The engine keeps Java's packet capture (capture.ts): per network card
every frame it sent (when it leaves on the cable) or received (also one
for another card), once, with time and number; switches log nothing;
frames the packet loss dropped are kept apart; the capture outlives a
stop and a new start, as in Java, until clearCapture. Per port instead
of per MAC, at most 10,000 frames a card.
expectSameCaptures compares each card's capture with Java's list of that
card in the ARP, WLAN, DNS, TCP, NAT, RIP and single-client DHCP runs.
It found that ScenarioNetwork ignored a switch's learning time (Java ran
300 s, the core the scenarios' 5 s); the oracle sets it now, and the TCP
and DHCP runs are recorded again. A core test found that a new start
numbers frames from 1 again and the capture took them as seen.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
WlanRun pings on the new network wlan-lan: over wireless cables, over a
plain cable, and to and from a notebook on the switch's SSID without a
cable. The engine matches Java without a change: it runs the links only.
Swift's wirelessAssociations, which join a host to its SSID's switch at
run time, are not taken. ScenarioNetwork sets wireless cables, wireless
cards and SSIDs; a corpus test pings over the wireless cables of
multicast_server.fls.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The core translates before the IP thread, as Java's NAT threads sit
between card and Ethernet: out, the WAN address and the entry's port (ICMP
keeps its identifier); in, by remote end, port and protocol, else by the
port forwarding, which then makes a dynamic entry; a ping at the WAN
address is answered by the router, packets from outside for the LAN are
dropped, both sides pass the router's firewall, TTL 1 going out gets time
exceeded. Dynamic entries expire at Java's check each second, 300 s times
the delay factor old (55 minutes at 90 %), refreshed only by what comes in.
The home router forwards as a router does.
Measured and not taken: when a LAN endpoint sends to a second remote end,
Java gives it the entry's WAN port but no entry, and the answer stays at
the router; the core adds an entry for each new remote end on that port.
The oracle now records every card of a node (the second card of routers
was missing); the ARP and RIP runs are recorded anew and still match.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ScenarioNetwork builds home routers (their first card the WAN port) and
their port forwarding (portForwarding, through GatewayFirmware.setStaticNAT).
NatRun plays six stories on nat-home's network and writes the frames, the
calls and the router's NAT table: a ping, a UDP exchange, the same socket
to a second server (whose answer never reaches the LAN: Java keeps one
entry per LAN endpoint, for the first remote only), a TCP connection out,
a forwarded one in, and pings at the WAN address and past it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
RipRun runs rip-chain's three routers in Java FILIUS, each router's random
set by reflection to the core's ripSeed, and writes the frames, every
change of each RIP table and the calls. converge: the tables fill, a ping
at 45 s. expire: R3 falls silent at 10 s; R2 marks its network 16 hops only
at its own next beacon, R1 passes it on at once, the ping gets ICMP 3/0.
Both match: per sending card the same messages at the same times, every
table through the same states. The first 5 s are compared by the state the
tables reach, since Java's routers start one after the other.
Measured on the JDK: Java's pause divides nextFloat() by 3 in float; the
core does too now (a millisecond off in one draw of about 10,000 before).
Scenarios name their RIP routers in the new field rip.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The resolver runs on the events now (a query, its dnsQueryTimeout, its
answer at the socket); the blocking resolveDns waits for it, and a
recursive server uses it: its own records, else the zone's name server
from its records, else its own DNS server, else a referral
(DNSServer.answerWithRemoteData).
DnsRun plays five stories on the new dns-recursive network. Measured and
not taken: a server that is its zone's name server asks itself, which asks
itself again, without end (the client times out, then every level floods
its own DNS server); Java guards only the step through its own DNS server.
The core asks itself in neither step. That story is compared up to the
client's query (JAVA_UNTIL).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Every packet that comes in goes through its node's IP thread (UDP, TCP) or
ICMP thread (ICMP), as Java's IPThread and ICMPThread: deliver, answer an
echo request or forward, one after the other; while one waits for ARP, the
next wait behind it. A packet a node sends to itself is delivered at once,
as Java's IP.senden does. The transport send threads and these are one
mechanism now (handToProtocolThread).
ArpRun has three stories through the router (routedUdpBehind,
routedPingBesideUdp, routedPingBehindPing); the oracle writes ICMP frames
with type, code, addresses and sequence number. All seven ArpRun stories
match Java FILIUS.
Measured and not taken: Java's ARP wakes one waiter with notify, not the
one whose answer came; the core wakes all that wait for the address.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
TcpRun loses exactly one frame (loseFirst: the card's output buffer drops
it; the sender's packet log still shows it, as with setDrop) in three new
stories: lostSynAck, lostFin, lostLastFin.
Measured: Java never repeats a SYN+ACK or a FIN. The server gives a
handshake up after one timeout (3 RTT) and closes with a FIN from SYN_RCVD
(its boundary check sometimes lets it wait twice; it means once). After its
FIN a socket waits at most five times, each wait ended by a segment or the
timeout; then it is gone. The core does the same (TCP_CLOSING_WAITS, with
TIME_WAIT in the loop) and shows CLOSED where Java leaves the state
standing (a named Java bug).
One departure: a repeated SYN in SYN_RCVD gets the SYN+ACK once more and
the wait starts anew; Java closes there, but its client never repeats the
SYN. lostSynAck is compared up to the second SYN (JAVA_UNTIL); the rest is
an engine test.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Sends to addresses no card has on the new arp-routed network (ping on the
LAN and behind a router, two UDP datagrams, one behind a failing one):
frames and calls. ScenarioNetwork builds routers and their interfaces.
Measured: ARP asks twice, a round trip each, then host unreachable; the
thread that waits is the ping's caller, the router's IP thread, or the
node's transport thread per protocol, which blocks the next send.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A resolver's lookup on the new dns-lan network (A, MX, a referral by NS,
an unknown name, a missing server): every frame and the call's result.
ScenarioNetwork installs a scenario's DNS servers. The Swift comparison
reads only scenarios with a Swift result.
Measured: Java's text messages (ID=... QR=... header, question and record
lines), no question in the answer, RCODE 0 always, a referral in the answer
section, an iterative resolver waiting 5 RTT per query.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Three stories of socket calls on tcp-lan's network (exchange, closedPort,
lostData): every frame with TCP flags and numbers, when each call started,
returned or failed, and each socket's states. Network building and the
packet log move from DhcpRun to ScenarioNetwork.
Measured: 3 RTT per attempt (8250 ms at 90 %), an orderly close after a
failed send, TIME_WAIT of 3 RTT, the passive side closing by itself on FIN,
one SYN only (Java's retry loop ends after the first timeout).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Builds a core scenario's network in Java FILIUS, starts it as the action
mode does and writes every frame each card saw (the Lauscher) and the
addresses at the end. References for the four DHCP scenarios: one
DISCOVER/REQUEST per attempt, an ARP check that waits one RTT, DECLINE on a
conflict, link-local 169.254.1.x after the fifth failure.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
engine.ts ports Swift's "DHCP server and automatic client": an active
server listens on port 67 from the start and offers a static assignment
before its range, round from the address after the last one offered;
offers last 10 s, dynamic leases a day, and a server whose offer another
server beat blacklists that address for a while. A client starts as an
event: DISCOVER, OFFER, a check by ARP, DECLINE, REQUEST, ACK or NAK,
2.5 s per wait, and after ten failures its old address back. The snapshot
carries the DHCP settings of clients and servers.
Four scenarios match the Swift engine. One Swift error is named in the
oracle test: a client run directly is advanced a second time, to FINISH,
by the server's answer within its own DISCOVER, and the outer run then
writes back its stale copy of the status (Swift's value types); the
address and every trace agree, only the reported status is wrong in
Swift. Swift's three DHCP engine tests are ported, with tests for
stopping and the message format. 15 mutation probes, 14 caught; the other
concerns the home router, which runs only once NAT is ported.
Two departures from Java remain, to be fixed next: from the event queue
the client's ARP check does not wait, so an address conflict is never
found, and each answer that wakes the client makes it send its state's
message again (a second DISCOVER and REQUEST, the second answered with
NAK). Java waits a round trip and sends once per attempt.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
engine.ts ports Swift's "Java-compatible TCP sockets" in one piece, since
half a state machine could not be checked: server and client sockets (a
free port from Java's random numbers on their own seed), the three-way
handshake with its result (connected, invalid socket, unreachable, timed
out), segments of 1460 bytes, messages up to the push, a timeout of
3 × 2.5 s with at most three attempts, closing through every Java state
including TIME_WAIT (2.5 s), sequence numbers modulo 2^32.
Three scenarios match the Swift engine trace for trace: a LAN with data
over three segments, closing from both ends, a closed port and an
unreachable host; two networks through a router with a firewall rule,
data lost until the connection closes, and an accepted socket that does
not hold the port of its closed listener; CLOSING, reached by a lost FIN.
The engine delivers a FIN within the same close, so both ends close at
once only with packet loss; an empty message does not arrive in Swift (it
does not advance the sequence number), nor here.
Swift's four TCP engine tests are ported, with tests for stopping, a
foreign SYN+ACK at a listener and the flag names. 17 mutation probes, 16
caught; the other (delivered, not waited, not connected gives unreachable)
cannot be reached through Swift's engine, which never sends a reset and
always acknowledges correctly.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
engine.ts ports the UDP part of Swift's "UDP sockets and Router RIP": a
socket bound to a given port or to a free one drawn from 49152–65534 with
Java's random numbers on Swift's seed; sending, also as a limited broadcast
on every interface (or the bound one) with TTL 1; delivery into the queue
of every socket on the destination port, which then answers to the last
sender; receiving, with a timeout that advances the virtual clock; closing.
Stopping forgets sockets and queues. RIP stays not ported; Swift's DHCP
handling of a datagram has nothing to do until DHCP is in the snapshot.
The oracle gets the steps bindUdp, sendUdp, receiveUdp and closeSocket on
both sides and reports the sockets. Two new scenarios — a LAN, and two
networks through a router with a UDP firewall rule and a broadcast that
does not cross it — match the Swift engine trace for trace, the random
ports included. Swift's four UDP engine tests are ported, with a test for
stopping. 11 mutation probes were all caught (one only after the stop test
was added).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
tools/swift-oracle builds the unchanged Swift runtime (pinned 3b5ded0)
in Docker and replays JSON scenarios; the TypeScript engine must produce
the same trace events, results and end state. Five scenarios, 884
events, identical.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
