig: c must not be nil")
	}
	mask := configRequestMask(c.Name, h.caps, h.canRead(), h.authed)
	if mask == 0 {
		return nil
	}
	cards, err := configGroupCards(h.repo.DB(), mask)
	if err != nil {
		return fmt.Errorf("handler: reqconfig %q: %w", c.Name, err)
	}
	h.resp = append(h.resp, cards...)
	return nil
}

func (h *handler) emitIGots() error {
	// Emit igot for all non-phantom blobs so the client can discover
	// everything the server has. Cluster generation is a client-side
	// optimization for push; the server always advertises all blobs.
	rows, err := h.repo.DB().Query(`
		SELECT uuid FROM blob WHERE size >= 0
		AND NOT EXISTS(SELECT 1 FROM shun WHERE uuid=blob.uuid)
		AND NOT EXISTS(SELECT 1 FROM private WHERE rid=blob.rid)`,
	)
	if err != nil {
		return fmt.Errorf("handler: listing blobs: %w", err)
	}
	defer rows.Close()

	var uuids []string
	for rows.Next() {
		var uuid string
		if err := rows.Scan(&uuid); err != nil {
			return err
		}
		// remoteHas is populated from client igot cards in handleIGot.
		// Skip if the client already has this blob as public (non-private).
		// If the client has it as private, we still emit the public igot so
		// the client can clear its private status (private→public transition).
		if e, ok := h.remoteHas[uuid]; ok && !e.isPrivate {
			continue
		}
		uuids = append(uuids, uuid)
	}
	if err := rows.Err(); err != nil {
		return err
	}

	// BUGGIFY: 10% chance truncate igot list to test multi-round convergence.
	if h.buggify != nil && h.buggify.Check("handler.emitIGots.truncate", 0.10) && len(uuids) > 1 {
		uuids = uuids[:len(uuids)/2]
	}

	for _, uuid := range uuids {
		h.resp = append(h.resp, &xfer.IGotCard{UUID: uuid})
	}

	if h.syncPrivate {
		if err := h.emitPrivateIGots(); err != nil {
			return err
		}
	}
	return nil
}

// emitPrivateIGots emits igot cards with IsPrivate=true for all blobs in
// the private table. Only called when the client sent pragma send-private
// and has the 'x' capability.
//
// Shunned private artifacts are still announced: fossil's send_private
// (src/xfer.c) selects from the private table with no shun filter, unlike
// the public igot paths. #349 asked for the filter, but matching fossil
// wins; the receiving side drops a shunned file on arrival either way.
func (h *handler) emitPrivateIGots() error {
	rows, err := h.repo.DB().Query(
		"SELECT b.uuid FROM private p JOIN blob b ON p.rid=b.rid WHERE b.size >= 0",
	)
	if err != nil {
		return fmt.Errorf("handler: listing private blobs: %w", err)
	}
	defer rows.Close()

	var uuids []string
	for rows.Next() {
		var uuid string
		if err := rows.Scan(&uuid); err != nil {
			return err
		}
		// Skip if the client already has this blob as private.
		// If the client has it as public, we still emit the private igot so
		// the client can update its private status (public→private transition).
		if e, ok := h.remoteHas[uuid]; ok && e.isPrivate {
			continue
		}
		uuids = append(uuids, uuid)
	}
	if err := rows.Err(); err != nil {
		return err
	}

	// BUGGIFY: 10% chance truncate private igot list to test multi-round convergence.
	if h.buggify != nil && h.buggify.Check("handler.emitPrivateIGots.truncate", 0.10) && len(uuids) > 1 {
		uuids = uuids[:len(uuids)/2]
	}

	for _, uuid := range uuids {
		h.resp = append(h.resp, &xfer.IGotCard{UUID: uuid, IsPrivate: true})
	}
	return nil
}

// sendAllClusters emits igot cards for all cluster artifacts that are
// not still in unclustered (i.e., already fully clustered themselves).
func (h *handler) sendAllClusters() error {
	rows, err := h.repo.DB().Query(`
		SELECT b.uuid FROM tagxref tx
		JOIN blob b ON tx.rid = b.rid
		WHERE tx.tagid = 7
		  AND NOT EXISTS (SELECT 1 FROM unclustered WHERE rid = b.rid)
		  AND NOT EXISTS (SELECT 1 FROM phantom WHERE rid = b.rid)
		  AND NOT EXISTS (SELECT 1 FROM shun WHERE uuid = b.uuid)
		  AND NOT EXISTS (SELECT 1 FROM private WHERE rid = b.rid)
		  AND b.size >= 0
	`)
	if err != nil {
		return fmt.Errorf("handler: listing clusters: %w", err)
	}
	defer rows.Close()

	bracketed := false
	for rows.Next() {
		var uuid string
		if err := rows.Scan(&uuid); err != nil {
			return err
		}
		// Clusters are always public (query excludes private table).
		if e, ok := h.remoteHas[uuid]; ok && !e.isPrivate {
			continue
		}
		if !bracketed {
			h.resp = append(h.resp, &xfer.CommentCard{Text: "sending-clusters"})
			bracketed = true
		}
		h.resp = append(h.resp, &xfer.IGotCard{UUID: uuid})
	}
	if err := rows.Err(); err != nil {
		return err
	}
	// fossil's send_all_clusters brackets the cluster igots with comments,
	// and sends neither when there is no cluster to announce (§5.3).
	if bracketed {
		h.resp = append(h.resp, &xfer.CommentCard{Text: "end-of-clusters"})
	}
	return nil
}

// handleConfigPush answers a pushed config card. fossil's page_xfer refuses
// one from a requester without Admin with a reset, this error and a break
// (§9.3). It applies an Admin's through configure_receive, which this server
// does not implement, so it refuses that too, naming why, rather than drop
// the configuration while the push appears to succeed.
func (h *handler) handleConfigPush() {
	if !auth.IsAdmin(h.caps) {
		h.breakLoop("not authorized to push configuration")
		return
	}
	h.breakLoop("configuration push is not supported by this server")
}

func (h *handler) emitCloneBatch() error {
	// Capture the snapshot bound on the first round of a clone session.
	// Without an upper bound the batch query keeps picking up blobs that other
	// writers commit between rounds, the completion sentinel
	// CloneSeqNoCard{SeqNo:0} never fires, and the client loops to MaxRounds
	// (issue #17). The bound rides back to the client via CookieCard and
	// gets echoed on every subsequent round.
	if h.cloneSnapMax == 0 {
		var maxRid int
		if err := h.repo.DB().QueryRow("SELECT COALESCE(MAX(rid), 0) FROM blob WHERE size >= 0").Scan(&maxRid); err != nil {
			return fmt.Errorf("handler: capture clone snapshot: %w", err)
		}
		h.cloneSnapMax = maxRid
		h.resp = append(h.resp, &xfer.CookieCard{Value: fmt.Sprintf("%s%d", cloneCapPrefix, maxRid)})
	}

	budget := DefaultCloneBatchBytes
	// BUGGIFY: 10% chance reduce the budget to one byte to stress pagination.
	// The budget is consulted before each artifact and never suppresses the
	// first one, so this yields exactly one artifact per round.
	if h.buggify != nil && h.buggify.Check("handler.emitCloneBatch.smallBatch", 0.10) {
		budget = 1
	}
	truncate := h.buggify != nil && h.buggify.Check("clone.emitCloneBatch.truncate", 0.10)

	// cloneSeq is the rid of the next blob to send and is inclusive, matching
	// §8.2's "iterate from supplied RID through current maximum" and
	// canonical's `while( seqno<=max )` loop in page_xfer(). handleControlCard
	// is the sole writer and floors it at 1, but that is enforced far from
	// here with nothing compile-time behind it, so this stays a check rather
	// than an assumption.
	cursorAtEntry := h.cloneSeq
	if cursorAtEntry <= 0 {
		return fmt.Errorf("handler: clone cursor must be positive, got %d", cursorAtEntry)
	}
	rows, err := h.repo.DB().Query(
		"SELECT rid, uuid, size FROM blob WHERE rid >= ? AND rid <= ? AND size >= 0 ORDER BY rid",
		h.cloneSeq, h.cloneSnapMax,
	)
	if err != nil {
		return fmt.Errorf("handler: clone batch: %w", err)
	}
	defer rows.Close()

	count := 0
	// sent records every rid emitted this round, including delta sources a
	// dependent's chain pulls in ahead of their ascending position (issue
	// #141). It caps re-emission at once per round the way canonical fossil's
	// per-request "sent" bag does; across rounds the ascending cursor resumes,
	// and a source pulled in early in a prior round may be re-sent, which the
	// receiver dedupes by hash.
	sent := map[int]bool{}
	// sentTop counts top-level artifacts sent this round. Each carries a whole
	// delta chain, so it is distinct from count (the card total) and is the
	// quantity the truncate liveness guard below reasons about.
	sentTop := 0
	// bytesSent accumulates the wire cost of the cards emitted this round,
	// mirroring canonical's `while( mxSend > blob_size(pOut) )`: the budget is
	// tested before each artifact and the artifact (with its chain) that
	// crosses it is still sent whole. That ordering is what guarantees forward
	// progress when a single artifact is larger than the entire budget — a
	// post-send test would emit nothing and stall the cursor.
	bytesSent := 0
	var lastTopRID int
	// nextRID is the cursor reported back to the client: the rid of the first
	// blob this batch did not send, or 0 once the snapshot is exhausted.
	nextRID := 0
	for rows.Next() {
		var rid int
		var uuid string
		var fullSize int
		if err := rows.Scan(&rid, &uuid, &fullSize); err != nil {
			return err
		}

		// A higher-rid delta source that a dependent's chain already pulled in
		// this round is on the wire; skip it before it can consume budget or
		// advance the cursor a second time.
		if sent[rid] {
			continue
		}

		isPriv := content.IsPrivate(h.repo.DB(), int64(rid))
		if isPriv && !h.syncPrivate {
			continue // skip private blob, don't count toward batch
		}

		if bytesSent >= budget {
			nextRID = rid
			break
		}

		// §7.2/#141: emit the artifact source-first. A row stored as a delta
		// goes out as a delta card preceded by the source it names, reclaiming
		// the delta-transmission bandwidth a plain (expanded) cfile gives up.
		// content.Deltify deltifies the OLDER artifact against the NEWER one,
		// so a delta's source almost always has a *greater* rid than the delta
		// itself and would not yet have been sent under this loop's ascending
		// order; buildCloneArtifact walks the chain and emits each source ahead
		// of its dependent so no delta ever forward-references a card that has
		// not arrived. The delta rides a cfile that names its source, as in
		// canonical's send_compressed_file.
		cards, rids, cost, err := h.buildCloneArtifact(rid, uuid, fullSize, sent)
		if err != nil {
			return err
		}
		// Respect the client's decode bound: the round the client decompresses
		// must stay within xfer.MaxDecompressedBytes. If this artifact's chain
		// would push the round past that bound and the round already carries
		// content, flush now so the chain leads a round of its own. Without this
		// the budget check above only bounds the accumulated *filler*, and a
		// large artifact riding behind sub-budget filler carried the round to
		// budget-plus-artifact, past what the client could decode (issue #109).
		//
		// The first artifact of a round (count == 0) is always sent, so a chain
		// larger than the bound still makes forward progress -- it is simply
		// unclonable, a property of the artifact alone and not of what precedes
		// it in the round.
		if count > 0 && bytesSent+cost > xfer.MaxDecompressedBytes {
			nextRID = rid
			break
		}

		h.resp = append(h.resp, cards...)
		for _, r := range rids {
			sent[r] = true
		}
		h.filesSent += len(rids)
		lastTopRID = rid
		count += len(rids)
		sentTop++
		bytesSent += cost
	}
	if err := rows.Err(); err != nil {
		return err
	}
	// BUGGIFY: truncate — remove the last cfile card to simulate an incomplete
	// batch. `sentTop > 1` is a liveness constraint, not just a chaos-injection
	// guard. The removed card is the deepest dependent of the last artifact's
	// chain (emitted last in source-first order, so nothing else on the wire
	// depends on it), and nextRID is set to that artifact's top-level rid. With
	// more than one top-level artifact sent this round that rid strictly
	// exceeds the cursor we entered with, so the client re-requests from it
	// rather than looping on the same batch until MaxRounds.
	if truncate && sentTop > 1 {
		for i := len(h.resp) - 1; i >= 0; i-- {
			if _, ok := h.resp[i].(*xfer.CFileCard); ok {
				h.resp = append(h.resp[:i], h.resp[i+1:]...)
				h.filesSent--
				// The dropped card held lastTopRID, so that rid becomes the
				// next one owed to the client rather than being skipped.
				nextRID = lastTopRID
				break
			}
		}
	}
	// The cursor must strictly advance or signal exhaustion. Catching a stall
	// here names the round that caused it; letting it through only surfaces
	// later as an opaque MaxRounds failure on the client.
	if nextRID != 0 && nextRID <= cursorAtEntry {
		return fmt.Errorf(
			"handler: clone cursor did not advance: next %d, cursor %d", nextRID, cursorAtEntry)
	}
	// SeqNo 0 signals completion so the client stops requesting.
	h.resp = append(h.resp, &xfer.CloneSeqNoCard{SeqNo: nextRID})
	return nil
}

// maxCloneChainWalk bounds the delta-chain walk in walkCloneChain. A real
// chain is only as deep as an artifact's revision count and always terminates
// at a full-content tip; this cap exists solely so a corrupt or adversarial
// delta graph cannot spin the walk forever. content.Deltify forbids cycles,
// but the send path does not trust its store to be uncorrupted here. On
// reaching the cap the deepest artifact is emitted as full content, which is
// always safe (it references no source card).
const maxCloneChainWalk = 4096

// cloneArtifact is one artifact the clone will emit. An empty srcUUID means
// emit full content as a compressed cfile card; a set srcUUID means emit a
// delta against that source, which buildCloneArtifact guarantees precedes this
// artifact on the wire.
type cloneArtifact struct {
	rid      int
	uuid     string
	usize    int
	priv     bool
	srcUUID  string
	expanded bool // stored as a delta but sent as full content
}

// cloneBlobRow holds the columns walkCloneChain needs to decide whether a
// delta source can be forwarded to the receiver this clone.
type cloneBlobRow struct {
	uuid string
	size int
	priv bool
}

// loadCloneBlob reads the blob row for rid. ok is false when no such row
// exists (a source named by a dangling delta link, which is treated as
// unforwardable rather than an error).
func (h *handler) loadCloneBlob(rid int) (cloneBlobRow, bool, error) {
	var row cloneBlobRow
	err := h.repo.DB().QueryRow("SELECT uuid, size FROM blob WHERE rid = ?", rid).
		Scan(&row.uuid, &row.size)
	if errors.Is(err, sql.ErrNoRows) {
		return cloneBlobRow{}, false, nil
	}
	if err != nil {
		return cloneBlobRow{}, false, fmt.Errorf("handler: clone blob info rid %d: %w", rid, err)
	}
	row.priv = content.IsPrivate(h.repo.DB(), int64(rid))
	return row, true, nil
}

// storedContent returns rid's content as the blob table stores it: the
// on-disk blob format, compressed, of the full artifact or of its delta.
func (h *handler) storedContent(rid int) ([]byte, error) {
	var stored []byte
	err := h.repo.DB().QueryRow("SELECT content FROM blob WHERE rid = ?", rid).Scan(&stored)
	if err != nil {
		return nil, fmt.Errorf("handler: stored content of rid %d: %w", rid, err)
	}
	return stored, nil
}

// walkCloneChain follows rid's delta chain toward its full-content tip and
// returns the artifacts to emit, ordered DEPENDENT-FIRST (rid itself first).
// The walk stops at the first source that is already available to the receiver
// (emitted earlier this round) or that cannot be forwarded this clone —
// outside the snapshot bound, a phantom, or a private source under a public
// clone — and tags the artifact below such a source for full-content emission
// so it never depends on a card that will not arrive.
func (h *handler) walkCloneChain(rid int, uuid string, usize int, sent map[int]bool) ([]cloneArtifact, error) {
	chain := make([]cloneArtifact, 0, 8)
	curRID, curUUID, curUSize := rid, uuid, usize
	curPriv := content.IsPrivate(h.repo.DB(), int64(curRID))
	for step := 0; step < maxCloneChainWalk; step++ {
		src, err := content.DeltaSource(h.repo.DB(), libfossil.FslID(curRID))
		if err != nil {
			return nil, fmt.Errorf("handler: delta source for rid %d: %w", curRID, err)
		}
		m := cloneArtifact{rid: curRID, uuid: curUUID, usize: curUSize, priv: curPriv}
		if src == 0 {
			return append(chain, m), nil // full-content tip
		}
		info, ok, err := h.loadCloneBlob(int(src))
		if err != nil {
			return nil, err
		}
		forwardable := ok && info.size >= 0 && int(src) <= h.cloneSnapMax &&
			(!info.priv || h.syncPrivate)
		if !forwardable {
			m.expanded = true
			return append(chain, m), nil // source unsendable: emit curRID full
		}
		m.srcUUID = info.uuid
		chain = append(chain, m)
		if sent[int(src)] {
			return chain, nil // source already on the wire this round
		}
		curRID, curUUID, curUSize, curPriv = int(src), info.uuid, info.size, info.priv
	}
	// Walk cap hit: force the deepest artifact to full content so its delta
	// does not reference a source we never emitted.
	chain[len(chain)-1].srcUUID = ""
	chain[len(chain)-1].expanded = true
	return chain, nil
}

// buildCloneArtifact assembles the wire cards that deliver artifact rid to the
// receiver, ordered source-first: the chain tip (or the lowest full-content
// anchor) leads, each delta follows the source it names, and rid itself is
// last. It returns the cards, the artifact rids they cover (for the caller's
// sent bookkeeping), and the accumulated wire cost. It does not consult the
// byte budget — the caller decides whether the whole sequence fits the round.
func (h *handler) buildCloneArtifact(
	rid int, uuid string, usize int, sent map[int]bool,
) ([]xfer.Card, []int, int, error) {
	chain, err := h.walkCloneChain(rid, uuid, usize, sent)
	if err != nil {
		return nil, nil, 0, err
	}
	if len(chain) == 0 {
		panic("handler.buildCloneArtifact: walkCloneChain returned no artifacts")
	}
	cards := make([]xfer.Card, 0, 2*len(chain))
	rids := make([]int, 0, len(chain))
	cost := 0
	// chain is dependent-first; emit it reversed so every source precedes the
	// delta that names it.
	for i := len(chain) - 1; i >= 0; i-- {
		m := chain[i]
		rids = append(rids, m.rid)
		// §7.6, Appendix A.2: the refusal takes a SHA3 artifact's place.
		if sha3Barred(m.uuid, h.peerVersion) {
			cards = append(cards, sha3Refusal())
			continue
		}
		var data []byte
		if m.srcUUID == "" {
			data, err = content.Expand(h.repo.DB(), libfossil.FslID(m.rid))
			if err != nil {
				return nil, nil, 0, fmt.Errorf("handler: expanding rid %d: %w", m.rid, err)
			}
		} else {
			data, err = blob.Load(h.repo.DB(), libfossil.FslID(m.rid))
			if err != nil {
				return nil, nil, 0, fmt.Errorf("handler: loading delta rid %d: %w", m.rid, err)
			}
		}
		if m.priv {
			cards = append(cards, &xfer.PrivateCard{})
		}
		// Every clone artifact rides a cfile, a delta naming its source, as
		// canonical's send_compressed_file does for clone version 3 and as the
		// draft requires (§7.1, §8.2). USIZE is the full artifact's size even
		// for a delta, as fossil sends blob.size.
		// An artifact sent as it is stored goes out as its stored bytes,
		// as send_compressed_file sends the blob table's content (§7.2).
		var stored []byte
		if !m.expanded {
			if stored, err = h.storedContent(m.rid); err != nil {
				return nil, nil, 0, err
			}
		}
		cards = append(cards, &xfer.CFileCard{
			UUID: m.uuid, DeltaSrc: m.srcUUID, USize: m.usize, Content: data,
			StoredBlob: stored,
		})
		cost += cloneCardOverheadBytes + len(data)
	}
	return cards, rids, cost, nil
}