The build auto-detect hint still said HORIZON_DLSS=1 enables the backend;
the player-facing control is now the mod settings toggle (default on),
with HORIZON_DLSS=1|0 as an override.
Co-Authored-By: Kimi Code <noreply@moonshot.ai>
The More page (N → Огонь, ремонт, эффекты…) gains a DLSS row next to
Акустика. Default on: a world without a saved choice leaves the engine's
own default (on where supported). The row reads render.dlss-status (API
level 33): with a DLSS device it toggles live; without one it says
«недоступен», and an on-request is kept and applied at the next launch —
a graceful no-op, never an error. The request is saved in the world's
layout (`dlss true|false`, older saves load unchanged) and re-applied on
world load; the engine persists it in horizon-graphics.txt for the next
start's device opening. /airdef capabilities reports the DLSS state too.
Graphics-only: no resolution control (the player's own Horizon settings
stay in charge) and no sim changes. Engine gitlink → b2cc11c45
(agent/dlss-toggle): the dlss graphics setting, the live evaluation gate,
the dlss control command and the level-33 SDK imports.
Co-Authored-By: Kimi Code <noreply@moonshot.ai>
tools/airdef.py game always built with --features mods, so an RTX user had
to run the manual cargo line from docs/DLSS-TESTING.md. The feature string
now comes from dlss_build_features(): dlss is added only when the OS is
Windows, a Win32_VideoController name has both NVIDIA and RTX, the DLSS SDK
has include/nvngx*.h and lib/Windows_x86_64/x64 (DLSS_SDK, a repo DLSS/ dir
or %ProgramFiles%\NVIDIA\DLSS), and VULKAN_SDK has Include/vulkan/vulkan.h
(dlss_wgpu's build script needs both SDKs). Otherwise the build stays
mods-only; on Windows + RTX without the SDK it prints how to get v310.9.1
instead of failing. GPU/SDK probe errors mean "no DLSS" and never abort the
build; the HORIZON_DLSS runtime switch is untouched.
Tests: tools/test_airdef_build.py covers non-Windows, no RTX, missing DLSS
SDK, missing Vulkan SDK, the full toolchain, a failed GPU probe, the SDK
layout check and the feature string reaching cargo. python3 tools/airdef.py
game still builds mods-only on macOS (fresh target dir: 3m11s, exit 0).
Co-Authored-By: Kimi Code <noreply@moonshot.ai>
The temporary in-session shader switch (harness build, reverted) renders the
same frozen tick with both effects.wgsl versions: old-vs-new differs by at
most 1-2/255, no pixel over 5/255, and no more than two frames of the same
shader differ. Frame time is below the in-game noise floor on M5 (32
screen-filling late-life fireballs: +0.37 ms median, sigma 3.6). Includes
the hd_bench polygon_noon run and test counts.
Co-Authored-By: Kimi Code <noreply@moonshot.ai>
In the HD smoke path fade = 1 - smoothstep(0.08, 0.28, k) is exactly 0
from k = 0.28: every step's density is zero, absorb is 0, so the march
adds no light and leaves transmittance alone while still sampling fbm
and shading the volume. Wrap the loop in an `if fade > 0.0` guard; the
glow after it is untouched (heat is still live there) and is itself
skipped once its heat is spent. Density, steps, sampling and lighting
are unchanged.
Co-Authored-By: Kimi Code <noreply@moonshot.ai>
Measure-first profile naming the culprit with data; no worker implemented
(step 3.2 is out of scope). docs/measurements/issue-147-cpu-profile/: README +
report-server.json (default server-thread path) + report-inframe.json
(--in-frame control), M5/Metal 2560x1440 HD, same-recording battle with
overlapping raids + explosions + dense smoke, each run serialized under the
shared flock /tmp/airdef-bench.lock (earlier concurrent runs discarded).
Instrumentation: host per-tick tick_mods spans (ground/world/snapshot/wasm/
models/output/total, HORIZON_MOD_CPU_PROFILE=1, attributed by tick+frame),
mod Sim::step phase timing via an adapter-injected monotonic clock
(game.monotonic-nanos, API level 32; airdef-core stays SDK-free), raid_bench
--scheduler + per-frame span/frame correlation + mod phase delta.
Findings: on the default path the battle frame is GPU-bound (GPU avg 73.7ms ~
frame p95 88ms; 127/128 frames >33ms); the mod path is ~2.5ms/tick, per-frame
p99 7.7ms, 0 frames >25ms, host world answers <0.1ms, model prep ~0.05ms; the
872ms CPU spike is a render/GPU hitch (mod 2.15ms). In-frame is CPU-bound
(CPU p99 175ms); the mod path reaches >25ms in 27 frames, and step_drones
dominates Sim::step (52%, max 47ms/tick), A* only 15%. The schedulers' mod
peaks do not coincide; only the render-dominated catch-up spike is common.
Recommendation: a pure A* route worker and a world-snapshot worker are NOT
justified (A* leads only the tiny default-path mod; world answers never lead).
Default-path peaks are GPU -> steps 1-2 (GPU/geometry) apply, a CPU worker
would not help. If targeting in-frame CPU, the unit is the whole Sim::step
(Wasmtime on its own thread), not a route worker — still not implemented.
Harness fix (engine gitlink 48eff0fcb): control-socket chat/chat-log now reach
the client mod on the server-thread path (chat-box parity), enabling the
required default-path battle measurement; sim rules untouched.
Tests: airdef-core 326 passed/0 failed/8 ignored; airdef-mod 113+1+2+5 passed
incl. same_seed_same_result and combat_trace_is_deterministic;
tools/test_raid_bench.py 11 passed; horizon-mods suite + wasm ABI test ok.
Engine: d073783f5 (spans + clock) + 48eff0fcb (chat harness) on
agent/126-cpu-profile. Parent: 92c058f60 (profiler + driver) + this commit.
Co-Authored-By: Kimi Code <noreply@moonshot.ai>
Mod side: Sim::step times plan_step, adapt, step_drones, step_moves,
update_sight, step_sams and record, plus the whole step and its untimed
remainder, through sim::profile. The simulation stays SDK-independent (also
built as airdef-core): the adapter installs the host monotonic clock
(game.monotonic-nanos) in init, native tests use std::time::Instant. The
counters are fixed-size, per-thread, read no randomness and change no
decision; /airdef stats reports them as PROFILE chat lines, reset per world.
Driver: raid_bench gains --scheduler in-frame|server-thread (the two
single-player paths, measured separately), HORIZON_MOD_CPU_PROFILE=1, the
per-tick mod-path spans joined to the frame trace (per-frame totals, >25/33 ms
counts, what the worst CPU frames carry), and the mod phase profile read at
idle and battle end with the battle delta. report.json carries mod_cpu,
mod_cpu_ticks, mod_phases_* and the derived correlation. Tests for the new
helpers.
Engine gitlink: mod-path CPU spans + game.monotonic-nanos (API level 32).
Co-Authored-By: Kimi Code <noreply@moonshot.ai>
tools/raid_bench.py (+ tests): the #147 step-0 driver replaces the #126
scratch script. One battle at fixed ticks (waves inside resident terrain —
far waves stall at the unknown-chunk edge and no fight happens), fingerprint
chat lines at control ticks, then the raid's own recording replayed. Both
render paths are benched on the SAME recording in ONE process via the
engine's diagnostic `model-instancing off|on` toggle (cross-process battles
diverge in event timing: 16 vs 32 kills at one control tick), pairs in
alternating order; idle cold/warm, mid-battle window, reload (F3+T), replay
seek, away-camera checks, and static off/on stills for pixel parity
(hd_bench motion-mask methodology). report.json carries gitlinks, device,
mod_models counters, CPU/GPU p50/p95/p99/max, raw frame times, >25/33 ms
frame counts, guest fuel and streaming queues.
Engine gitlink: one mass mod-model group GPU-instanced (retained mesh per
pack generation, current/previous pose/id/light instance stream shared by
main, motion and shadow passes, per-vertex fallback to parity, runtime
diagnostic toggle, stats raw + guest fuel reporting).
Measurements: docs/measurements/issue-147-instancing/ (M5/Metal, 2560x1440;
RTX 3060 pending). docs/REPO-MAP.md and docs/ENGINE.md pin updated.
Co-Authored-By: Kimi Code <noreply@moonshot.ai>
Re-applied the genuinely new #149 features onto kimi-work, REUSING the already-integrated
#146 infrastructure (KnowledgeSnapshot { tick, contacts, marks }, knowledge_now(),
Plan.knowledge, RouteJob.known, versioned price_grid_known and restore). No #146 scaffolding
is duplicated: one snapshot per order, prices over known.tick, versioned restore are kept
exactly as they were. The only additive change to KnowledgeSnapshot is observations().
New #149 layer:
- observe::Observation + observe::merge: Contact and Mark become one public hint, matched by
public features only (area overlap + exact category); coarse Unknown/contradicting marks stay
separate, the contact's kind/fix are never rewritten, recon.truth and real ids are not used.
One Observation::rank scale replaces the two ranks in radar_strikes, choose_target and the
adaptive picks (kimi-work's slice signatures kept).
- price_grid_known prices marks: a trusted launcher/gun/jammer mark prices its published
category envelope weighted by confidence (Confirmed x1, Probable x0.5); a radar mark feeds
the network belief; a mark matched to a contact is priced once through the contact; a
doubtful mark prices nothing.
- Bounded refresh of the unreleased portion: ew::Refresh with absolute deadline
REFRESH_WINDOW=600; step_refresh re-aims queued Plan/Sead drones once on a report (typed
Intent::Refresh, SEAD priority) within SWITCH_CAP and assignment budgets; no report never
delays a launch.
- Bearing window: Intel.outcomes bounded VecDeque (48 entries, 12000-tick window); best_bearing
adapts exploration share by typed BearingReason (few/stale/conditions-changed/exploiting);
conditions are a fingerprint of the public picture, never hidden state. bearing_report() and
bearing/refresh_*/observations fields in EnemyPictureReport. Legacy fixed share kept.
- examples/smart_bench runner (simple/current/candidate, seeds 1/7/42, open/hills/config-change/
recon-lost, JSON+lines); Sim::smart_legacy reproduces current Smart; harness::run_with_hook.
Light modes and bounded jobs unchanged; policy reads no health/charges/hidden switches/positions.
Tests: smart_tests + observe::tests (metamorphic hidden-layout, public dedup, marks in prices,
one version per order in both modes, bounded refresh fires-once/never-pauses, bearing window and
reasons, durable window + Intent::Refresh). The relief-version regression fixture's wall is
widened in z so it crosses any east-west route after the candidate's assignment-order change.
cargo test passes: airdef-core 326 lib + 8 scenarios; airdef-mod 113 lib + core_boundary + 5 scenarios.
Co-Authored-By: Kimi Code <noreply@moonshot.ai>
The heavy WASM scenario (wasm_flat_budget, 64 integrated complexes vs two
`airdef wave oreshnik 40 2600`) lost intercepts at the tick-2102 checkpoint:
522 (main) -> 302 (#133) -> 243 (#134). Decomposed by A/B on one fixture, one
host (2bcce64), one SDK, each slice built from its own commit:
- 522 -> 302 (#133): Track.vel became an estimate from noisy measured places
(it was the target's true velocity); the forecast, the aim lead and the
interceptor's launch Fix all read it. Forcing the class estimate to Unknown
leaves the count unchanged (native 296 ~= 294), so class/roles is not the
cause. The pre-#133 522 used the target's true velocity, which the #133
contract "class and speed only from measurements" forbids; this portion is
the intentional cost of honest measurement-only tracks and is kept.
- 302 -> 243 (#134): the network cue was extended from the separate engagement
radar to every fire-quality radar, including each weapon's embedded radar.
Cueing every relayed contact each tick flooded the embedded fire control with
one-tick plots whose noisy velocity degraded guidance. Restricting the cue
back to separate radars recovers the whole drop: 243 -> 325.
- own_fire_within / FIRE_QUALITY / channels are NOT causes here: relaxing the
own-fire gate or the channel limit each changed nothing (243 -> 243). Embedded
radars always measure Fire plots of quality >= 64 and have no channel limit.
- #140 does not affect this loadout, by fact: the fixture deploys no separate
radar, so every profile is embedded with channels = 0; the channel-age bug
binds only where a profile has channels, and the channels-off A/B was flat.
Fix: the cue looks only through a separate fire-quality radar
(sam.kind == SamKind::Radar); the embedded radar keeps its beam/revisit/lock
acquisition as before #134. All radar profiles, channel limits, own-fire
authority and the measurement-only class/speed contract are preserved; no
target truth is injected.
Measured after the fix (WASM, fuel budget 300M/tick): checkpoint 325 intercepts
/ 2011 hits, fuel max 118M; final (tick 3302, two late Smart raids) 385 / 2472,
fuel max 176M; zero fuel failures. The first late raid reports the #135 recon
flight ("Разведка противника: впереди налёта высланы разведчики"); the second
is silent (recon budget spent) — the fixture now expects exactly that instead
of the stale empty chat that stopped the scenario.
Tests:
- airdef-mod/tests/saturation.rs: deterministic native replica of the checkpoint
(same sim, host block tracer) asserting the restored count and a regression
threshold, plus the #140 loadout fact (no separate radar, channels = 0).
- Sim::launch_audit (why.rs): on-demand tolerance diagnostic counting why each
(system, drone) pair holds fire — no track / stale / no authority / role /
doctrine / reach / claimed / not ready / not aimed / terrain / ready.
- sensor_tests: a network plot cues a separate fire radar but not the embedded one.
- wasm_flat_budget baselines re-measured with provenance comments; the asserted
fixture SHA in tools/airdef.py updated.
Verification: `python3 tools/airdef.py test-budgets` passes both stages and the
two late raids (283 steps, fuel failures = 0). airdef-core 292 + 8, mod lib 104,
mod scenarios 5, saturation 2 — all green. A/B record:
docs/measurements/issue-145-intercept-regression.json.
Co-Authored-By: Kimi Code <noreply@moonshot.ai>
Integration note (shipped engine 8e711f969): the tick-2102 checkpoint baseline is
unchanged and passes on the shipped engine (325 intercepts / 2011 hits). The final
stats baseline was re-measured on the shipped engine with #146 integrated (one
knowledge snapshot per Smart order changes the two late raids): 414 / 2463, world
8885 sections / 420671 rays / 3442845 cells. The asserted fixture SHA in
tools/airdef.py was updated to match. python3 tools/airdef.py test-budgets passes
all 283 steps with zero fuel failures (tick max 184608537 / 300000000).
Engine a3045b61b: the motion buffer's coordinate contract is explicit and
decoded per history grid — native TAAU's stable history no longer receives
the two frames' jitter difference (a still surface reads the same output
pixel on every phase), 1:1 TAA keeps the correction its jittered history
needs, and the NGX input and flags are untouched for separate verification
on the DLSS path (#137). Numerical reprojection tests on all jitter phases
plus an ignored GPU A/B smoke on the real fs_taau.
Co-Authored-By: Kimi Code <noreply@moonshot.ai>
order_raid read one picture and one list of recon marks for the whole
order, but the Plan kept neither: plan_step priced each kind's first
route on whatever picture_of_defences() held at the tick the budget
reached that kind, so one raid could hold several versions of the map
at once. A retarget job carried its version's contacts, yet
price_grid_known aged them at self.tick, so two jobs of one version
reached on different ticks could disagree about a contact's own sensor.
KnowledgeSnapshot fixes the observation tick together with the contacts
and marks read at it. The plan stores the order's snapshot and passes it
explicitly to price_grid_on/price_grid_known; radar_strikes takes the
order's marks instead of re-reading the recon memory. A retarget job
shares its adaptive tick's snapshot and prices age at the snapshot's
tick, never at the current one: current time still cancels a stale job
(prune_air, job_live), separately from the version's own arithmetic, and
an explicitly new retarget gets a new snapshot and may use a fresh
report. The save rebuilds each version's snapshot from the contacts and
the marks reported by it; the section format is unchanged. Nothing here
reads the force: only published kinds and the contacts' own observations.
SDK-free regressions: a Smart order of two kinds whose land answers
starve while a visible contact is disclosed mid-wait keeps every kind's
targets and routes on the order's snapshot (and differs from an order
that knew the contact from the start); two jobs of one version worked
out across the emission-belief boundary price the same world, a new
version with a fresh report believes again, and a new version without
one does not.
Co-Authored-By: Kimi Code <noreply@moonshot.ai>
A station counted any own plot, however old, as a busy refine/fire
channel: a network heartbeat kept a stale contact's track fresh, so two
long-expired own fire plots still held both channels of the engagement
radar and a new contact nearby got only a coarse search plot, even with
the search radar running.
Channel occupancy now follows the station's own tracking.
Track::holds_channel requires an own (never relayed) refine/fire plot no
older than the station's track_ttl — three missed turns, the same value
that already bounds a contact without a fresh measurement. The planner
counts channels with the same Sam::own_channels_used(now) the panel,
monitor, map and "why" views read, so the UI and the allocation cannot
disagree about a stale channel. A stale channel frees for a new contact
without the network's survey track being removed; the hard per-profile
limit and the own measurement's provenance are unchanged, and a relayed
plot still never holds a channel.
Test: two contacts take both channels at the station, then leave its
reach while the search radar beside them keeps relaying them; a third
contact arrives nearby and gets its own fire channel (the test fails on
the old counting), and the relayed survey tracks of the old contacts
stay.
Co-Authored-By: Kimi Code <noreply@moonshot.ai>
The replay's sounds, the raid map's counts and the struck objects now read
the recording's event and scene journals by time instead of every pass over
the whole journal: partition points over the journals (with a stable time
index built at close only when a journal is out of order), the map's kill
and hit prefix indices, each object's first hit, and marker/time-line paths
that stop at the map's shape cap. Blasts, smoke, debris and the fire replay
read bounded windows back from the moment.
The sets, order and names stay exactly as they were: tests compare every
reader against the full journal filter, including tied times, out-of-order
journals and arbitrary forward and back seeks. The replay_bench example
measures a 10-minute raid with nearly full journals (playback, reverse seek,
pause, MP4); its result is docs/measurements/issue-150-replay-ranges.json.
Co-Authored-By: Kimi Code <noreply@moonshot.ai>
Engine 1cadd6338: the DLSS frame now submits world+linearize -> NGX ->
delinearize -> film/tonemap/UI/capture/present in one queue.submit with no
CPU sync, so the final frame, screenshot and export read the current
resolved frame instead of the previous one. Native TAA history readiness
is tracked separately from NGX, so the first fallback frame after DLSS
does not blend a stale or empty native history; a cut/resize clears both.
HD-RENDERER.md records the corrected order and the history separation.
Real-RTX acceptance (docs/DLSS-TESTING.md) is still pending: no RTX is
available here, so this is verified by the ordering/history unit tests and
the native horizon-client + horizon-mods suites, not a DLSS GPU capture.
Co-Authored-By: Kimi Code <noreply@moonshot.ai>
A battery's "why" now starts with one typed reason shared by the monitor,
the HUD and the chat: only observation, waiting for a track, busy channels,
no ammunition, or ready. It carries the age and the station of the mark it
is about, and `/airdef why <n> source` moves the monitor to that station.
The reason is computed from the same predicates the launch decision uses,
and the detailed decision settles it, so a headline cannot promise or deny
a shot the simulation does not take.
The after-action review builds 4-5 measured steps per contact from the
recording's scene journal and events: mark, handoff, lock, launch, outcome.
The menu lists them per battery with a filter and opens a replay frame on
a click (the outcome clip, else the map held at the launch or mark);
`/airdef chains [battery]` prints the same list. Missing transitions stay
missing and marked as no launch, never guessed, and the normal HUD still
shows only the battery's own picture.
Co-Authored-By: Kimi Code <noreply@moonshot.ai>
An automatic optical contact is kept only within the optics' own reach and let go after a few failed probes. It stays out of the next picks for a while, so a visible alternative can be tried without constant switching; unknown world geometry gets a longer, still bounded wait. A player's command still waits for its contact and says so in the status line.
Ready posts are now served oldest-probe first through a simulation-wide probe sequence, so posts whose phases coincide all advance under the unchanged CHECKS_PER_TICK budget.
Tests: an out-of-range contact with a visible alternative, a blocked contact with a visible alternative, five same-phase posts, the bounded unknown-geometry wait, and the manual pin's status and out-of-reach behaviour.
Co-Authored-By: Kimi Code <noreply@moonshot.ai>
Sampling slowed by the pose budget skipped whole short stages: the six-tick
tube cover opening, the ejection and the radar switch fell between samples,
and thinning then halved already recorded transitions along with poses.
Poses still share the adaptive sam_step grid and the unchanged 32768-key
budget, but the recording is now sampled every tick for facts: a discrete
change (spent_cells, cap_cell, radar_on, mode, role) and every frame of the
cover's own movement are kept on the tick they happen, with the stage's start
backdated one tick so the opening still spans exactly OPEN_TICKS. These keys
carry an event mark and decimation halves poses only, keeping events, the
first and the last key.
Events get their own explicit 4096-key budget instead of raising the pose
budget: when it runs out the oldest stop being critical and thin as poses
while the newest stay exact.
Test: 64 turning radars plus an S-400 at sam_step 16, ejection on tick 4103
(cover 0.8333333 a tick before, spent cell on the ejection tick), a cancelled
opening and a radar switch off the pose grid, all surviving further thinning;
a 64-radar chattering recording stays exactly at the event budget with the
newest switch exact.
Co-Authored-By: Kimi Code <noreply@moonshot.ai>
A look no longer reports a wreck as a lead: a first look at one makes no
mark, and one the picture already held is radioed home as destroyed and
takes its mark out of the picture. A hidden kill still never rewrites a
mark, so the old assessment stands until its age runs out.
The channel now follows actual work, not a switch: a radar emits only
while radar_up, a jammer only while its transmitter is on the air. A
silent, packed, failed or destroyed station is just a unit the flyby saw.
A pending wreck observation is durable; older saves without its flag read
as before, every look being of a unit alive when it was seen.
Co-Authored-By: Kimi Code <noreply@moonshot.ai>
The real-.hmod scenario still expected the pre-#134 role switch to
retune radar 4 and return its role as feedback. After #134 the profile
is chosen at deployment and dispatch refuses the old syntax, so step 14
failed with no feedback.
Both stale checks now expect the intended refusal, naming
deploy radar_search|radar_track|radar_height and on/off. The deployed
station is not transformed back for the test.
Verified with python3 tools/airdef.py test-wasm on the pinned host
2bcce64236 and the real airdef.hmod: 30 steps pass.
Co-Authored-By: Kimi Code <noreply@moonshot.ai>
The overlay clipped the network's tracks to the nearest 64 by distance
before projecting them, and the truth and interceptor layers the same way.
A crowd behind the camera filled the budget, so a raid could show nothing
even with a target right in front of the player; the native reproduction
(one contact ahead, 64 behind) came out with zero markers and a header
saying all 65 were out of frame.
Now visibility comes before the draw budget. Each layer orders its
candidates nearest first, projects at most MAX_SCAN = 4 x MAX_MARKERS of
them (bounded work, no sort of the whole picture beyond that prefix), skips
those the frame does not hold without spending a marker, and draws the
nearest MAX_MARKERS visible ones. Each layer keeps its own draw budget, so
the interceptors and the truth markers cannot crowd out the network's.
The header separates the two ways a target stays unmarked: "за кадром"
(behind the camera or past an edge) and "сверх бюджета" (past the scan or
draw budget), alongside "в кадре".
Regression tests: 64 behind plus one ahead for the network, truth and
interceptor layers; a mixed crowd proving the nearest visible keep the
markers and the farthest is budget-capped, not offscreen; the projection
count bounded by MAX_SCAN; every case stays inside the HUD's 1024 shapes.
Co-Authored-By: Kimi Code <noreply@moonshot.ai>
Holding C draws square markers over the targets the player can see, each
with the class the radars estimated and the distance; an unidentified
track stays a "?", never the target's type. Stale forecast tracks fade
orange and say their age, the player's own interceptors wear small yellow
squares, and the sandbox truth mode adds white markers of the real drones
apart from the network's estimates. Letting the key go clears the HUD.
The markers project through the camera of the last drawn frame
(render.project, API 31), so they sit on the world as it is drawn and
follow the view frame by frame; the overlay stays off under a replay, the
grid J, the monitor K and the map M, which own the screen themselves. A
full raid stays inside the HUD's 1024-shape budget, keeping the nearest
markers and counting the rest behind the camera or past the edges.
Bump the engine pin to 2bcce6423 (API 31, render.project) and name API 31
in the capabilities diagnostic to match.
Co-Authored-By: Kimi Code <noreply@moonshot.ai>
The core recon flights and partial reports (9d8f3da57) had no player-facing
or diagnostic consumer: enemy_picture().recon/recon_sorties were read only by
core tests. This adds the missing layer on top, without touching the mechanics,
models or textures.
- Map: a tracked recon aircraft is a recognisable yellow hollow diamond with a
"разведчик" label, a role line in the selection panel and a filter legend; in
the sandbox truth mode it gets an amber marker of its own.
- Battle UI: the summary panel and the map show the flight's coarse phase
(разведка → донесение → завершение) only while the player's network actually
holds a recon aircraft — what the visible aircraft are doing, not what the
enemy learned. The truth mode adds a recon count line.
- Diagnostics and after-action review: /airdef recon and the menu page
"Разведка противника (разбор)" (from Main and Replays) list the enemy's marks
best-first with the reason each outranks the rest — priority = freshness×4 +
quality×3 + observed role (recon::priority_parts, cross-checked against
Mark::priority by a test) — plus the sortie phases and the drones aimed at a
mark. The omniscient enemy picture is never shown in the normal HUD; it is
explicit debug/review only.
recon_view reads only the current report types (EnemyPictureReport, ReconReport,
ReconSortieReport); the sim side only widens visibility and adds priority_parts
and the is_recon_drone/recon_status read helpers. Tests: recon_view, panel,
command tree, priority_parts cross-check. README and IMPLEMENTATION-STATUS updated.
Co-Authored-By: Kimi Code <noreply@moonshot.ai>
- ClassEst::step_optical: the shared estimate's rule for one passive look —
an unsure look (over OPTICAL_CLASS_MAX_UNCERTAINTY) names nothing, a sure
one names a class, firms an agreeing probable, replaces a contrary guess
- Track::refine_class_optical: the minimal common method — class only; the
plot, age, source, quality and own fire rights stay untouched
- Sim::apply_optical_classes: the explicit adapter, fed by each observation
made this tick through OpticalObservation::estimate; a relay of the
refined track carries the class as it was and gains nothing from it
- the post still holds no track, emits nothing and authorizes no shot;
the refined class reaches the existing #133 track UI through the picture
- tests: network refinement, unsure look, relay/stale look add no
confidence, equal observables refine equal classes, the step rules
Co-Authored-By: Kimi Code <noreply@moonshot.ai>
- a radar's fresh track reaches a post as a cue only: it points the head
but never enters the post's tracks, seen or own set
- a searching post's sector follows the turning head, not the fixed base
heading; switching contacts drops the old contact's picture
- dead and blinded posts spend no probe budget and never observe
- the icon alias is part of the generator, so the placeholder item uses
the radar texture without a duplicate PNG or a hand-written asset
- part diagnostics read "Опт" for a passive post's sensor
- tests: cue wiring, turning sector, contact switch, dead post, graphics
preset; generator alias
Cherry-picked from 3697a1764aec7a7bdc6311918453dd51e26ecabf onto the
#132/#133/#134 base: the cue-wiring test's track literal now carries the
shared class/source/product/quality contract instead of a DroneKind.
Co-Authored-By: Kimi Code <noreply@moonshot.ai>
A smart raid may send a small, bounded flight of unarmed Gerbera aircraft
ahead of its strike. They fly physically through zones round the marked
objectives (and revisit a doubtful mark when nothing is unchecked), look
at units inside a zone, and radio home an area, a public category, the
channel, age and quality only after a delay and only while the aircraft
is alive. A sortie lost before its report sends nothing; marks already
delivered stay, and the picture remembers them across raids.
The planner now reads recon marks as well as disclosure contacts. A mark
carries a synthetic id, an area, a category and a radar role only once a
second independent look refines it -- never a real id, SamKind, exact
point, health, magazine or link. Aims are chosen by observed role,
freshness and quality; a fresh modest mark can outrank a stale doubtful
big one, and a hidden unit is in neither list. In-flight recon aims keep
their old area: hidden moves, re-roles and damage do not rewrite a mark.
Bounded assignments, the route-job queue, the original wave objectives
and the no-eternal-pause rule are preserved. Recon aircraft take ids from
their own range, do not spend the strike's combat RNG and are not part of
the strike's replay. Save/load keeps marks, sorties, pending reports and
checked zones. No new models, textures or weapons.
Co-Authored-By: Kimi Code <noreply@moonshot.ai>
- new Optical unit kind with its own catalog row and deploy item
- bounded one-contact line-of-sight checks against land and blocks
- day/night reach and hold timing; obstacle and unknown geometry block sight
- independent class observation with source, time and uncertainty, never DroneKind
- map/monitor/panel statuses and a line to the selected contact
- tests: day/night, wall/unknown, ageing, no kind leak, bounded work, passivity
Cherry-picked from 3ae21123964a9d3234bb2b3efc9ff926d64fc77d onto the
#132/#133/#134 base: the cue test's track literal now carries the shared
class/source/product/quality contract instead of a DroneKind.
Co-Authored-By: Kimi Code <noreply@moonshot.ai>
The maximum menu raid keeps a supersonic contact in the air past 5000 ticks under the new fire-authority rules; the recording closes only when the sky is clear. Run the scenario to 8000 ticks, still well inside its 20-minute cap, and keep the replay assertion.
Co-Authored-By: Kimi Code <noreply@moonshot.ai>
Give the separate radar three real profiles over SamKind::Radar and the
typed catalog variants:
- radar_search: coarse, slow, long-range search plots; no fire authority.
- radar_height: raised mast, local refinement of low plots; no fire authority.
- radar_track: local search, fast revisit, two refined/fire channels.
- radar_multi remains the compatible universal profile, not best everywhere.
Tracks now carry the measuring station, product (search/refine/fire), and
quality (separate from identification), and keep the station's own plot when a
fresher network relay replaces the published position. A relay never refreshes
the measurement or authorizes a launch. Ordinary launches need the battery's
own confirmed fire plot; launcher-only batteries use a supporting radar's own
plot, and the existing heavy far shot keeps its cue-radar exception.
Per-tick refinement/fire work is bounded by profile channels and local search
radius. Off, damage, network and save/replay behavior are preserved: new saves
write the stable profile ID, old role-only lines migrate, and embedded SAM
radars keep their own fire authority. The deploy menu, status lines, map,
monitor, radar scope and /airdef why explain the profile, source, age, quality,
channels and refusal reason.
Models/textures are untouched: all profiles reuse the existing radar
silhouette (the separate art zone remains for its owner).
Tests: profile single/pair/full scenarios and duplicate station, channels,
relay source/age/quality, Off/damage, save migration, replay, embedded and
Multifunction radars, catalog/generator checks, airdef-mod suites, core lib
(non-balance on one thread; the four active balance tests in release) and core
scenarios.
Co-Authored-By: Kimi Code <noreply@moonshot.ai>
A radar track is what the network measured, and it never carried the
target's own answers: the tactical Track loses `kind: DroneKind` and the
true velocity, and gains what the measurements say instead.
- Track now holds a ClassEst (unknown / probable / firm) plus an observed
velocity: none from the first blip, then the change of the measured
places between real measurements, smoothed so one noisy pair does not
flap the class. The classifier's bands sit between the game's speed
clusters (Class::of_speed); a new contrary measurement replaces the
hypothesis, an agreeing one firms it.
- The class estimate is separate from track confirmation (Track::ident):
a track can be confirmed with its type still unknown. The estimate
grows only on real measurements — a forecast and a relayed copy add
nothing — and it travels with the measurement it grew from, ageing
with the track and gone when the track drops.
- Roles, last chance, the reserve and the interceptor count read the
estimate, never the target's class. Explicit policy for the unknown:
an unidentified track stays a candidate under the general rules (and
the system with no roles holds its fire at everything); the map, the
monitor, the summary panel and /airdef why read the same estimate and
name the reason — «Трек подтверждён · тип не определён», «вероятно
крылатая ракета». The exact model stays in the sandbox's truth mode.
- A gun or a launcher lets go of a target that classified out of the
roles since it was taken; a first blip no longer rules a contact out
as a threat to guarded objects (it has no measured course yet).
Old saves and replays are untouched (tracks are derived, not persisted);
the simulation stays deterministic (the classifier is pure arithmetic on
the measurement noise). The doctrine/first-blip/last-chance tests gained
classified tracks where the contract provably changed; new tests cover
the classifier's steps, kind-independent estimates, first-blip blindness,
and relays and forecasts adding no confidence.
Co-Authored-By: Kimi Code <noreply@moonshot.ai>
- mini radar: class-shaped blips that merge into one with a count
- map: larger track markers and trails, hover card with distance, type,
status, target and interception
- explicit network vs sandbox truth: true drones, routes and aims drawn
apart from the network tracks, with their own trails
Co-Authored-By: Kimi Code <noreply@kimi.com>
Engine a861c1ed2 (kimi-work-124), based on ed751714e: mod-model faces resolved once per pack generation, frames only transform and light them; motion #124 from the same faces; counters and span; reload/unload invalidate the cache. HD-RENDERER status + section, and docs/measurements/issue-126-first-stage with the A/B method and numbers (M5/Metal, Polygon raid: entity_geometry CPU 5.71 -> 1.57 ms avg, prepare 23.5 -> 7.05 ms; static scene pixel-identical; reload/unload smoke). Not GPU instancing; 5600X/3060 acceptance pending. Issue #126 stays open.
Co-Authored-By: Kimi Code <noreply@moonshot.ai>
Stage 1 of #125: prove or disprove real hardware ray-tracing support without
touching the client renderer. The probe pins the same wgpu 30.0.1 stack as the
engine, including its vendored third_party/wgpu-hal patch, builds one BLAS/TLAS
and traces a 16x16 ray-query compute frame, checking every ray against a CPU
point-in-triangle expectation.
Verified on macOS aarch64, Apple M5 / Metal: BLAS/TLAS and the ray-query
pipeline build, 32 hits and 224 misses all match, exit 0. RT_PROBE_BACKEND=gl
reports no RT adapter and exits 2; RT_PROBE_NO_OPTIN=1 shows request_device
rejects EXPERIMENTAL_RAY_QUERY without the experimental opt-in and exits 2.
tools/rt-probe/target/ is ignored; the probe is a standalone workspace and is
not wired into the game build.
Co-Authored-By: Kimi Code <noreply@moonshot.ai>
The battery deploy pieces (s400_launcher, s500_launcher, radar_search/
track/height) were described three times: the menu table and the
configuration match in lib.rs, and a chained id list in commands.rs. One
catalog decision instead: catalog/content.json gains a variants section
(mode/role reuse the existing SamMode/RadarRole save keys), gen_catalog.py
generates catalog::UnitVariant rows, and the loader rejects duplicates,
unknown kinds/modes and id collisions with kinds or items. The battery
menu, its click handler and the deploy completion read the generated
table; the historical row order and labels are unchanged.
No variant carries its own balance: per-variant ammo profiles stay open
pending a confirmed profile and an explicitly labelled game abstraction
(PAC-3 note on #91), so no new specs, no new matches and no save/replay
change - a variant is only a deploy key over the same kind.
Checks: gen_catalog.py --check and gen_models.py --check clean, the
python catalog/asset and sound tests pass, airdef-core 234 unit + 8
scenarios, airdef-mod 69 unit + 1 boundary + 5 scenarios (save/load and
state round trips included), zero failures.
Co-Authored-By: Kimi Code <noreply@moonshot.ai>
Sam::status, Comps::describe/worst and Sam::parts_text formatted
player-facing Russian in the simulation. The adapter now owns that
text: battery_view::sam_status for the panel and map, why_view::parts
and why_view::worst for the monitor and the status line. The core keeps
the data and gains only the accessors the adapter needs
(preparing_cell, public cells/reload_time/magazine/gun facts).
Texts are unchanged; the adapter tests pin the existing strings.
Co-Authored-By: Kimi Code <noreply@moonshot.ai>
Advance to engine ed751714e. Document the optional SDK build, ordered color conversion/submission, native fallback and the remaining MSVC-link and RTX-runtime acceptance.
Co-Authored-By: Kimi Code <noreply@moonshot.ai>
Advance to engine 7a73e8d26: separate previous-position streams and a depth-read-only motion pass for mod models, with matching alpha cutouts and a debug view.
Co-Authored-By: Kimi Code <noreply@moonshot.ai>
Draw the player marker before network tracks so a close blip is not hidden. Add a shape-order regression and verify a live raid with truth disabled.
Co-Authored-By: Kimi Code <noreply@moonshot.ai>
Advance to engine 17382445a and document the six-block near/far transition verified on the forest aerial view.
Co-Authored-By: Kimi Code <noreply@moonshot.ai>
Advance to engine c962cd472. Document the render-pixel current-to-previous contract, native TAA/TAAU consumer and the remaining DLSS/object-motion/linear-HDR work.
Co-Authored-By: Kimi Code <noreply@moonshot.ai>
Advance the engine to 94b572cd2, backporting upstream f7b77fbc8 with its GPU growth/readback regression. Keep API30 unchanged.
Co-Authored-By: Kimi Code <noreply@moonshot.ai>
Thin old SAM keyframes in place, reduce allocation capacity and adapt later sampling when a recording exceeds its fixed pose budget. Preserve recent poses and cover the full ten-minute window in native and real WASM stress tests.
Co-Authored-By: Kimi Code <noreply@moonshot.ai>
Stage 2 of #127. A smart drone now carries why it is aimed where it is
(Intent: plan, sead, react, replace) apart from the aim bookkeeping
(unit, since, switches, confidence). Assignments are throttled per
adaptive tick and per contact, a fresh disclosure replaces an aim only
after a hold, within a switch cap and when the change is worth a new
route, and a killed aim is replaced as a separate, lower-priority
decision. The queue is bounded and taken by priority (plan > sead >
react > replace); unknown land is no longer free in a replanned route.
The autosave now carries the air state: the drones in flight, the launch
queue, every pending job with its id, aim, unit, version, priority,
intent and relief version, the contacts the jobs were chosen from and
the relief snapshots they plan over. The format is a versioned `air 1`
section; older saves read as before, scenarios do not carry the sky,
stale, ghost and duplicate jobs are cancelled on load, and the section is
bounded (512 drones, 512 route points, 4096 contacts, 8 reliefs, 900 KB).
A periodic save waits while a raid is only being planned.
enemy_picture() now explains targets and pending jobs (aims, aimed_by,
pending_jobs, reason, work_per_tick) in English, and intel_bench records
the queue and intent evidence next to completion, losses and failures.