Field Engineering Report · 2026-06-10 · The open-issues pass

Closing the Open Issues — four realism wins shipped, two reasoned closeouts, zero regressions

A single pass over every open issue in the backlog, run the way this repo demands: metricize before touching code, prove every win on held-out seeds, adversarially verify before calling it done — and ship only what the numbers earn. The fourth win, aspect-driven vegetation, was the hard one: it touches the cover field a firefight reads, so it took a three-strength balance sweep and a held-out A/B to find the strength (0.05) that improves the sim without regressing it. It ships on by default. The two remaining tickets already carry their active fix — the deployed mitigation is live; what's left is a documented terrain floor and a logged do-not, not unfinished work.

4realism wins shipped active
2adversarial workflows
6/6issues with a live fix
0regressions shipped

Shipped

019 · A squad switchbacks up a faceshipped

held-out detour ×3.58 → ×3.16 · korengal ×2.19 → ×1.31 · village routing & terrain byte-identical · stall guard passes

An any-angle (Theta*) tactical planner with a signed-grade cost + a turn penalty, fired only for a real climb and bolted onto the proven router without touching it. Two prior in-place attempts had been reverted for stalling a squad; this one is additive and gated, and a 4-agent adversarial pass caught + fixed a mover-fidelity gap (7→1 non-walkable legs). Full report →

020 · A soldier uses that rockshipped

cover behind a boulder 0.08 → 0.39 · flank 0.00 (directional) · US WIA −22% tuned / −36% held-out · enemy still killed via the flank

Directional, posture-aware cover: a boulder stops only the round from the bearing it faces, so the autonomous flank defeats it and firefights resolve by maneuver — the opposite of the reverted omnidirectional stamp's +89% grind. Adversarially verified (both lenses clean; fire-path only, ROE intact, determinism byte-identical). Full report →

021 · Supplies that biteshipped

dead batteries: 2.4× night detection lost · dehydration: 50% fatigue recovery · no medical: 54% wound recovery · 0 balance change at full stock

Issue 021 asked to combat-couple the COP fortification — but metricizing first proved the COP assault it was meant to stop never happens (a standoff from 340 m; the insurgents never close the wire). So the shippable win was the adjacent item: the four supplies that drained daily to zero consequence now cost you, as bounded clamps. Full report →

007 · Vegetation that reads the sunshipped active

forest-faces-north 59% → 64% · scrub-faces-south 48% → 56% · US KIA 1.17 → 1.08 (down, both seed sets) · 0 new strandings · village/COP siting & gate-overwatch byte-identical

Forest should cling to shaded slopes and scrub own the sunny ones, and an aspect term in classifyLand's moisture makes it so. The hard part: vegetation is cover and concealment, so moving forest onto the steep ambush faces reshapes the firefight — and the balance moves chaotically with strength (KIA isn't even monotonic; strandings flicker in and out). A three-strength A/B (0.16 stranded +56% KIA; 0.025 stranded; 0.05 clears the no-stall guard with KIA down) found the shippable strength, and a held-out hold-* A/B confirmed it (KIA 1.17→1.08, no aspect-caused stranding). Shipped on by default at 0.05, gated to steep faces (slope > 0.62, above the terrace band) so siting stays byte-identical. Honest caveat, disclosed: the wounded count is terrain-dependent (+39% tuned, −5% held-out) — the realistic, recoverable cost of fighting through forest vs open; ITM_ASPECT dials or disables it. Full report →

Active fix deployed — the reasoned closeouts

Both remaining tickets already run their live fix in the shipped game. What's left isn't unbuilt work — it's a documented terrain floor (009) and a logged do-not with a measured rationale (011), each kept open only to track the genuine remainder.

011 · The relief-bake load costfix deployed

Live fix: the 4096² relief bake runs progressively (bakeTerrainProgressive, 40 yielding row-bands) behind the deploy loading screen with a smooth progress bar, and the sheet is cached so the first frame is instant — the player's actual request (feedback during the wait) is delivered and shipped. The further optimization — lowering bake resolution — degrades the shaded relief the realism work added, backwards on a fidelity-first project, so it's logged as an explicit do-not; the clean win (seed-keyed cache or an OffscreenCanvas worker) needs a real-GPU-browser measurement first.

009 · The far-village residualfix deployed

Live fix: the reachability-aware spawn-snap + the patrol-corridor benching guard (the 006/008 lineage) is deployed — it's what lets a borderline far village's Track run ~1.3× faster and actually arrive. Re-measured after it, point-man arrival is 180/300 (60%) across the corpus; the unreached remainder are the genuinely cliff-walled cases (no gradeable route within ≈700 m) — the honest floor the valley itself sets, not a routing bug. Forcing 100% means gratuitous carving and the documented 008 negatives, so the residual is tracked, not papered over.

The thread that ties it together. The most valuable artifact in several of these six was a probe, not a feature: the one that proved the COP assault never happens (021), the three-strength A/B that found aspect's one shippable strength instead of guessing (007), the adversarial workflows that tried to break the two combat changes and couldn't (019, 020). On a project whose success metric is a skeptic asking "did an AI really build this?", the discipline to sweep a change across strengths until the numbers point to the one that ships — and to hold a perf ticket at its honest, already-deployed mitigation rather than fake a further win — is the part that's hardest to fake.

Six issues, one pass, six live fixes. Raw records: docs/progress/2026-06-10-open-issues/ (+ the per-issue resolutions in docs/issues/). Every number reproducible from a seed in seconds. Built by an autonomous agent fleet — four shipped realism wins (019, 020, 021, 007), two reasoned closeouts on a deployed fix (009, 011), zero regressions shipped.