Two-player teams, enforced by the simulation:
- Match puts every player on a team (SimEntity.team); Weapons.damage_player
refuses damage between teammates from every weapon, the boathook and blasts.
- A player whose health runs out with a teammate standing goes down instead of
out: 60 health to bleed (one point every half second), crawling at 1 m/s, no
firing, building, looting or items. Hits still take it; drowning too.
- A standing teammate within 2 m holding interact for 6 s revives them with 30
health. Letting go or moving away starts it over.
- Downed players are out when their health runs out again or when no teammate
is left standing, credited to whoever downed them. A team falls together and
shares its place; the tie-break puts standing ahead of downed.
- Downs and revives go through Match events and the kill feed ("X knocked Y
down"); the HUD says what to do while down.
Wire: FLAG_DOWNED in the player record and the own block (flags now u16),
so prediction crawls too. Protocol version 5.
Bots: never target a teammate, land with their teammate, regroup when they
stray past 30 m, crawl toward a standing teammate when down, go to revive a
downed teammate within 60 m when not in a fight, and turn from a downed target
to its standing teammate when it's in view.
Options: --team-size=2 on the server, --duos offline, and Solo/Duos under
Crew in the Practice menu. The figure has a downed crawl pose.
Tests: test_duos (downing, crawl, revive, interrupted revive, bleed-out,
finishing, team out, falling together, no friendly damage including kegs,
drowning while down, hash, match placement and tie-break), bots as teammates,
the downed flag on the wire, the feed, server and menu options. make
duo-matches plays 10 bot duos matches.
Closes#30
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
make stability (tools/stability.sh, StabilityReport) runs one dedicated
server for 50 bot matches back to back, unthrottled, sampling its resident
memory, and fails on a crash, hang, engine or script error, a match with
no winner, or memory that grows after the first five matches.
The first run had no crash, hang or error, but showed two leaks:
- ServerLoop kept every tick's duration in an array (8 bytes a tick,
about 21 MB a day at 30 Hz) and copied and sorted all of it for every
progress line. TickTimes keeps exact statistics in fixed memory: a
one-microsecond histogram up to 65 ms, plus the rare slower ticks.
- Resident memory grew about 28 MB a match (905 MB after 42 matches):
glibc gives each thread its own malloc arena, and every next island is
built on a worker thread. The server launchers (packaging/server,
packaging/dist, so also the .deb, Docker image and systemd unit) and
tools/godot.sh for --server now set MALLOC_ARENA_MAX=1 unless already
set. Tick times were unchanged (p50 501 vs 507 us, p99 1635 vs 1625).
After: 50 matches (seed 4242) in 756 s, all with a winner, exit 0, no
errors; static memory +1.6 KiB and resident +145 KiB a match, flat at
204 MB from match 26 on. No match crashed, so there is no crash seed to
turn into a regression test.
Tests: TickTimes against a sorted list, its fixed memory, the launchers'
setting, and StabilityReport. make check, dist-check and
server-package-check pass.
Closes#57
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
make balance (tools/balance.sh, tools/balance.gd, BalanceReport) plays 200
bot-only matches in four processes under nice and reports match length,
players standing over time, causes of death, eliminations by weapon and by
difficulty, judged against the new BALANCE_* targets in Tuning and the
GDD's Balance section.
Before: median match 5.0 min, 3.4 of 24 standing at 2 min, the boathook
41% of eliminations. After, over seeds 1001-1200: median 13.1 min (p10 9.2,
p90 14.6), top weapon the rivet repeater at 31.7%. Passes.
Changes, each traced from single-match runs (docs/BALANCE.md):
- Fix: bots leapt from the barge on a timer and landed hundreds of metres
short of their spot; they now leap when it's in gliding reach or the
barge starts drawing away from it.
- Bots spread their landings, only brawl whoever hit them, pick out new
opponents within 30 m, roam locally and hold ground, run from a gun when
unarmed, and (normal and hard) break off when badly hurt.
- Aim error widened; rivet repeater damage 17 to 15.
- Tide: lower early levels, longer holds; every match ends within 17.75 min.
Tests: BalanceReport, the new bot behaviours, the tide length. Tests that
leaned on old bot luck now pin or set up what they check. make match moves
to seed 29, where the scripted player still builds.
Closes#56
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
docs/SECURITY_REVIEW.md lists the scope, nine findings and what was done.
- High, fixed: a peer that connects and stops answering kept its ENet slot
for up to 15 s after being turned away, so one address could fill every
slot. Flood refusals now disconnect at once (reason in the disconnect
data, which NetClient now honours for REASON_BUSY), peers told to leave
get 1 s to acknowledge, and when slots run short the longest-waiting
silent guest makes room.
- Medium, mitigated: ENet holds half-open slots for spoofed CONNECTs that
the script never sees. The host now has 512 slots (NetServer.peer_slots).
- Medium, fixed: SourceLimiter, handshakes per source and the list
server's per-address cap count an IPv6 /64 as one source.
- Low, fixed: CommandLog refuses a header with non-numeric fields instead
of raising errors.
- tools/godot.sh unsets DISPLAY and WAYLAND_DISPLAY so the engine's error
alerts can never reach the host desktop.
- A test fixture's 10.0.0.9 is now the documentation address 192.0.2.9.
Tests: 4 new handshake tests with real non-acknowledging ENet peers,
test_source_limiter (3), a CommandLog type test. make check, make fuzz,
make client-check, make discovery-check and make soak-selfcheck pass.
Closes#58
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Harvest.swing traces the swing as a shot is traced (Weapons.trace, lag
compensated). A player reached before terrain, a solid, a prop's grown
box or a built piece takes the boathook's damage through
Weapons.damage_player, so armor, the kill feed and stats count it. A level
swing meets the head and takes the headshot multiplier, as a shot does.
- Unarmed bots go for an opponent in view within BOT_BRAWL_METRES (6 m) and
swing within BOT_SWING_METRES; inside a hut they swing back without
walking in.
- Tests: harvest +5 (hit, armor first, out of reach, wall between, player in
front of a tree), bot brain +2 (closes in and lands hits; leaves a distant
opponent alone).
- test_world_replication: an unarmed bot now meets the scripted player early
on seed 2, so the test gives it full armor on landing so it lives to build;
all assertions unchanged.
Closes#73
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Build ghost: ClientSession shows a BuildGhost in build mode while the
client plays, asking Building.check() of its own world and predicted
player with the mouse's newest look, so it agrees with the server.
- Flooded basins: ClientMirror.follow_flood() spreads the client's flood
as the server's tide tick does, from the tide and the pieces it has.
WaterView passes a mask of flooded cells (Flood.pours says when it
changed) to the sea shader, which leaves dry cells undrawn.
- Shots, impacts and swings: a new FX message (NetFx) goes unreliably on
the state channel next to each snapshot, with what happened within
NET_FX_RADIUS_METRES of what the client watches or involving that
player. ClientGame keeps them in `fx` for views and sound. They change
no state, so a lost one is a missed flash; sending them reliably would
add to the traffic a bad link can least afford (N8). The fuzzer covers
the new parser.
- Captures: client_build and client_build_refused show the ghost in its
allowed and refused tints; client_basin rings shore ground with
breakwaters and plays until the tide stands over it, and the server,
the client's flood and the drawn sea all keep the basin dry.
- Tests: NetFx round trips, limits and malformed input; a landed player's
shots and impacts reach a client, each matching a server event; the
client keeps a ringed basin dry through the first rise, floods it with
no ring, catches up after a long frame; the water mask follows pours.
The held item is still not drawn; that needs the models (A4, #44).
Closes#78
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The soak's drops under loss came from stalls: the server's island (#79) and
each client process making islands on its main thread fed ENet second-long
round-trip samples, its retransmission timer grew, and two losses of one
reliable message in the busy drop that followed reached the 15 s timeout.
- ClientGame.build_in_background makes a new match's island on a
WorkerThreadPool task and holds the packets that arrive meanwhile, then
takes them in order. The networked client and the soak's clients use it;
the soak's own one-island-a-frame holding is gone.
- NetServer.close(grace_msec) sends the shutdown notice reliably and
services the socket while clients acknowledge it. The headless server
waits up to 2 s. A notice sent once was lost under 2% loss, and that
client timed out instead of hearing the server stop.
- Tests: background island holding and ordering, freeing mid-build, a
graceful shutdown reaching four clients behind 50% loss (it fails with
the old one-shot notice), and the scripted UDP match and client capture
wait for the client's island the way a real 20 s lobby would.
The soak at 2% loss and +/-20 ms now keeps all 24 clients for both matches,
with 0 of 3,758 world hashes differing and every client told of the stop.
It still fails on client memory growth between matches in two of three
processes (17 and 20 MB), which is not yet explained, so #39 stays open.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
When a match's results begin, Match starts the next match's Simulation on a
WorkerThreadPool task (its seed is fixed by then) and collects it when the
lobby opens. The same task plans the island's tide levels
(Simulation.plan_tide) and runs a prepare_next hook, which ServerLoop uses
to build the bots' NavGrid and BotWorld for the island. All of these read
only the new simulation, so the results are the ones the main thread would
have made: tests compare the state hash, terrain, solids, props, tide
levels and navigation with fresh ones, and the ten-matches test still
matches its references. A Match freed while its task runs waits for it.
Measured: a 24-bot server paced in real time through the results, lobby,
countdown and 3 s of the next drop has a worst tick of 5.4 ms; it was 1.8 s.
In the 24-client soak at 2% loss, moving the island alone left a 0.82 s tick
(the bots' navigation) and moving that too left 74 ms at the next match's
start (planning the tide), which this also moves. The soak hasn't been run
again since that last step.
Closes#79
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- packaging/server/tidewall-server: a wrapper that turns SIGTERM and SIGINT
into a clean stop through the server's new --stop-file, since the engine
can't catch signals. The server tells everyone it shut down, logs its
statistics and exits 0; the wrapper kills it only after 15 s.
- --config=PATH reads options from a file (packaging/server/server.cfg);
options after it on the command line win.
- --health-port answers GET /health with JSON (phase, players, tick, uptime)
on loopback by default. HealthServer is non-blocking and bounded.
- packaging/docker: a debian:13 image that checks the engine against
tools/engine.sha512, imports the game at build time, runs as an unprivileged
user, and works with a read-only root and a small /tmp. compose.yaml is
project and container tidewall-server, read-only, no capabilities.
- packaging/systemd: a hardened unit with a dynamic user.
- make server-package-check: compose up, healthy, non-root, read-only, a real
headless client joins over UDP into the drop, then SIGTERM stops it cleanly.
This also covers H7 (graceful shutdown on SIGTERM).
Closes#38Closes#70
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
make soak runs a real server and 24 headless clients in three processes for
two whole matches in real time. Each client reaches the server through its own
UdpRelay, which adds 60 ms, +/-20 ms jitter and 2% loss each way, so ENet runs
unchanged over a bad link. SoakPlayer drives each client from what a client
knows: it drops, glides, heads for high ground, harvests and tries to build.
tools/soak_report.gd checks world hashes against the server's, errors,
unreadable packets, every client playing every match with its connection
intact, memory between matches, and the p99 tick with the network's share.
make soak-selfcheck proves a drifted client world is caught.
Server changes: --matches=N, --world-hash-every=N, tick timing that includes
after_tick, and memory in match and progress lines.
What the soak found and fixed:
- Snapshot building for 24 clients took the p99 tick to 13.8 ms, over the
12 ms budget, against a 2.8 ms simulation. Encoded records are now shared
between clients for the tick, unchanged records skip the field diff, and
clients are spread over the snapshot interval by handle. p99 is 7.3 ms.
- ENet peers had a timeout limit of 0, which gives up on a peer as soon as any
reliable message has waited the minimum timeout. Now ENet's default, 32.
Not fixed: one or two clients still lose their connection about 15 s into the
second match's drop in each run (see the issue), so the soak still fails and
N8 stays open.
Refs #39
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- A hostile test client crafts INPUT packets byte by byte and sends them
through NetGame.take_input() into a real match: extreme axes, the same
tick claimed repeatedly, the whole lead window at once, past, far-future
and maximum ticks, fire toggled every tick with slot swaps, builds aimed
anywhere with junk kinds and materials and no materials, and
pickup/drop/select tricks. Speed stays within sprint speed per tick,
shots within the weapon's rate, pieces within build reach and paid for,
and every item is conserved.
- Fix: a command tick beyond the simulation's lead window no longer moves
NetGame's dedupe marker, so one bogus far-future tick can't lock out a
player's real commands.
- Connections are rate-limited per address (32 per 10 s, 8 handshakes in
progress) with a new REASON_BUSY and a message for it.
- ParserFuzz mutates real messages from a bot match and random bytes and
feeds them to every network decoder, the snapshot receiver and the list
server. 20,000 per make check; make fuzz runs 1,000,000 (about 40 s):
no errors, 2.9 KB of memory growth.
- ARCHITECTURE.md: "Trusting nothing from clients".
Closes#40
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- LAN discovery: a listening server answers QUERY packets on its game port
plus one (ServerBeacon, off with --no-lan). ServerBrowser scans eight game
ports from 24650 by broadcast.
- A self-hostable list server: --list-server [--port=N] (ListServer,
default UDP 24660). Game servers register with --list=HOST[:PORT] and
--name=NAME; they are listed under their source address only after
echoing a CHALLENGE token. Entries expire after 45 s, 8 per address,
256 in all; every source is rate-limited (SourceLimiter).
- NetDiscovery messages: every QUERY is padded to 512 bytes and no answer
is bigger, so neither service amplifies spoofed requests.
- tidewall --browse [--list=...] [--lan-target=IP] [--lan-port=N] prints
the servers found as JSON until menus (A10) show them.
- People who connect mid-match or during results watch a living player
(Match.watch) and take places ahead of bots in the next lobby, instead
of being told the server is full. Bots fill only the places left when
the match starts.
- Tests: codec round trips, size limits and a 40,000-message fuzz; list
server with fake servers (challenge, stolen and wrong tokens, renewal,
expiry, caps, flood); LAN scan over loopback; watchers over UDP; lobby
backfill. make discovery-check runs it all as real processes.
Closes#37
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Built pieces, props' health, opened chests and the tide's schedule now reach
clients as WORLD messages on the reliable channel, and each player is told
what they carry as OWN messages.
- NetWorld: the WORLD and OWN wire formats, and world_hash(), a digest of
everything replicated this way.
- WorldSender: one message after each tick in which something changed, the
same for everyone, and the whole state for someone joining or a new match.
StructureSet and PropSet keep the change lists it reads.
- ClientWorld: applies them to the client's world, so pieces and broken props
are solid for prediction. A piece's health follows from its building and
damage, so building costs nothing on the wire.
- Placing ahead of the server: a provisional piece goes up with the command
and is replaced by the server's, or taken back once a snapshot shows the
server had its chance. Snapshots carry the newest WORLD number for that.
- Protocol version 3.
A test plays a whole scripted match on a real island with a ClientGame
attached and compares the two worlds' hashes throughout.
Closes#36
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The client's side of a networked match, in src/net/:
- ClientGame takes the server's notices and snapshots, generates the island
from the seed and answers with commands. It works on packets, so it runs
over a NetClient or a test's simulated link.
- ClientClock decides which tick the next command is for, from the server's
echo of the newest command tick it has received, holding three commands
waiting at the server.
- ClientPrediction runs each command on a copy of the client's own player
with the server's movement code, and compares every snapshot with what it
predicted. A difference is replayed from the server's state and eased in.
- SnapshotInterpolator draws everyone else four ticks in the past, between
snapshots, on a clock steered gently by arrivals.
Snapshots now carry the client's own block: exact position and velocity,
the water around the player, stamina and state flags, so a prediction can
match the server to the bit. The command echo is now the newest tick
received rather than taken. Protocol version 2.
Snapshots go out every second tick. The interval was three ticks, which is
10 a second, not the 20 the documents claimed; 15 a second is the nearest
even spacing and what interpolation is tuned for.
Drop.ride() holds the barge ride for the server and the client alike, and
NetGame's joining, input and snapshot steps can be called without a socket.
Measured in test/net/test_prediction_link.gd over a simulated link with
100 ms each way, 20 ms jitter and 2% loss: no corrections to own movement
in about 1,400 snapshots per client, and the other player drawn 0.4 mm on
average from where a perfect link would draw them.
Closes#34
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
NetSnapshot defines the game traffic after the handshake: delta snapshots of
quantized entity records, the client's input packet and a reliable match
notice. SnapshotSender builds each client's snapshot against the newest one
it acknowledged, tells it only about entities near the player it is watching,
and keeps every snapshot under 1,200 bytes, nearest first, with a staleness
bonus so nothing starves. SnapshotReceiver rebuilds the picture, ignores old
and repeated snapshots and asks for a fresh start if it is ever sent a delta
on something it doesn't hold.
NetGame joins the transport to the match: people who connect become
participants, their inputs go through ServerLoop.submit() once per tick, and
snapshots go out 20 times a second. `--server --listen` makes the headless
server take network players and wait in the lobby for the first of them.
Measured by test on a 24-player match start with every player a client, a
tenth of packets lost and late acknowledgements: 2.4 KB/s mean and 2.5 KB/s
for the worst client, packet headers included, against a 40 KB/s limit.
Closes#33
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
`make match` plays one whole match from the lobby to the results and back to the
lobby: one scripted player and 23 bots on a local server, with no script errors.
- CommandLog records the seed, the rules and every command the server is given.
MatchReplay plays a log again with nobody thinking and ends on the same state
hash; `make match` saves the log, reads it back and checks that.
- ServerLoop.submit() is now the one way commands reach the simulation, so a
recording server sees them all.
- ScriptedPlayer plays a human seat: it drops, loots, fights and flees as a bot
does, and also cuts driftwood and builds a breakwater and a wall.
- Bots draw randomness from their own per-bot generators, not the simulation's
streams. A bot thinking no longer changes the simulation's state, which a
replay without bots needs.
- MatchView and EntityView draw a match from one Simulation: terrain, sea,
points of interest, props, pieces, players, the barge, items and chests, with
stand-in shapes for the art pass to replace.
- `make match-captures` renders eight stages of the recorded match.
Closes#31
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
BotBrain plays a match with the commands a player sends: it picks a landing
marker near the barge's route, loots points of interest (following their own
walking routes into huts, lofts and piers), fights opponents it can see,
bandages itself, and gets to ground the coming tide will leave dry. BotWorld
is the map every bot shares: loot markers, routes, and each tide phase's safe
ground, which is the dry land joined to the island's highest walkable ground.
Three difficulties (Tuning.BOT_DIFFICULTIES) set reaction time, sight range
and aim error. The server mixes them unless --bot-difficulty says otherwise.
NavGrid labels connected regions when it is built, so a goal that is cut off
is turned down without a search of the whole island (about 70 ms each).
make bot-matches plays 20 bot-only matches and checks winners, drownings and
difficulty. Seeds 101 to 120: 20 of 20 with a winner, 7 of 460 deaths were
drownings (1.5%), easy bots placed worst (13.7 against 11.8 and 12.0) and
hard bots took the most eliminations (186 against 137 and 132).
Closes#28
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
NavGrid is the map bots plan on: a point per 2 m terrain sample, solid on cliffs and
where solids stand between step and head height, and costlier where the water is
deeper than wading. Path search uses Godot's AStarGrid2D. The grid syncs every tick:
new solids (walls, pieces) block points at once, removed ones are found by a rolling
scan, and points turn wet as the tide and flood reach them. Each change is stamped
with its tick.
NavBot follows a path with ordinary PlayerCommands: it steers at waypoints, sprints
on long dry stretches, jumps when stalled, and marks a point blocked after repeated
stalls. Every half second it checks its remaining path against the change stamps and
plans again if water or a wall now crosses it. One path is planned per server tick
across all bots.
The server builds the grid while the lobby waits and its bots now roam between
random reachable places, replacing the placeholder wander bot.
Tests: 150 of 150 random reachable targets reached on 50 islands in the sweep; the
test runs 2 per island and needs 95%. New walls and rising water both cause a new
path around them; a broken wall opens the grid again; with no dry way round, the bot
swims. 23 roaming bots on an island: p99 tick 1.6 ms against the 12 ms budget.
Closes#27
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Match (src/sim/match.gd) runs a server's matches back to back: lobby,
countdown, drop, play and results, then a fresh island and the lobby again.
It is stepped once per server tick and has no scene-tree or clock dependency.
- MatchRules holds max and min players, team size, the lobby, countdown and
results timers (defaults in Tuning.MATCH_*) and whether bots fill empty
places.
- Participants are server-chosen handles. Joining during a match waits for
the next; leaving during the drop or play places the leaver as eliminated.
- Placements, winners (last team standing) and events for phase changes,
joins, leaves, places, match start and end.
- Match 0 uses the server's seed; later matches derive their seed from it and
their index. The next match's Simulation is built when the lobby opens.
- ServerLoop now drives a Match (immediate rules by default, so the smoke
match is unchanged: same state hash), rebuilds bots each match, and counts
server ticks across matches. The headless server logs match events.
Tests: test_match (13) covers every transition, leaving in each phase, an
empty lobby, an empty match, teams, late joiners, and ten consecutive matches
that each start identical to a freshly built simulation. A ServerLoop test
checks the server goes back to the lobby and starts the next match.
Closes#24
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Seeded loot at every point of interest's loot markers, rolled once per match
from the "loot" stream by LootSet.plan() and placed by Simulation.place_loot().
- Each marker holds a sea chest (chance by location) or floor loot (rolls on
the location's weapon/consumable/ammo/nothing weights). Weapons roll a tier
from the location's tier weights and come with their ammunition; chests
hold a weapon (best of two tier rolls), a consumable and spare ammunition.
- Interact opens the nearest closed chest in reach unless a pickup is nearer,
and spills its contents as pickups.
- Washed-up crate props hold one roll of the crate table and spill it when
broken by any means.
- If the rolls leave fewer POI weapons than LOOT_MIN_POI_WEAPONS (one per
player in a full lobby), extra weapon rolls top up floor spots.
- Chest and crate state is part of state_hash(). ServerLoop places loot.
Tests: test_loot (12), including a statistics run of 1,000 loot seeds over
four generated islands that checks tier shares against the tables, that
shrines beat hamlets, and that every match has a weapon per player at POIs
and 1.25 per player counting crates.
Closes#23
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Simulation.start_drop() plans a seeded straight route across the island at
BARGE_ALTITUDE, through a point near the middle pulled onto land, and puts
every player aboard. Riders move with the barge; jump leaps off with some of
its speed. Freefall steers and tops out at FREEFALL_MAX_SPEED; jump opens
the wing-sail, which also opens by itself WING_AUTO_DEPLOY_HEIGHT above the
ground or water. Gliding sinks slowly and steers at GLIDE_SPEED, and the
sail folds just above the ground or sea so landing does no damage. Anyone
aboard at the end of the route is thrown off. Drop state is hashed.
Every server match now starts with the drop. Bots leap at a time fixed by
their player ID. The carrier's name and look are open in #72.
Closes#14
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
`tools/godot.sh -- --server` runs a headless dedicated server with
--seed, --port, --players, --bots, --ticks, --unthrottled and
--log-every. ServerLoop drives the simulation and times every tick
without a scene tree, so tests run it directly. A placeholder WanderBot
gives smoke runs realistic load until the v0.4 bots land.
Logs are JSON lines that all carry the seed. An ErrorCounter logger
makes any engine or script error exit 1; a bad command line exits 2.
`make smoke` (now part of `make check`) runs 1,000 ticks with 24 bots,
prints min/mean/p50/p95/p99/max tick time against the 12 ms budget,
reruns to confirm the state hash reproduces, and proves a deliberate
script error and a bad config both exit non-zero.
Closes#6
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>