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.
Write both measurement CSVs with LF instead of CRLF so the range no longer
fails `git diff --check` on trailing whitespace; the parsed data is unchanged.
Add a matched before/after capture: one copied world, scenario and seed, the
same client/director and the same export command. The baseline hmod is built
from dc36d67 and the final one from the #129 code. Both launch shots start at
frame 94; the first centre illumination is frame 101 (0.233 s) before and
frame 121 (0.900 s) after, against the 0.25/1.00 s cold rules. Raw probe,
hashes and the measurement method are in issue-129-pair.json. The two MP4s
in docs are CRF26 storage copies; the CRF16 originals stay in the scratch
homes.
These are replay exports, not live footage. The capturing agent still has no
image input, so composition is left for human review; the frames before the
launch are not pixel-identical because these are two live sessions.
Co-Authored-By: Kimi Code <noreply@kimi.com>
The stage contract and profiles are in 4b013fb. This commit adds the
measured balance (the #120 grid on identical seeds and the 12-seed raid),
the replay/export video with four stills, and updates the user and handoff
documentation. The visual files are provided for human review; the agent
that captured them has no image input and checked only the encoded
parameters, shot boundaries and the first large plume.
Co-Authored-By: Kimi Code <noreply@kimi.com>
The interceptor's launch is one sequence: the cap opens, the round is
thrown out unlit and slows under linear gas drag (`cold` ticks), the motor
lights where it has reached, it turns with a growing rate while thrust
spools up (`boost` ticks), and full guidance starts only at `cold + boost`.
A rail launch lights at once but spools over the same `boost` ticks; the
electric P1-SUN has no rocket spool or plume.
S-400/500 now hold the unlit ejection for 20/24 ticks (1.00/1.20 s at
20 Hz); IRIS-T/Tor keep shorter 8-tick throws. Ignition, its sound, flame
and recorded MotorIgnition event all use the missile's current position.
The RM48U strike follows the same 20-tick choreography, counts the throw
in its flight time and starts its recorded smoke at ignition.
Maximum speed, turn, accel, pk and every collision rule are unchanged.
Co-Authored-By: Kimi Code <noreply@kimi.com>
The planner caches were keyed by (kind, relief Arc pointer, start, aim): two jobs with the same relief but different known-units snapshots could reuse each other's price grid and route. Key relief sampling by its Arc pointer but key prices and routes by the whole knowledge version: adaptive tick plus the known-units and relief Arc pointers. Regression creates two same-start, same-aim jobs on one relief with different known snapshots and checks their routes differ, while a third job of one version still reuses the route cache. Core release, adapter and WASM contract pass.
The relief snapshot was still global: a smart plan finishing while a retarget job waited could replace Sim::known_relief, so the pending route was planned on the newer map. RouteJob now carries Arc<KnownRelief> for its knowledge version; Sim::known_relief keeps the same Arc only to seed new jobs. retarget_relief reads the job's snapshot, and relief sampling, price grids and route caches are keyed by that version. The new regression makes a job wait, swaps the global relief for another grid with a wall, and checks the pending route and sample stay those of the original version; outside its grid is Unknown. Core release, adapter and WASM contract pass.
A mass retarget used to clear every selected drone's route before the planner's per-tick work ran out: the rest flew straight for ever. Route jobs now carry the drone id, new aim, target unit, knowledge version and the intel snapshot, stay in a FIFO across ticks and are drained under the same WORK_PER_TICK. An old safe route is kept until its replacement is ready; a cancelled job falls back to a straight run. Jobs are replaced in place per drone, cancelled when the drone or unit is gone, and retarget routes reuse the original plan's known relief. An in-flight drone no longer turns to a unit behind its turning circle. Core release tests, adapter tests and the real WASM contract pass.
step_supply sent every free truck into the return branch whenever any
S-400 had a pending ground_return, even when that return could not
proceed (the system moving, damaged, reloading or preparing, its round
already claimed by another truck, the route marked unreachable, or no
depot with room). The branch then found no target, cleared the job and
stopped the truck, so independent typed supply requests starved.
The return now keeps its priority only while it is executable: a
carrier holding a reserved round looks for a depot with room in reach,
an empty one for a system ready to hand a round over. Otherwise the
claim stays untouched and only real RM48U demand (a requesting system
with depot stock, or ground rounds aboard for one) pulls the truck
into the typed supply branch; without such demand it holds its watch
exactly as before. Redirecting a reserved round to a waiting request
clears return_cargo once the last one leaves the truck, and a stale
return_cargo with an empty hold no longer sends the truck to a depot
handover it cannot complete.
Regressions: three sim tests (moving claimant, pickup claimed by
another truck with damaged/save-restore phases, unreachable route) and
one full-drive lib test with a mid-flight state round trip; the rounds
are conserved in every phase and the actual return completes once the
claimant is back. All three sim tests fail on the old code.
Checks: core release 200 passed/8 ignored + 8 scenarios; adapter
67 unit + 1 boundary + 5 scenarios; real WASM test-wasm 30 steps.
The Windows log for #122 showed the mod stopping in the UI event
(fuel budget 1e9, remaining 0) with known_units <- order_raid <-
launch_wave <- ui_select: order_raid recomputed the planner snapshot
known_units() for every drone, and the snapshot itself rescanned the
whole force and the dedup list per intel entry. A second big Smart
order cost drones x intel-log x force-size inside one event.
Reproduced in actual WASM with measured per-callback fuel through the
exact ui_select(BTN_LAUNCH) handler (diagnostic guest, trigger.patch):
before, a 111-drone second order at revealed=13172/known=144 exhausted
exactly 1e9 and stopped the mod (an 11-drone order at the same state:
247805865); after, the same scenario passes with max command fuel
32946177/1e9, failures=0, and /stats at tick 2602 byte-identical.
A stacked single-point layout dies to one warhead (known=0, cheap
order), so the scenario deploys spread clusters with a moved player.
Fix: one known_units snapshot per order_raid passed to radar_strikes
and choose_target; known_units is a single O(log + force) set pass;
price_grid_on reuses its snapshot. Emitter/network targeting and
carrier choice semantics are unchanged (native equivalence test
against the former predicate; WASM checkpoint counters equal).
Limits unchanged; HD/recording untouched; wave size unrestricted.
Committed regressions: native snapshot-equivalence and
second-smart-order tests; the heavy WASM fixture now orders two Smart
raids through the wave screen's launch function (airdef preset +
airdef raid) with EW and a combat intel log, four accepted orders
asserted; the light contract covers preset and the empty-wave refusal.
Core release 195 + 8 scenarios; adapter 65 + 1 boundary + 5 scenarios;
light WASM 30 steps; heavy WASM 283 steps (tick max 182413686/300M,
command max 6939810/1e9, failures 0). Windows GUI acceptance remains
external and is not claimed.
Co-Authored-By: Kimi Code (Moonshot AI) <noreply@moonshot.cn>
The final engine 6a53a626c with the API30 AirDef build was run in the
real game (macOS, Apple M5, Metal, --offscreen, a clean HORIZON_HOME and
copied test worlds), not only in the sandbox: the monitor and the film's
main view report ready with the picture actually drawn (screenshots show
the panel and picture area, not a registered target), an empty set-views
drops the request, a world change bumps the generation and clears it, and
AirDef's next submission prepares the camera again in the new world.
Fixes nothing: no defect showed. It also records what the run does not
prove (a human visual check, Windows/other GPUs, two runs with the same
local id on the GPU, a mod reload on the GPU), the per-tick
set-views/set-main-view ordering that can honestly read awaiting-frame
between tick and note (not flicker), and that a title trip back into the
same world keeps the request because the generation is per world.
Co-Authored-By: Kimi Code <noreply@moonshot.cn>
Pin the engine at 6a53a626c (API 30) and show render.camera-status: the
calling run's set-views cameras, whether the last noted frame was a world
frame, the request's world generation, each camera's reason and the run's
own main-view selection. As the other lines, this is a host milestone:
availability, request, preparation and why, not a GPU frame or a visible
picture. A unit test pins the line: readiness stays per camera and apart
from the run's main-view selection.
docs/ENGINE.md, the 4 November continuation and the status table now say
what prepared means (a picture encoded into the camera's target, or a
world frame drawn from it as this run's screen view, only for the run
whose selection the screen used), the per-run/per-world resets, and what
remains: GPU acceptance.
Verified: python3 tools/airdef.py check (gitlink, packer, mod needs 30,
fork 30, game 30); test-wasm 27 steps; airdef-mod 65 unit, 1 boundary and
5 scenarios; python3 -m unittest discover -s tools 28 tests.
Co-Authored-By: Kimi Code <noreply@moonshot.cn>
Wider same-seed sampling of the finite-burn/coast change: baseline 2689d93
(pre-#120) against the current tree, 6 variants x 8 seeds of a 46-drone raid
(batt10/20/50, near200, far1500, hills20). Harness-only: Result carries the
tally's fuse/expired/orphaned/energy_lost counts and a new ignored
sim::balance::issue_120_grid test; no simulation parameters changed.
Findings: kills are equal or up and leaks equal or down in every variant;
fuse misses drop (far1500 70 -> 38 over 8 seeds); hills terrain losses fall
(scrubbed 6 -> 2). MotorEnergyLost never fires in the 48 scenarios and
expired stays <= 1: coast ends before drag reaches the 25% guidance
threshold. Leaks stay concentrated in the ballistic/fast classes
(Iskander/Kinzhal/Kh22/Zircon). No confirmed defect, so no tuning.
Full airdef-core suite: 193 + 8 passed, 0 failed (baseline side 169 + 4).
Co-Authored-By: Kimi Code <noreply@moonshot.cn>
The armed/support check in load_specs named the ten armed kinds, so a new
unit type would still have required a code edit and repeated the coupling
#91 is removing. Drop it: optional missile/gun rows are still validated
generically, and the existing kind-coverage check against content.json
already ties specs to the catalog.
validate_specs now takes the document, content identities and sound
events, and a test shows a hypothetical unit kind is accepted from a
specs row plus its content entry alone. No data or balance changes.
Verified: gen_catalog --check fresh, test_gen_assets 10 tests, core
release 195 passed + 8 scenarios (7 ignored), adapter 64 unit +
1 boundary + 5 scenarios, WASM contract 27 steps.
The old engine_voice returned early for flame == 0 with
(DRONE_LANCET, 0.8, 1.6), so the later `P1sun => 1.7` match arm was
unreachable. The catalog migration mistook that arm for the played value
and stored 1.7, which changed the interceptor drone's flight sound.
specs.json is back to 1.6 and specs_generated.rs is regenerated.
Adds missile_engine_voices_keep_their_played_values, pinning P1-SUN to
(DRONE_LANCET, 0.8, 1.6) and M48N6 to (MISSILE_FLIGHT, 1.0, 0.88).
The continuation and handoff text now describe the early return instead
of the dead arm, and tools/airdef_assets/catalog.py gets its final
newline back.
Verified: gen_catalog --check and test_gen_assets, core release
195 passed + 8 scenarios (7 ignored), adapter 64 unit + 1 boundary +
5 scenarios, WASM contract 27 steps.
catalog/specs.json is now the single copy of the numeric rows for
DroneSpec, MissileSpec and SamSpec: speed, guidance, warheads, models'
parts, RCS, seekers, launch and motor sounds, mount drives, gravity,
the Oreshnik split height and logistics limits, plus the named part
sets. tools/gen_catalog.py generates airdef-mod/src/sim/specs_generated.rs
(included by mod.rs, so the mod and airdef-core compile the same rows)
and unit_specs_generated.rs (included by units.rs). The hand-written
spec matches are gone; DroneKind::spec, MissileKind::spec and
SamKind::spec read the generated constants.
load_specs validates the schema, sound references, kind coverage against
content.json, named-part references and unused rows. gen_catalog.py
--check already runs in the CI generated-assets step, so a JSON/Rust
drift fails the build.
Every value was compared with the literals at 632061d field by field.
The only mismatch was the P1-SUN motor voice pitch in JSON (1.6 instead
of the actual 1.7, the flame == 0 branch in voices.rs); the JSON was
corrected and behaviour is unchanged. Save keys, raid order, RNG and
gameplay traces are untouched.
Verified: core release 194 passed + 8 scenarios (7 ignored), adapter
64 unit + 1 boundary + 5 scenarios, WASM contract 27 steps, catalog and
model generators fresh, asset/sound/model-bound/report tests pass.
Pairs of the same replay exported with acoustics on and off on the
scratch world (no HORIZON_NO_AUDIO): acoustics and capture ownership
both ready. The open field has no reverb send by design, so it shows
occlusion/air absorption only; the stone court's on-track is quieter
and darker, and its residual after a gain fit is low-passed and
stereo-decorrelated, i.e. the wet path. A clean reverb tail/RT60 in a
real MP4, echoes alone and listening remain open. No code changes.
Co-Authored-By: Kimi Code <noreply@moonshot.ai>
Engine 60cc7b11d removes a failed or cancelled take's video, raw sound
and log, carries the log's last lines in the error, and only creates
<name>.mp4 after a successful finish. A real failing encoder on the
scratch world left no temporaries and kept the earlier film byte for
byte; successful audio and no-audio control films still verified.
Co-Authored-By: Kimi Code <noreply@moonshot.ai>
horizon 60cc7b11d (on top of 57f8b2853, API 29) removes a failed or
cancelled take's video, raw sound and log, carries the log's last lines
in the error, and only creates <name>.mp4 after a successful finish.
Co-Authored-By: Kimi Code <noreply@moonshot.ai>
A frame hook ends an active film or export when the world is gone. The game
stops ticking at the title screen but keeps rendering, so an export used to
keep recording the title with its playback clock frozen (a sec=10 request
produced 3898 frames) and went on after re-entering the world. Session
cleanup runs once with WorldChange and cancels the capture.
Real short exports on a scratch home and world (pin 57f8b2853, API 29,
Metal): h264+AAC both 2.07 s with non-silent audio, reported frames equal
ffprobe nb_frames; cancel, SDK refusal, encoder failure and restart all
leave no capture running; 1920x1080 works. The engine still leaves
*.audio.raw and *.ffmpeg.log behind on an encoder error (game/film.rs).
Co-Authored-By: Kimi Code <noreply@moonshot.ai>
Preserve the diagnostic trigger, scenario, before/after logs and scope of equivalent counters. Update ground strike guidance, Windows acceptance status and profiler semantics.
Co-authored-by: Codex <codex@openai.com>
Show typed depot cargo separately, identify an empty nearby depot, and document loading, reconnaissance, two ID namespaces and reverse pickup.
Co-authored-by: Codex <codex@openai.com>
Keep the corrected historical coverage claim, archive the real flat-world WASM run, explain trap diagnostics and the small HD trail density change.
Co-authored-by: Codex <codex@openai.com>
Archive the original budget report as insufficient proof of overlapping raids. Pin callback/fuel diagnostics and an opt-in flat render world that services actual guest queries.
Co-authored-by: Codex <codex@openai.com>
chexit