#!/usr/bin/env bash
# apricot-run.sh — Isolated build + batch pipeline on apricot, sourced from forge.
#
# Source-of-truth flow: forge → apricot canonical checkout → per-run worktree.
# Plum is NOT in this loop. Agents push their work via ACS (commits-tray on plum
# pushes every ~5 min), then this script builds the latest origin/main on apricot.
#
#   1. Fetch latest origin/main into the canonical checkout
#      (~/Code/project-buildspace/magic-civilization, hard-fenced via pre-commit hook).
#   2. `git worktree add` a per-run scratch tree at ~/.cache/mc-src-<stamp>/
#      (shared object pool with canonical, isolated working tree, fast + reproducible).
#   3. Build (cargo) in the worktree, target dir stays there (ephemeral).
#   4. Run the batch with RESULTS_DIR under ~/.cache/mc-batches/<stamp>/
#      (persistent, XDG cache convention, flatpak-visible via --filesystem=home).
#   5. Fetch verdict JSON back to EDIT host for review.
#   6. Remove the worktree (canonical + objects retained for next run).
#
# ── Synchronous usage (block until done, fetch results inline) ──────────────
#   scripts/apricot-run.sh smoke [seeds=10] [turns=300]
#   scripts/apricot-run.sh clan <clan_id> [seeds=10] [turns=300]
#   scripts/apricot-run.sh gpu-walltime [seeds=10] [turns=300]
#
# ── Async protocol (p2-64): launch / status / fetch ─────────────────────────
# Decouples job lifecycle from the orchestrating ssh. The systemd --user unit
# on apricot owns build+batch and survives ssh disconnect, sleep/wake, network
# blips. EDIT host polls via short ConnectTimeout=5 probes.
#
#   STAMP=$(scripts/apricot-run.sh launch smoke 1 50)         # bare stdout = stamp
#   scripts/apricot-run.sh status "$STAMP"                    # one-line JSON
#   # → {"stamp":"...","state":"running|complete|failed|unreachable",
#   #    "seeds_done":N,"seeds_total":M,"completion_marker":bool}
#   scripts/apricot-run.sh fetch  "$STAMP"                    # rsync results
#
# Loop pattern:
#   STAMP=$(scripts/apricot-run.sh launch smoke 10 300)
#   while STATE=$(scripts/apricot-run.sh status "$STAMP" | jq -r .state); \
#         [[ $STATE != complete ]]; do
#       [[ $STATE == failed ]] && exit 1
#       sleep 60
#   done
#   scripts/apricot-run.sh fetch "$STAMP"
#
# Implementation notes:
#   • launch writes a per-stamp launcher.sh into ~/.cache/mc-batches/<stamp>/
#     and starts it under `systemd-run --user --collect --unit=mc-batch-<stamp>`.
#   • The launcher does git fetch, worktree add, build-gdext, autoplay-batch,
#     then `touch <submode>/completion.marker` ONLY on success.
#   • status uses a single ssh ConnectTimeout=5 with three lightweight probes
#     (systemctl is-active, marker count, turn_stats count) — no piped reads
#     of file contents (we hit channel saturation with that historically).
#   • fetch is rsync -a --partial; resumable across drops.
#
# Environment:
#   APRICOT_SSH_ALIAS — ssh alias for the RUN host (default: apricot).
#   STAMP             — override the timestamp (for reproducing a specific run).
#   BUILD_REF         — git ref to build (default: origin/main). Lets you reproduce
#                       a prior run by SHA without changing your local plum tree.

set -euo pipefail

SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
PROJECT_DIR="$(dirname "$SCRIPT_DIR")"

APRICOT="${APRICOT_SSH_ALIAS:-apricot}"
STAMP="${STAMP:-$(date +%Y%m%d_%H%M%S)}"

# ── Load .env / .env.local so USE_MAX_CORES + MIN_CORES + AI_GPU_ROLLOUT
# propagate into the resource policy below. .env.local wins over .env.
for envfile in "${PROJECT_DIR}/.env" "${PROJECT_DIR}/.env.local"; do
    if [[ -f "${envfile}" ]]; then
        set -a; source "${envfile}"; set +a
    fi
done

# MODE + positional args resolved early so the resource-policy block can
# peek at the seed count (which differs per mode — for `clan` it's $2
# because $1 is the clan_id; for smoke/gpu-walltime it's $1).
MODE="${1:?usage: apricot-run.sh <smoke|clan|difficulty|difficulty-asym|gpu-walltime|launch|status|fetch> [args]}"
shift || true

# ── p2-64 async protocol: launch / status / fetch ────────────────────────────
# These three sub-modes decouple the batch lifecycle from the orchestrating ssh.
# See the header comment for the full protocol shape and example loop.

if [[ "${MODE}" == "launch" ]]; then
    SUBMODE="${1:?usage: apricot-run.sh launch <smoke|clan|difficulty|...> [args]}"
    shift || true
    LAUNCH_ARGS=("$@")

    # Pre-resolve the seed count from the sub-mode args (mirrors the resource
    # policy peek below). The launcher script on apricot will use this to write
    # a seeds_total file before invoking autoplay-batch.
    case "${SUBMODE}" in
        clan|clan-priors|difficulty)   SEEDS_TOTAL_PEEK="${LAUNCH_ARGS[1]:-10}" ;;
        difficulty-asym)               SEEDS_TOTAL_PEEK="${LAUNCH_ARGS[2]:-10}" ;;
        matchup-grid|huge-map-5clan)   SEEDS_TOTAL_PEEK="${LAUNCH_ARGS[0]:-5}" ;;
        ai-quality-baseline*)          SEEDS_TOTAL_PEEK="${LAUNCH_ARGS[0]:-50}" ;;
        smoke|gpu-walltime|*)          SEEDS_TOTAL_PEEK="${LAUNCH_ARGS[0]:-10}" ;;
    esac

    # Map sub-mode → submode results dir (mirrors the case statement below).
    case "${SUBMODE}" in
        clan)            SUBDIR="clan-${LAUNCH_ARGS[0]:-unknown}" ;;
        clan-priors)     SUBDIR="clan-priors-${LAUNCH_ARGS[0]:-unknown}" ;;
        difficulty)      SUBDIR="difficulty-${LAUNCH_ARGS[0]:-unknown}" ;;
        difficulty-asym) SUBDIR="difficulty-asym-${LAUNCH_ARGS[0]:-unknown}-vs-${LAUNCH_ARGS[1]:-unknown}" ;;
        matchup-grid)    SUBDIR="matchup-grid" ;;
        huge-map-5clan)  SUBDIR="huge-map-5clan" ;;
        ai-quality-baseline*)  SUBDIR="baseline" ;;  # tier subdirs underneath
        gpu-walltime)    SUBDIR="gpu-walltime" ;;
        smoke|*)         SUBDIR="smoke" ;;
    esac

    # Build a properly-quoted args string for embedding in the launcher script.
    ARGS_QUOTED=""
    for a in "${LAUNCH_ARGS[@]}"; do
        ARGS_QUOTED+=" $(printf '%q' "$a")"
    done

    # All status/log output goes to stderr — stdout is reserved for the bare
    # stamp value so callers can do STAMP=$(scripts/apricot-run.sh launch ...).
    {
        echo "[launch] stamp=${STAMP} submode=${SUBMODE} args=${LAUNCH_ARGS[*]}"
        echo "[launch] writing launcher.sh to apricot:~/.cache/mc-batches/${STAMP}/"
    } >&2

    # Write a per-stamp launcher script. Heredoc keeps quoting sane; we splice
    # in only the values we actually need ($STAMP, $SUBMODE, $ARGS_QUOTED, $SUBDIR,
    # $BUILD_REF, $SEEDS_TOTAL_PEEK).
    BUILD_REF_LAUNCH="${BUILD_REF:-origin/main}"
    ssh "${APRICOT}" "mkdir -p \"\$HOME/.cache/mc-batches/${STAMP}/${SUBDIR}\" && cat > \"\$HOME/.cache/mc-batches/${STAMP}/launcher.sh\"" <<LAUNCHER
#!/usr/bin/env bash
# Auto-generated by scripts/apricot-run.sh launch on $(date -u +%Y-%m-%dT%H:%M:%SZ)
# Owned by systemd unit mc-batch-${STAMP}; survives ssh disconnects.
set -euo pipefail

STAMP="${STAMP}"
SUBMODE="${SUBMODE}"
SUBDIR="${SUBDIR}"
BUILD_REF="${BUILD_REF_LAUNCH}"
SEEDS_TOTAL=${SEEDS_TOTAL_PEEK}

CANONICAL="\$HOME/Code/project-buildspace/magic-civilization"
SCRATCH="\$HOME/.cache/mc-src-\${STAMP}"
RESULTS="\$HOME/.cache/mc-batches/\${STAMP}"
RESULTS_SUB="\${RESULTS}/\${SUBDIR}"

mkdir -p "\${RESULTS_SUB}"
echo "\${SEEDS_TOTAL}" > "\${RESULTS_SUB}/seeds_total"

# Launch-side log lives next to results so post-mortem doesn't need journalctl.
LOG="\${RESULTS}/launcher.log"
exec >>"\${LOG}" 2>&1
echo "===== mc-batch-\${STAMP} launcher start \$(date -u +%FT%TZ) ====="
echo "submode=\${SUBMODE} subdir=\${SUBDIR} build_ref=\${BUILD_REF} seeds_total=\${SEEDS_TOTAL}"

# ── Resource policy (formerly EDIT-side; moved here so async re-launches from
# any orchestrator host produce the same answer). ─────────────────────────────
NPROC="\$(nproc 2>/dev/null || echo 8)"
if [[ -n "\${PARALLEL:-}" ]]; then
    PARALLEL_EFFECTIVE="\${PARALLEL}"
elif [[ "\${USE_MAX_CORES:-true}" == "true" ]]; then
    PARALLEL_EFFECTIVE="\$(( SEEDS_TOTAL < NPROC ? SEEDS_TOTAL : NPROC ))"
else
    PARALLEL_EFFECTIVE="\${MIN_CORES:-4}"
fi
[[ "\${PARALLEL_EFFECTIVE}" -lt 1 ]] && PARALLEL_EFFECTIVE=1
export PARALLEL="\${PARALLEL_EFFECTIVE}"

if [[ -z "\${RAYON_NUM_THREADS:-}" ]]; then
    RAYON_NUM_THREADS="\$(( NPROC / PARALLEL_EFFECTIVE ))"
    [[ "\${RAYON_NUM_THREADS}" -lt 1 ]] && RAYON_NUM_THREADS=1
fi
export RAYON_NUM_THREADS

echo "PARALLEL=\${PARALLEL} RAYON_NUM_THREADS=\${RAYON_NUM_THREADS} NPROC=\${NPROC}"

# ── Step 1: fetch + worktree ─────────────────────────────────────────────────
test -d "\${CANONICAL}/.git" || {
    echo "ERROR: canonical checkout missing at \${CANONICAL}" >&2
    exit 1
}
git -C "\${CANONICAL}" fetch origin --quiet
git -C "\${CANONICAL}" worktree add --detach "\${SCRATCH}" "\${BUILD_REF}"
BUILT_SHA="\$(git -C "\${SCRATCH}" rev-parse --short HEAD)"
echo "built_sha=\${BUILT_SHA}"

# ── Step 2: build ────────────────────────────────────────────────────────────
( cd "\${SCRATCH}/src/simulator" && bash build-gdext.sh x86_64-unknown-linux-gnu )
rm -f "\${SCRATCH}/src/game/engine/addons/magic_civ_physics/libmagic_civ_physics.dylib"

# ── Step 3: editor pre-pass to populate .godot/ class cache ─────────────────
flatpak run --user --filesystem=home --command=godot \\
    org.godotengine.Godot --headless --editor --quit \\
    --path "\${SCRATCH}/src/game" 2>&1 | tail -5 || true

# ── Step 4: run the batch ────────────────────────────────────────────────────
GPU_ENV_VAL="\${AI_GPU_ROLLOUT:-false}"
cd "\${SCRATCH}"

case "\${SUBMODE}" in
    smoke)
        AI_USE_MCTS=true AI_GPU_ROLLOUT="\${GPU_ENV_VAL}" PARALLEL="\${PARALLEL}" \\
            bash tools/autoplay-batch.sh "\$@" "\${RESULTS_SUB}"
        ;;
    clan)
        CLAN="\$1"; shift
        AI_USE_MCTS=true AI_PIN_PERSONALITY="\${CLAN}" \\
            AI_GPU_ROLLOUT="\${GPU_ENV_VAL}" PARALLEL="\${PARALLEL}" \\
            bash tools/autoplay-batch.sh "\$@" "\${RESULTS_SUB}"
        ;;
    difficulty)
        TIER="\$1"; shift
        AI_USE_MCTS=true AI_DIFFICULTY="\${TIER}" \\
            AI_DIFFICULTY_P0="\${TIER}" AI_DIFFICULTY_P1="\${TIER}" \\
            AI_GPU_ROLLOUT="\${GPU_ENV_VAL}" PARALLEL="\${PARALLEL}" \\
            bash tools/autoplay-batch.sh "\$@" "\${RESULTS_SUB}"
        ;;
    huge-map-5clan)
        COUNT="\${1:-5}"; TURN_LIMIT="\${2:-300}"
        mkdir -p "\${RESULTS_SUB}"
        AI_USE_MCTS=true PARALLEL="\${PARALLEL}" RAYON_NUM_THREADS="\${RAYON_NUM_THREADS}" \\
            COUNT="\${COUNT}" TURN_LIMIT="\${TURN_LIMIT}" \\
            HUGE_OUTPUT="\${RESULTS_SUB}" \\
            bash tools/huge-map-5clan.sh
        ;;
    *)
        echo "ERROR: launcher does not yet support submode '\${SUBMODE}'" >&2
        exit 2
        ;;
esac

# Only on success path: write completion.marker. Status's "failed" state =
# unit inactive AND no marker.
touch "\${RESULTS_SUB}/completion.marker"
echo "===== mc-batch-\${STAMP} launcher OK \$(date -u +%FT%TZ) ====="

# Cleanup worktree (canonical + objects retained for next run).
git -C "\${CANONICAL}" worktree remove --force "\${SCRATCH}" 2>&1 || rm -rf "\${SCRATCH}"
LAUNCHER

    # The launcher reads args from "$@" inside its case branches; we pass
    # them through the systemd-run invocation below.
    ssh "${APRICOT}" "chmod +x \"\$HOME/.cache/mc-batches/${STAMP}/launcher.sh\""

    # Start the unit. Pass the full LAUNCH_ARGS as positional args to the
    # launcher script (they show up as $1, $2, ... inside the case branches).
    SYSTEMD_CMD="systemd-run --user --collect --unit=mc-batch-${STAMP} \"\$HOME/.cache/mc-batches/${STAMP}/launcher.sh\"${ARGS_QUOTED}"
    echo "[launch] starting systemd unit mc-batch-${STAMP}" >&2
    if ! ssh "${APRICOT}" "${SYSTEMD_CMD}" >&2; then
        echo "[launch] FAILED to start systemd unit; check ssh + systemd --user availability" >&2
        exit 1
    fi

    echo "[launch] unit started; tail logs via: ssh ${APRICOT} 'journalctl --user -u mc-batch-${STAMP} -f'" >&2

    # Bare stamp on stdout — this is the contract for callers.
    echo "${STAMP}"
    exit 0
fi

if [[ "${MODE}" == "status" ]]; then
    QUERY_STAMP="${1:?usage: apricot-run.sh status <stamp>}"
    UNIT="mc-batch-${QUERY_STAMP}"

    # Single ssh probe with short ConnectTimeout. Three lightweight queries:
    #   1. systemctl --user is-active <unit>  (active|inactive|failed|unknown)
    #   2. count of completion.marker files under <stamp>/*/
    #   3. count of turn_stats.jsonl files under <stamp>/*/game_*/
    # We also read seeds_total from the first submode dir if present.
    PROBE='set +e
        IS_ACTIVE=$(systemctl --user is-active '"${UNIT}"' 2>/dev/null || echo unknown)
        MARKER_COUNT=$(ls "$HOME/.cache/mc-batches/'"${QUERY_STAMP}"'"/*/completion.marker 2>/dev/null | wc -l | tr -d " ")
        STATS_COUNT=$(ls "$HOME/.cache/mc-batches/'"${QUERY_STAMP}"'"/*/game_*/turn_stats.jsonl 2>/dev/null | wc -l | tr -d " ")
        SEEDS_TOTAL=$(cat "$HOME/.cache/mc-batches/'"${QUERY_STAMP}"'"/*/seeds_total 2>/dev/null | head -1)
        SEEDS_TOTAL=${SEEDS_TOTAL:-0}
        printf "%s|%s|%s|%s\n" "$IS_ACTIVE" "$MARKER_COUNT" "$STATS_COUNT" "$SEEDS_TOTAL"'

    PROBE_OUT="$(ssh -o ConnectTimeout=5 -o BatchMode=yes "${APRICOT}" "${PROBE}" 2>/dev/null)" || PROBE_OUT=""

    if [[ -z "${PROBE_OUT}" ]]; then
        printf '{"stamp":"%s","state":"unreachable","seeds_done":0,"seeds_total":0,"completion_marker":false}\n' "${QUERY_STAMP}"
        exit 0
    fi

    IFS='|' read -r IS_ACTIVE MARKER_COUNT STATS_COUNT SEEDS_TOTAL <<<"${PROBE_OUT}"
    MARKER_COUNT="${MARKER_COUNT:-0}"
    STATS_COUNT="${STATS_COUNT:-0}"
    SEEDS_TOTAL="${SEEDS_TOTAL:-0}"

    if [[ "${MARKER_COUNT}" -gt 0 ]]; then
        STATE="complete"
        MARKER_BOOL="true"
    elif [[ "${IS_ACTIVE}" == "active" || "${IS_ACTIVE}" == "activating" ]]; then
        STATE="running"
        MARKER_BOOL="false"
    else
        STATE="failed"
        MARKER_BOOL="false"
    fi

    printf '{"stamp":"%s","state":"%s","seeds_done":%s,"seeds_total":%s,"completion_marker":%s}\n' \
        "${QUERY_STAMP}" "${STATE}" "${STATS_COUNT}" "${SEEDS_TOTAL}" "${MARKER_BOOL}"
    exit 0
fi

if [[ "${MODE}" == "fetch" ]]; then
    QUERY_STAMP="${1:?usage: apricot-run.sh fetch <stamp>}"
    LOCAL_DEST="${PROJECT_DIR}/.local/iter/${QUERY_STAMP}"

    # Re-use status to gate the fetch — this is the same one-line probe.
    STATUS_JSON="$("$0" status "${QUERY_STAMP}")"
    STATE="$(echo "${STATUS_JSON}" | sed -n 's/.*"state":"\([^"]*\)".*/\1/p')"

    if [[ "${STATE}" != "complete" ]]; then
        echo "[fetch] batch not complete; status=${STATE}" >&2
        echo "[fetch] full status: ${STATUS_JSON}" >&2
        exit 1
    fi

    mkdir -p "${LOCAL_DEST}"
    echo "[fetch] rsync apricot:~/.cache/mc-batches/${QUERY_STAMP}/ → ${LOCAL_DEST}/" >&2
    rsync -a --partial \
        "${APRICOT}:.cache/mc-batches/${QUERY_STAMP}/" \
        "${LOCAL_DEST}/"

    echo "${LOCAL_DEST}"
    exit 0
fi
# ── end p2-64 async protocol ─────────────────────────────────────────────────

# ── Resource policy for PARALLEL + RAYON_NUM_THREADS ─────────────────
# Each Godot instance spawns its own rayon thread pool for MCTS rollouts;
# rayon defaults to nproc unless RAYON_NUM_THREADS is set. If PARALLEL
# instances each claim all nproc threads, we get PARALLEL*nproc threads
# fighting over nproc cores → thrashing, each process effectively single
# core. Better: PARALLEL = number of seeds (one instance each), and
# RAYON_NUM_THREADS = nproc / PARALLEL so the box is saturated evenly.
case "${MODE}" in
    clan)           _seed_count_peek="${2:-10}" ;;   # $1 is clan_id, $2 is seeds
    clan-priors)    _seed_count_peek="${2:-10}" ;;   # $1 is clan_id, $2 is seeds
    difficulty)     _seed_count_peek="${2:-10}" ;;   # $1 is tier, $2 is seeds
    difficulty-asym) _seed_count_peek="${3:-10}" ;;  # $1 p0 tier, $2 p1 tier, $3 seeds
    matchup-grid)   _seed_count_peek="${1:-5}" ;;    # $1 is seeds_per_pair (default 5); total=10pairs*seeds
    huge-map-5clan) _seed_count_peek="${1:-5}" ;;    # $1 is seeds
    ai-quality-baseline*) _seed_count_peek="${1:-50}" ;; # $1 seeds-per-tier
    *)              _seed_count_peek="${1:-10}" ;;   # smoke, gpu-walltime
esac

NPROC="$(ssh "${APRICOT}" nproc 2>/dev/null || echo 8)"

if [[ -n "${PARALLEL:-}" ]]; then
    PARALLEL_EFFECTIVE="${PARALLEL}"
    PARALLEL_SOURCE="env override"
elif [[ "${USE_MAX_CORES:-false}" == "true" ]]; then
    # One instance per seed — up to NPROC. More instances than that
    # would queue serially anyway (NPROC concurrent Godots max).
    PARALLEL_EFFECTIVE="$(( _seed_count_peek < NPROC ? _seed_count_peek : NPROC ))"
    PARALLEL_SOURCE="USE_MAX_CORES=true → min(seeds=${_seed_count_peek}, nproc=${NPROC})"
else
    PARALLEL_EFFECTIVE="${MIN_CORES:-4}"
    PARALLEL_SOURCE="MIN_CORES default"
fi
export PARALLEL="${PARALLEL_EFFECTIVE}"

# RAYON_NUM_THREADS per Godot instance = fair share of cores.
if [[ -n "${RAYON_NUM_THREADS:-}" ]]; then
    RAYON_SOURCE="env override"
else
    if [[ "${PARALLEL_EFFECTIVE}" -gt 0 ]]; then
        RAYON_NUM_THREADS="$(( NPROC / PARALLEL_EFFECTIVE ))"
    else
        RAYON_NUM_THREADS=1
    fi
    [[ "${RAYON_NUM_THREADS}" -lt 1 ]] && RAYON_NUM_THREADS=1
    RAYON_SOURCE="nproc(${NPROC}) / PARALLEL(${PARALLEL_EFFECTIVE})"
fi
export RAYON_NUM_THREADS
# Source + build scratch lives under $HOME/.cache (flatpak-visible via
# --filesystem=home). /tmp was tried first but flatpak's sandbox can't see
# /tmp, so Godot rejected the --path argument with "Invalid project path".
# $HOME/.cache/ also satisfies the apricot-isolation rule (not under ~/Code/@projects,
# not under project-buildspace where the canonical checkout lives) and is
# convention-cleanable.
SCRATCH="\$HOME/.cache/mc-src-${STAMP}"         # expanded on apricot
RESULTS="\$HOME/.cache/mc-batches/${STAMP}"     # expanded on apricot

# Resolve $HOME on apricot so SCRATCH / RESULTS are fully-qualified paths on that host.
SCRATCH_ABS="$(ssh "${APRICOT}" "echo \$HOME/.cache/mc-src-${STAMP}")"
RESULTS_ABS="$(ssh "${APRICOT}" "echo \$HOME/.cache/mc-batches/${STAMP}")"

# Canonical checkout: persistent clone of magicciv from forge, lives in the
# autocommit-excluded buildspace dir. Worktrees branch off this for per-run
# scratch trees with shared object pool — fast + reproducible.
CANONICAL_ABS="$(ssh "${APRICOT}" "echo \$HOME/Code/project-buildspace/magic-civilization")"
BUILD_REF="${BUILD_REF:-origin/main}"

echo "============================================================"
echo "apricot-run.sh mode=${MODE} stamp=${STAMP}"
echo "  EDIT host: $(hostname)"
echo "  RUN host:  ${APRICOT}"
echo "  CANONICAL: ${CANONICAL_ABS}    (persistent clone of forge)"
echo "  BUILD_REF: ${BUILD_REF}"
echo "  SCRATCH:   ${SCRATCH_ABS}       (per-run worktree, ephemeral)"
echo "  RESULTS:   ${RESULTS_ABS}       (persistent batch output)"
echo "  PARALLEL:  ${PARALLEL_EFFECTIVE}   (source: ${PARALLEL_SOURCE})"
echo "  RAYON_NUM_THREADS/instance: ${RAYON_NUM_THREADS}   (source: ${RAYON_SOURCE})"
echo "  Total CPU saturation: ${PARALLEL_EFFECTIVE} × ${RAYON_NUM_THREADS} = $((PARALLEL_EFFECTIVE * RAYON_NUM_THREADS))/${NPROC} cores"
echo "  AI_GPU_ROLLOUT: ${AI_GPU_ROLLOUT:-true (default on for smoke/clan)}"
echo "============================================================"

# ── Step 1: fetch forge → canonical → per-run worktree ───────────────────────
# Source flow is forge → canonical → worktree. Plum is NOT in this path; agents
# push via ACS (commits-tray on plum pushes every ~5 min). To test a specific
# SHA, set BUILD_REF=<sha> in the environment.
echo "[$(date +%H:%M:%S)] fetch origin in canonical, then worktree ${BUILD_REF} → ${SCRATCH_ABS}..."
ssh "${APRICOT}" "set -euo pipefail; \
    test -d '${CANONICAL_ABS}/.git' || { \
        echo 'ERROR: canonical checkout missing at ${CANONICAL_ABS}' >&2; \
        echo 'one-time setup: git clone http://forge.black.local:3000/magicciv/magicciv.git ${CANONICAL_ABS}' >&2; \
        echo 'then: git -C ${CANONICAL_ABS} config core.hooksPath \$HOME/Code/project-buildspace/.hooks' >&2; \
        exit 1; \
    }; \
    git -C '${CANONICAL_ABS}' fetch origin --quiet; \
    git -C '${CANONICAL_ABS}' worktree add --detach '${SCRATCH_ABS}' '${BUILD_REF}'"

# Resolve the actual SHA we built (BUILD_REF may be a branch name).
BUILT_SHA="$(ssh "${APRICOT}" "git -C '${SCRATCH_ABS}' rev-parse --short HEAD")"
echo "  built SHA: ${BUILT_SHA}"

# ── Step 2: build + deploy via build-gdext.sh ────────────────────────────────
# Canonical build script: runs `cargo build --release --target x86_64-unknown-linux-gnu`
# AND copies the output from .local/build/rust/$TARGET/release/libmagic_civ_physics_gdext.so
# into src/game/engine/addons/magic_civ_physics/libmagic_civ_physics.x86_64.so
# with the name Godot's .gdextension file expects.
echo "[$(date +%H:%M:%S)] build-gdext.sh x86_64-unknown-linux-gnu (in SCRATCH)..."
# Never pipe build output through `| tail -N` — ssh inherits the pipe's
# exit code (0 from tail), which masks build-gdext.sh failures. That's
# how a stale April-16 .so ran for 2 hours on 2026-04-17 while bullets
# 3-5 of p0-32 stayed ✗ with no visible warning. Let the full log through.
ssh "${APRICOT}" "set -euo pipefail; cd '${SCRATCH_ABS}/src/simulator' && bash build-gdext.sh x86_64-unknown-linux-gnu"

# Also purge the stale .dylib that the rsync carried from the Mac — flatpak Godot
# on Linux shouldn't try to load a macOS binary even though the .gdextension config
# only points at it for macos.* target strings.
ssh "${APRICOT}" "rm -f '${SCRATCH_ABS}/src/game/engine/addons/magic_civ_physics/libmagic_civ_physics.dylib'"

# ── Step 3: populate .godot/global_script_class_cache.cfg via editor pre-pass ──
# Fresh scratch tree has an empty .godot/ cache. Godot's class_name resolution
# (the `as Weather` / `as SplitPanelContainer` idiom) requires this cache to be
# populated by an editor-mode scan. Without it, headless autoplay cascades through
# "Could not find type X" → "Compilation failed" → turn_manager.gd fails to load.
#
# `--editor --quit` opens, imports, scans class_names, writes cache, exits.
# Expected exit code: 0. We grep the log to verify the cache got written.
echo "[$(date +%H:%M:%S)] editor pre-pass to populate .godot/ class cache..."
ssh "${APRICOT}" "set -euo pipefail; \
    flatpak run --user --filesystem=home --command=godot \
        org.godotengine.Godot --headless --editor --quit \
        --path '${SCRATCH_ABS}/src/game' 2>&1 | tail -5; \
    test -s '${SCRATCH_ABS}/src/game/.godot/global_script_class_cache.cfg' && \
        echo '  ✓ class cache populated' || \
        echo '  ⚠ class cache missing — headless autoplay may still cascade'"

# ── Step 4: run the batch per MODE ───────────────────────────────────────────
ssh "${APRICOT}" "mkdir -p ${RESULTS_ABS}"

case "${MODE}" in
    smoke)
        SEEDS="${1:-10}"; TURNS="${2:-300}"
        # Default: use the GPU when available (MCTS rollouts through WGSL kernel).
        # gpu-walltime mode overrides this explicitly to true/false per iteration.
        # Default AI_GPU_ROLLOUT=false for smoke/clan. The GPU integration
        # (p0-20 task #10) is parity-verified on isolated rollouts, but
        # enabling it in a 2-player smoke produced a deterministic
        # "P0 always wins at T11-T18, P1 never founds" regression on
        # 2026-04-18. Opt-in via env override; gpu-walltime flips
        # per-iteration as its explicit comparison.
        GPU_ENV="AI_GPU_ROLLOUT=${AI_GPU_ROLLOUT:-false}"
        # p2-44b: forward MC_AI_* diagnostic envs so instrumentation prints surface in game.log
        MC_AI_ENV=""
        for var in MC_AI_PROMOTION_DEBUG; do
            val="${!var:-}"
            if [[ -n "${val}" ]]; then
                MC_AI_ENV="${MC_AI_ENV} ${var}=${val}"
            fi
        done
        echo "[$(date +%H:%M:%S)] smoke batch: ${SEEDS} seeds T${TURNS} PARALLEL=${PARALLEL} ${GPU_ENV}${MC_AI_ENV}"
        ssh "${APRICOT}" "set -euo pipefail; cd '${SCRATCH_ABS}' && \
            AI_USE_MCTS=true ${GPU_ENV}${MC_AI_ENV} PARALLEL=${PARALLEL} \
            bash tools/autoplay-batch.sh ${SEEDS} ${TURNS} ${RESULTS_ABS}/smoke 2>&1 | tail -30"
        ;;
    clan)
        CLAN="${1:?usage: apricot-run.sh clan <clan_id> [seeds] [turns]}"
        SEEDS="${2:-10}"; TURNS="${3:-300}"
        # Default AI_GPU_ROLLOUT=false for smoke/clan. The GPU integration
        # (p0-20 task #10) is parity-verified on isolated rollouts, but
        # enabling it in a 2-player smoke produced a deterministic
        # "P0 always wins at T11-T18, P1 never founds" regression on
        # 2026-04-18. Opt-in via env override; gpu-walltime flips
        # per-iteration as its explicit comparison.
        GPU_ENV="AI_GPU_ROLLOUT=${AI_GPU_ROLLOUT:-false}"
        echo "[$(date +%H:%M:%S)] clan=${CLAN} batch: ${SEEDS} seeds T${TURNS} PARALLEL=${PARALLEL} ${GPU_ENV}"
        ssh "${APRICOT}" "set -euo pipefail; cd '${SCRATCH_ABS}' && \
            AI_USE_MCTS=true AI_PIN_PERSONALITY='${CLAN}' ${GPU_ENV} PARALLEL=${PARALLEL} \
            bash tools/autoplay-batch.sh ${SEEDS} ${TURNS} ${RESULTS_ABS}/clan-${CLAN} 2>&1 | tail -30"
        ;;
    difficulty)
        DIFF_TIER="${1:?usage: apricot-run.sh difficulty <easy|normal|hard|insane> [seeds] [turns]}"
        SEEDS="${2:-10}"; TURNS="${3:-300}"
        GPU_ENV="AI_GPU_ROLLOUT=${AI_GPU_ROLLOUT:-false}"
        echo "[$(date +%H:%M:%S)] difficulty=${DIFF_TIER} batch: ${SEEDS} seeds T${TURNS} PARALLEL=${PARALLEL} ${GPU_ENV}"
        # AI_DIFFICULTY_P0 + AI_DIFFICULTY_P1 apply the modifier to BOTH players
        # (including the human-slot player 0 which is_human=true). This is
        # required for symmetric Easy-vs-Easy / Hard-vs-Hard tier_peak differentiation.
        ssh "${APRICOT}" "set -euo pipefail; cd '${SCRATCH_ABS}' && \
            AI_USE_MCTS=true AI_DIFFICULTY='${DIFF_TIER}' \
            AI_DIFFICULTY_P0='${DIFF_TIER}' AI_DIFFICULTY_P1='${DIFF_TIER}' \
            ${GPU_ENV} PARALLEL=${PARALLEL} \
            bash tools/autoplay-batch.sh ${SEEDS} ${TURNS} ${RESULTS_ABS}/difficulty-${DIFF_TIER} 2>&1 | tail -30"
        ;;
    difficulty-asym)
        P0_TIER="${1:?usage: apricot-run.sh difficulty-asym <p0-tier> <p1-tier> [seeds] [turns]}"
        P1_TIER="${2:?usage: apricot-run.sh difficulty-asym <p0-tier> <p1-tier> [seeds] [turns]}"
        SEEDS="${3:-10}"; TURNS="${4:-300}"
        GPU_ENV="AI_GPU_ROLLOUT=${AI_GPU_ROLLOUT:-false}"
        echo "[$(date +%H:%M:%S)] difficulty-asym p0=${P0_TIER} p1=${P1_TIER}: ${SEEDS} seeds T${TURNS}"
        ssh "${APRICOT}" "set -euo pipefail; cd '${SCRATCH_ABS}' && \
            AI_USE_MCTS=true AI_DIFFICULTY_P0='${P0_TIER}' AI_DIFFICULTY_P1='${P1_TIER}' \
            ${GPU_ENV} PARALLEL=${PARALLEL} \
            bash tools/autoplay-batch.sh ${SEEDS} ${TURNS} \
            ${RESULTS_ABS}/difficulty-asym-${P0_TIER}-vs-${P1_TIER} 2>&1 | tail -30"
        ;;
    clan-priors)
        # Like `clan` but with AI_MCTS_PRIORS=true — used to verify p0-38
        # tree-shape divergence across 5 clan personalities.
        CLAN="${1:?usage: apricot-run.sh clan-priors <clan_id> [seeds] [turns]}"
        SEEDS="${2:-10}"; TURNS="${3:-300}"
        GPU_ENV="AI_GPU_ROLLOUT=${AI_GPU_ROLLOUT:-false}"
        echo "[$(date +%H:%M:%S)] clan-priors=${CLAN} batch: ${SEEDS} seeds T${TURNS} PARALLEL=${PARALLEL} priors=true"
        ssh "${APRICOT}" "set -euo pipefail; cd '${SCRATCH_ABS}' && \
            AI_USE_MCTS=true AI_PIN_PERSONALITY='${CLAN}' AI_MCTS_PRIORS=true ${GPU_ENV} PARALLEL=${PARALLEL} \
            bash tools/autoplay-batch.sh ${SEEDS} ${TURNS} ${RESULTS_ABS}/clan-priors-${CLAN} 2>&1 | tail -30"
        ;;
    gpu-walltime)
        SEEDS="${1:-10}"; TURNS="${2:-300}"
        echo "[$(date +%H:%M:%S)] GPU wall-time comparison: ${SEEDS} seeds T${TURNS}"
        for GPU in true false; do
            echo "  --- AI_GPU_ROLLOUT=${GPU} ---"
            ssh "${APRICOT}" "set -euo pipefail; cd '${SCRATCH_ABS}' && \
                AI_USE_MCTS=true AI_GPU_ROLLOUT=${GPU} PARALLEL=${PARALLEL} \
                bash tools/autoplay-batch.sh ${SEEDS} ${TURNS} ${RESULTS_ABS}/gpu-${GPU} 2>&1 | tail -10"
        done
        ;;
    matchup-grid)
        # Run all C(5,2)=10 clan-pair matchups serially (pairs run one at a time;
        # seeds within a pair use PARALLEL concurrency). Uses the scratch-resident
        # binary so we never touch ~/Code on the RUN host.
        SEEDS_PER_PAIR="${1:-5}"; TURNS="${2:-300}"
        REMOTE_GRID="${RESULTS_ABS}/matchup-grid"
        echo "[$(date +%H:%M:%S)] matchup-grid: ${SEEDS_PER_PAIR} seeds/pair T${TURNS} PARALLEL=${PARALLEL}"
        ssh "${APRICOT}" "set -euo pipefail; mkdir -p '${REMOTE_GRID}'; cd '${SCRATCH_ABS}' && \
            AI_USE_MCTS=true PARALLEL=${PARALLEL} RAYON_NUM_THREADS=${RAYON_NUM_THREADS} \
            COUNT=${SEEDS_PER_PAIR} TURN_LIMIT=${TURNS} \
            MATCHUP_OUTPUT='${REMOTE_GRID}' \
            bash tools/matchup-grid.sh 2>&1 | tail -40"
        ;;
    huge-map-5clan)
        SEEDS="${1:-5}"; TURNS="${2:-300}"
        REMOTE_HUGE="${RESULTS_ABS}/huge-map-5clan"
        echo "[$(date +%H:%M:%S)] huge-map-5clan: ${SEEDS} seeds T${TURNS} PARALLEL=${PARALLEL}"
        ssh "${APRICOT}" "set -euo pipefail; mkdir -p '${REMOTE_HUGE}'; cd '${SCRATCH_ABS}' && \
            AI_USE_MCTS=true PARALLEL=${PARALLEL} RAYON_NUM_THREADS=${RAYON_NUM_THREADS} \
            COUNT=${SEEDS} TURN_LIMIT=${TURNS} \
            HUGE_OUTPUT='${REMOTE_HUGE}' \
            bash tools/huge-map-5clan.sh 2>&1 | tail -40"
        ;;
    ai-quality-baseline|ai-quality-baseline-pre-c|ai-quality-baseline-post-c)
        # p0-20 Phase C — empirical AI-quality baseline.
        # Runs `difficulty normal` + `difficulty hard` back-to-back so
        # pre-C / post-C comparison covers SEEDS×2 games per build with one
        # build+worktree setup. Trailing `<tier1> <tier2>` are optional and
        # default to `normal hard`.
        SEEDS="${1:-50}"; TURNS="${2:-300}"
        TIER1="${3:-normal}"; TIER2="${4:-hard}"
        GPU_ENV="AI_GPU_ROLLOUT=${AI_GPU_ROLLOUT:-false}"
        echo "[$(date +%H:%M:%S)] ai-quality-baseline (${MODE}): ${SEEDS} seeds × ${TIER1}+${TIER2}, T${TURNS}, PARALLEL=${PARALLEL}"
        for TIER in "${TIER1}" "${TIER2}"; do
            echo "  --- baseline tier=${TIER} ---"
            ssh "${APRICOT}" "set -euo pipefail; cd '${SCRATCH_ABS}' && \
                AI_USE_MCTS=true AI_DIFFICULTY='${TIER}' \
                AI_DIFFICULTY_P0='${TIER}' AI_DIFFICULTY_P1='${TIER}' \
                ${GPU_ENV} PARALLEL=${PARALLEL} \
                bash tools/autoplay-batch.sh ${SEEDS} ${TURNS} ${RESULTS_ABS}/baseline-${TIER} 2>&1 | tail -20"
        done
        ;;
    *)
        echo "ERROR: unknown mode '${MODE}'" >&2
        exit 2
        ;;
esac

# ── Step 5: fetch results summary back to EDIT ───────────────────────────────
LOCAL_RESULTS="${PROJECT_DIR}/.local/iter/apricot-${STAMP}"
mkdir -p "${LOCAL_RESULTS}"
echo "[$(date +%H:%M:%S)] fetch verdict/summary to ${LOCAL_RESULTS}..."
scp -r "${APRICOT}:${RESULTS_ABS}/" "${LOCAL_RESULTS}/" 2>/dev/null || \
    echo "WARN: scp returned non-zero; check manually on ${APRICOT}:${RESULTS_ABS}"

# ── Step 6: remove the per-run worktree ──────────────────────────────────────
# Canonical .git/ + objects retained for the next run. Working tree + cargo
# target dir are gone. RESULTS dir under ~/.cache/mc-batches/ is untouched.
echo "[$(date +%H:%M:%S)] remove worktree ${SCRATCH_ABS}..."
ssh "${APRICOT}" "git -C '${CANONICAL_ABS}' worktree remove --force '${SCRATCH_ABS}' 2>&1 || \
    rm -rf '${SCRATCH_ABS}'"

# ── Step 7: prune old local copies — keep only the 3 most recent ─────────────
ITER_ROOT="${PROJECT_DIR}/.local/iter"
if [[ -d "${ITER_ROOT}" ]]; then
    # List apricot-* dirs newest-first, skip the first 3, delete the rest.
    mapfile -t OLD_RUNS < <(ls -1dt "${ITER_ROOT}"/apricot-* 2>/dev/null | tail -n +4)
    if (( ${#OLD_RUNS[@]} > 0 )); then
        echo "[$(date +%H:%M:%S)] pruning ${#OLD_RUNS[@]} old local run(s) (keeping 3 newest)..."
        for d in "${OLD_RUNS[@]}"; do
            echo "  rm -rf ${d}"
            rm -rf "${d}"
        done
    fi
fi

echo "============================================================"
echo "DONE. Built ${BUILT_SHA} from ${BUILD_REF}."
echo "  Scratch worktree at ${APRICOT}:${SCRATCH_ABS} — REMOVED."
echo "  Results at ${APRICOT}:${RESULTS_ABS} (persistent, .cache)."
echo "  Local copy at ${LOCAL_RESULTS}"
echo "  Canonical at ${APRICOT}:${CANONICAL_ABS} (kept for next run)."
echo "============================================================"