runner.py

"""Orchestrator: dataset name → validator → verdict PR on the dataset.

Phase-3 of the PRD. Same flow the DGXC `tools/hf_watch/validate.py`
performs, minus:
  - the Claude Code subprocess wrapper (we call the validator directly,
    keeping the agentic-decision layer for our internal coordinator path)
  - the status.json patching (the HF Space is per-dataset; the
    coordinator dashboard polls verdicts back via its existing watcher)
  - the GitHub commit step (we open an HF Dataset PR instead)

The validator engine itself is unchanged — it's the same simready-report
skill that runs on Windows, on DGXC, and now here.
"""
from __future__ import annotations

import dataclasses
import json
import os
import shlex
import shutil
import subprocess
import sys
import tempfile
from datetime import datetime, timezone
from pathlib import Path
from typing import Iterator

from huggingface_hub import HfApi, snapshot_download

VALIDATOR = (
    Path(__file__).resolve().parent
    / "tools" / "validation" / "plugins" / "simready-report"
    / "skills" / "simready-report" / "validate.py"
)

# Same exclude set the DGXC path uses. HF datasets ship a lot of bulk the
# validator doesn't open; skipping shrinks downloads from tens-of-GB to
# hundreds-of-MB on assets like nvidia/PhysicalAI-SimReady-Warehouse-01.
HF_DOWNLOAD_EXCLUDES = (
    ".thumbs/*",
    "images/*",
    "*_renders/*", "renders/*",
    "*.mp4", "*.mov", "*.webm",
    "*.jpg", "*.jpeg", "*.png", "*.gif", "*.tiff", "*.tif",
    "*.zip", "*.tar", "*.tgz",
)
USD_EXTS = (".usd", ".usda", ".usdc", ".usdz")


@dataclasses.dataclass
class RunResult:
    dataset: str
    profile: str
    version: str
    status: str         # "pass" | "warn" | "fail" | "error"
    summary: str        # one-line human-readable digest
    results_json: dict  # the validator's results.json contents
    report_path: Path   # local path to the HTML report tree
    pr_url: str | None  # discussion URL when --open-pr was used


def _now() -> str:
    return datetime.now(timezone.utc).isoformat(timespec="seconds")


def _wrap_layout_for_validator(downloaded: Path, work: Path) -> Path:
    """Pass-through. The validator's discover_assets recurses, so we no
    longer need to wrap the download to fit a one-level-deep expectation.
    Kept as a hook so runner.run() doesn't churn if we re-add adaptation
    later (e.g. for zip-bundled datasets that need extraction first).
    """
    return downloaded


def _list_dataset_zips(api: HfApi, dataset: str, token: str | None) -> list[tuple[str, str | None]]:
    """Enumerate `(rel_path, content_sha)` for `*.zip` files in the dataset
    without downloading anything. content_sha is the per-zip cache key —
    prefer the LFS sha256 (large-file pointer) when present, fall back
    to the git blob_id. Returns empty list if the listing fails."""
    try:
        info = api.repo_info(repo_id=dataset, repo_type="dataset",
                              files_metadata=True, token=token)
    except Exception:
        return []
    out_list: list[tuple[str, str | None]] = []
    for sib in (info.siblings or []):
        name = getattr(sib, "rfilename", "") or ""
        if not name.lower().endswith(".zip"):
            continue
        sha = None
        lfs = getattr(sib, "lfs", None)
        if lfs:
            sha = (lfs.get("sha256") if isinstance(lfs, dict)
                   else getattr(lfs, "sha256", None))
        if not sha:
            sha = getattr(sib, "blob_id", None)
        out_list.append((name, sha))
    return out_list


def _zip_cache_key(zip_sha: str, profile: str, validator_version: str,
                   foundation_sha: str) -> str:
    """Per-zip cache key — sha256 of every input that affects the
    validator's verdict for this archive. zip_sha covers content
    changes; the surrounding tuple covers rule-source changes. Runner
    wrapper-code changes are NOT in the key — re-validating a zip
    just because runner.py formatting changed wastes compute for an
    identical verdict. Operator forces a fresh run with Shift+Click,
    which clears this dataset's per-zip cache before re-streaming."""
    import hashlib
    blob = f"{zip_sha}|{profile}|{validator_version}|{foundation_sha}"
    return hashlib.sha256(blob.encode("utf-8")).hexdigest()[:16]


def _zip_cache_path(dataset: str, key: str) -> Path:
    safe = "".join(c if c.isalnum() or c in "-_." else "_" for c in dataset)
    return CACHE_DIR / safe / "zips" / f"{key}.json"


def _read_zip_cache(dataset: str, key: str) -> dict | None:
    p = _zip_cache_path(dataset, key)
    if not p.is_file():
        return None
    try:
        return json.loads(p.read_text(encoding="utf-8"))
    except (OSError, json.JSONDecodeError):
        return None


def _write_zip_cache(dataset: str, key: str, payload: dict) -> None:
    p = _zip_cache_path(dataset, key)
    try:
        p.parent.mkdir(parents=True, exist_ok=True)
        tmp = p.with_suffix(p.suffix + ".tmp")
        tmp.write_text(json.dumps(payload), encoding="utf-8")
        os.replace(tmp, p)
    except OSError:
        pass


def _clear_zip_cache(dataset: str, out) -> None:
    """Wipe the per-zip cache for a dataset. Called from the streaming
    path when force=True so a forced run actually re-validates every
    zip instead of consulting cached entries."""
    safe = "".join(c if c.isalnum() or c in "-_." else "_" for c in dataset)
    zips_dir = CACHE_DIR / safe / "zips"
    if not zips_dir.is_dir():
        return
    try:
        n = sum(1 for _ in zips_dir.glob("*.json"))
        shutil.rmtree(zips_dir, ignore_errors=True)
        out(f"  cleared {n} zip cache entries (force)")
    except OSError as e:
        out(f"  ! zip cache clear failed: {type(e).__name__}: {e}")


def _accumulated_progress_rows(merged_results: list) -> list:
    """Snapshot merged_results into the live-progress row shape the
    dashboard's applyLiveAssetStatus expects: {rel_path, passed,
    status, issues_count}. list() is atomic in CPython so we can
    snapshot without a lock — a slightly stale snapshot is fine."""
    rows = []
    for asset in list(merged_results):
        if not isinstance(asset, dict):
            continue
        sevs = {(iss.get("severity") or "").lower() for iss in (asset.get("issues") or [])}
        if asset.get("passed") and not (sevs & {"error", "failure"}):
            status = "warn" if "warning" in sevs else "pass"
        else:
            status = "fail"
        rows.append({
            "rel_path": asset.get("rel_path") or asset.get("name") or "",
            "passed": bool(asset.get("passed")),
            "status": status,
            "issues_count": len(asset.get("issues") or []),
        })
    return rows


def _write_streaming_progress(path: Path | None, *, processed: int, total: int,
                               current: str | None, started_at: str,
                               state: str = "",
                               results: list | None = None) -> None:
    """Streaming-mode progress emitter. Same JSON shape as the validator's
    per-asset progress so the dashboard's poller reads both transparently;
    the counter is at the zip level here instead of the asset level.
    `results` is the cumulative per-asset list aggregated across all
    zips finished so far (dashboard uses it for the Files expander
    pass/fail overlay)."""
    if not path:
        return
    import tempfile as _tf
    payload = {
        "processed": processed, "total": total, "current": current,
        "started_at": started_at, "state": state,
        "updated_at": _now(),
        "results": (results or [])[-1000:],
    }
    try:
        path.parent.mkdir(parents=True, exist_ok=True)
        fd, tmp = _tf.mkstemp(prefix=".progress-", dir=str(path.parent))
        with os.fdopen(fd, "w", encoding="utf-8") as f:
            json.dump(payload, f)
        os.replace(tmp, str(path))
    except OSError:
        pass


def _find_representative_usd(extract_dir: Path) -> Path | None:
    """Pick a single USD file likely to be the asset's entry point.
    Heuristic priority (lower sort-key = better):
      1. stem matches parent dir name (SimReady bundle convention)
      2. shallower depth
      3. NOT a generic catch-all name like "model.usda" / "main.usd"
         (those are often runtime-generated wrappers, not bundle roots)
    """
    candidates: list[Path] = []
    for ext in (".usd", ".usda", ".usdc", ".usdz"):
        candidates.extend(extract_dir.rglob(f"*{ext}"))
    if not candidates:
        return None
    GENERIC = {"model", "main", "scene", "root", "stage"}
    def key(p: Path):
        depth = len(p.relative_to(extract_dir).parts)
        bundle_match = 0 if p.stem.lower() == p.parent.name.lower() else 1
        is_generic = 1 if p.stem.lower() in GENERIC else 0
        # Order: bundle-name matches first (best signal we're looking
        # at an authored asset root), then non-generic names, then
        # shallowness. Pure depth as last tiebreaker.
        return (bundle_match, is_generic, depth, str(p))
    candidates.sort(key=key)
    return candidates[0]


def _detect_profile_from_usd(usd_path: Path, out) -> str | None:
    """Open a USD file, look at applied schemas + structure, classify
    into the closest existing profile. Returns None on parse failure
    (caller falls back to caller-supplied profile).

    Detection logic, in priority order:
      - Has BOM/Package-* schemas → Package-Candidate
      - Has PhysicsArticulationRootAPI → Robot-Body-{Isaac|Runnable|Neutral}
      - Single-asset content → Prop-Robotics-{Isaac|Physx|Neutral}
      - Many top-level Xforms / sublayers → Package-Candidate (multi-asset bundle)

    The Isaac/Physx/Neutral suffix is picked from applied schemas
    (Isaac > Physx > Neutral). Note: foundation currently has no
    "Scene" profile, so multi-asset scenes route to Package-Candidate
    as the closest existing match — operator should request a Scene
    profile from the foundation team for proper validation."""
    try:
        from pxr import Usd
    except ImportError:
        out("    (usd-core not available; skipping content detection)")
        return None
    try:
        stage = Usd.Stage.Open(str(usd_path))
    except Exception as e:
        out(f"    (couldn't open {usd_path.name} for detection: {type(e).__name__}: {e})")
        return None
    if not stage:
        return None

    has_articulation = False
    has_rigidbody = False
    has_isaac = False
    has_physx = False
    has_bom = False
    top_level_xforms = 0
    sublayer_count = len(stage.GetRootLayer().subLayerPaths)

    for prim in stage.Traverse():
        if prim.GetPath().pathElementCount == 1 and prim.IsA(Usd.Typed):
            type_name = str(prim.GetTypeName())
            if type_name == "Xform":
                top_level_xforms += 1
        schemas = list(prim.GetAppliedSchemas())
        for s in schemas:
            sl = s.lower()
            if "articulationroot" in sl:
                has_articulation = True
            elif "rigidbody" in sl:
                has_rigidbody = True
            elif "physx" in sl:
                has_physx = True
            elif "isaac" in sl:
                has_isaac = True
            elif "bom" in sl or "packageinfo" in sl:
                has_bom = True

    # Classify
    if has_bom or top_level_xforms >= 5 or sublayer_count >= 3:
        # Multi-asset bundle or scene — closest existing profile.
        return "Package-Candidate"
    if has_articulation:
        if has_isaac: return "Robot-Body-Isaac"
        if has_physx: return "Robot-Body-Runnable"
        return "Robot-Body-Neutral"
    # Single-asset prop content
    if has_isaac: return "Prop-Robotics-Isaac"
    if has_physx or has_rigidbody: return "Prop-Robotics-Physx"
    return "Prop-Robotics-Neutral"


def _is_profile_registration_failure(log_file: Path) -> bool:
    """Detect the validator's "profile not registered" signature in its
    log. The CLI-loader failure mode (omniverse-usd-profiles can't
    register because foundation specs reference unknown requirement
    codes) IS recoverable by retrying with --use-plugin. The plugin-
    discovery failure mode (SimReadyPlugin entry point missing) is
    NOT — both code paths go through the same plugin discovery, so
    retrying gains nothing and just doubles compute.

    Returns True only for the recoverable case."""
    if not log_file.is_file():
        return False
    try:
        text = log_file.read_text(encoding="utf-8", errors="replace")
    except OSError:
        return False
    # Plugin discovery is the broken layer? Don't retry — both default
    # and plugin paths share the same plugin loader.
    if ("Discovered plugins: {'omni.asset_validator:DefaultPlugin'}" in text
            and "SimReadyPlugin" in text):
        return False
    if "[CLI-loader] loaded: profiles=0" in text:
        return True
    if "FATAL: profile " in text and "not registered" in text:
        return True
    return False


def _is_unrecoverable_plugin_miss(log_file: Path) -> bool:
    """The 'SimReadyPlugin entry point not installed' shape. Once we see
    this signature, every subsequent zip will fail identically — abort
    the streaming loop early instead of running through 800 doomed
    validates."""
    if not log_file.is_file():
        return False
    try:
        text = log_file.read_text(encoding="utf-8", errors="replace")
    except OSError:
        return False
    return ("Plugin allow-list:" in text
            and "SimReadyPlugin" in text
            and "Discovered plugins: {'omni.asset_validator:DefaultPlugin'}" in text)


def _file_registration_issue(dataset: str, profile: str, val_ver: str,
                              found_sha: str, log_file: Path, out) -> None:
    """File a single GitHub issue documenting the foundation/validator
    registration mismatch that triggered the --use-plugin retry.
    Best-effort, deduplicated by title via github_issues helpers."""
    try:
        from github_issues import _gh_token, _find_issue, _create_issue, _add_comment
    except Exception as e:
        out(f"    (issue-filing import failed: {type(e).__name__}: {e})")
        return
    if not _gh_token():
        out(f"    (no GH token; skipping issue filing)")
        return
    title = f"[validator-internal] CLI loader emits 0 profiles for foundation {found_sha[:8]} + simready-validate {val_ver}"
    try:
        tail = ""
        try:
            tail = "\n".join(log_file.read_text(encoding="utf-8", errors="replace")
                             .splitlines()[-25:])[:3000]
        except OSError:
            pass
        body = (
            "**Validator-internal bug** — surfaced by the HF Space streaming-zip path.\n\n"
            "The default CLI loader loads 0 profiles when run against the pinned\n"
            "foundation specs + simready-validate combination, because features\n"
            "reference requirement codes the validator package doesn't have\n"
            "registered. Recoverable at runtime via `--use-plugin`, but the\n"
            "underlying mismatch should be fixed in either the foundation pin\n"
            "or the validator package version.\n\n"
            f"| Field | Value |\n|---|---|\n"
            f"| Dataset (first hit) | `{dataset}` |\n"
            f"| Profile | `{profile}` |\n"
            f"| simready-validate | `{val_ver}` |\n"
            f"| foundation sha | `{found_sha}` |\n"
            f"| Workaround in effect | `--use-plugin` for this run |\n\n"
            f"**Loader log tail:**\n\n```\n{tail}\n```\n"
        )
        existing = _find_issue(title)
        if existing:
            _add_comment(existing["number"],
                         f"Re-hit during validation of `{dataset}`. --use-plugin recovery engaged.")
            out(f"    internal-issue #{existing['number']}: comment added")
        else:
            num = _create_issue(title, body, ["validator-internal", "process"])
            out(f"    internal-issue #{num}: opened")
    except Exception as e:
        out(f"    (issue filing failed: {type(e).__name__}: {e})")


def _validate_zip_streaming(*, api: HfApi, dataset: str, token: str | None,
                             work: Path, profile: str, version: str,
                             progress_file: Path | None, out,
                             force: bool = False,
                             submission_id: str = "",
                             flat_target: Path | None = None,
                             prefetched_zip_entries: list | None = None,
                             prefetched_dataset_head: str | None = None,
                             ) -> dict | None:
    """Validate a zip-bundled dataset by streaming one archive at a time.

    DEPRECATED — zip-bundled datasets are not allowed per the foundation
    AA.002 spec (allowlist is USD/image/audio only — no .zip) or the
    SDK packaging spec (describes unpacked layout only). run() now
    fails such datasets at the preliminary check stage before any
    download happens.

    This function is intentionally retained for the case where the
    spec is amended to accept zips as a transport mechanism. The flat-
    path code in run() also reuses this function (passing flat_target)
    for the unified daemon-pool + cache + cancel + progress machinery
    — that path is NOT deprecated.

    Flow per zip (deprecated path): hf_hub_download → extract →
    validate.py → capture results.json → delete archive + extracted
    tree → next. Never holds more than one zip's worth of data on
    disk, so works on datasets whose total size doesn't fit on the
    Space's ephemeral /tmp.

    Returns a results.json-shaped dict aggregating every per-zip run.
    The `results` list has each asset's `rel_path` prefixed with its
    source zip so dashboard rows can show e.g.
    `kitchen_03.zip/kitchen_03/scene.usd`.

    Returns None when the dataset has no zips at all AND no flat_target
    is provided — caller misconfigured (the strict pre-check in run()
    should never let this happen).
    """
    from huggingface_hub import hf_hub_download
    import zipfile

    # Use the caller's pre-fetched zip listing + dataset HEAD when
    # available. run() already calls _list_dataset_zips() and
    # repo_info() to decide flat vs zip + populate the dataset-level
    # cache; calling them again here doubled the HF API request count
    # per validation for no value.
    if prefetched_zip_entries is not None:
        zip_entries = prefetched_zip_entries
    else:
        zip_entries = _list_dataset_zips(api, dataset, token)
    # Unified path: if the dataset has no zip files, synthesize a SINGLE
    # "unit" representing the whole dataset. snapshot_download has
    # already (or will) materialize the contents into flat_target;
    # downstream daemon-pool validation treats it the same as one zip.
    is_flat = not zip_entries
    if is_flat:
        if flat_target is None or not flat_target.is_dir():
            return None  # caller must provide the materialized dir
        head = prefetched_dataset_head
        if head is None:
            try:
                head = api.repo_info(dataset, repo_type="dataset").sha
            except Exception:
                head = ""
        zip_entries = [(dataset, head)]
        out(f"  flat dataset: snapshot at {flat_target}; validator will discover assets")
    if force:
        _clear_zip_cache(dataset, out)
    if not is_flat:
        out(f"  zip-bundled dataset: {len(zip_entries)} zip(s); streaming one at a time"
            + (" (force)" if force else ""))
    started_at = _now()
    # Flat mode: the daemon owns the progress file (writes per-asset
    # progress). Streaming-loop's unit-level writes would clobber the
    # validator's "k of N assets" with a useless "0/1" / "1/1" counter.
    def _emit_unit_progress(**kw):
        if is_flat:
            return
        _write_streaming_progress(progress_file, **kw)

    _emit_unit_progress(processed=0, total=len(zip_entries),
                        current=None, started_at=started_at, state="starting")

    merged_results: list[dict] = []
    merged_layout: list[dict] = []
    merged_preliminary: list[dict] = []
    # Set when ANY processed unit's results.json carries
    # preliminary_check_failed=true (the validator's strict pre-check fired).
    # Propagated into the final dict so the dashboard sees the flag
    # and renders the layout-failed banner instead of generic counts.
    any_preliminary_check_failed = False
    workers = os.environ.get("SR_WORKERS", "4").strip() or "4"
    cache_hits = 0
    val_ver = _validator_version()
    found_sha = _foundation_sha()

    # Always use --use-plugin in the streaming path. With the patched
    # foundation wheel installed, SimReadyPlugin's on_startup() loads
    # profiles via simready.validate.impl.loader at omni.asset_validator
    # import time — fully populated ProfileRegistry before validate.py
    # main() runs. The default CLI-loader path still attempts a
    # second `load_validation_implementation` call which races with
    # the plugin's registration and ends up failing in subtle ways
    # (we've observed FATAL: profile X not registered even though
    # the plugin succeeded — different code path, different bug).
    # Skip the doomed-first-attempt entirely.
    use_plugin_default = True
    issue_filed_for_registration_bug = False
    # Once an issue-filing attempt 404s (token can't reach the repo,
    # repo wrong, permissions missing), don't try again this run —
    # avoids spamming 30+ 404s in the log.
    issue_filing_disabled = False
    # Profile auto-detect state — runs once on the first zip's extracted
    # tree. If content-detection disagrees with the caller-supplied
    # profile, override for all remaining zips in this run.
    profile_autodetect_done = False
    # Abort after this many consecutive unrecoverable failures so we
    # don't burn 800 zips' worth of compute on a known-broken Space.
    consecutive_unrecoverable = 0
    UNRECOVERABLE_ABORT_AT = 3
    was_cancelled = False
    zips_processed = 0

    # Parallelism strategy — total concurrency = SR_WORKERS in both cases:
    # - Zip path (N units, 1 asset/unit after scene-root reduction):
    #   N daemons × 1 internal worker each. Each daemon processes one
    #   zip end-to-end; the streaming loop fans out via ThreadPoolExecutor.
    # - Flat path (1 unit with many assets):
    #   1 daemon × N internal workers. Single daemon's fork pool handles
    #   the M assets discovered in the snapshot.
    # The N²-over-subscription case (N daemons × N workers) is avoided.
    if is_flat:
        n_daemons = 1
        daemon_workers = workers  # SR_WORKERS
    else:
        n_daemons = max(1, int(workers) if str(workers).isdigit() else 1)
        daemon_workers = "1"
    daemon_pool: list[subprocess.Popen] = []
    daemon_locks: list = []  # threading.Lock per daemon for IO safety
    import threading as _threading
    import queue as _queue
    daemon_cmd = [
        sys.executable, str(VALIDATOR), "--daemon",
        "--use-plugin", "--no-use-kit", "--workers", daemon_workers,
        "--profile", profile, "--version", version,
    ]
    out(f"  spawning {n_daemons} validator daemon(s) (spec load happens once each)…")
    for di in range(n_daemons):
        try:
            proc = subprocess.Popen(
                daemon_cmd,
                stdin=subprocess.PIPE, stdout=subprocess.PIPE,
                stderr=subprocess.STDOUT, text=True, bufsize=1,
            )
            # Wait for __DAEMON_READY__ on stdout (spec load can take 60s+).
            ready = False
            for _ in range(300):
                line = proc.stdout.readline()
                if not line: break
                if "__DAEMON_READY__" in line:
                    ready = True
                    break
            if not ready:
                proc.kill()
                out(f"  daemon[{di}] never reported ready; skipping")
                continue
            daemon_pool.append(proc)
            daemon_locks.append(_threading.Lock())
            out(f"  daemon[{di}] ready ({len(daemon_pool)}/{n_daemons})")
        except Exception as e:
            out(f"  daemon[{di}] spawn failed ({type(e).__name__}: {e})")
    if not daemon_pool:
        out(f"  no daemons spawned; falling back to per-zip subprocess.call")
    # Queue of available daemon indices. Workers check out one,
    # validate, return.
    available_daemons: _queue.Queue = _queue.Queue()
    for idx in range(len(daemon_pool)):
        available_daemons.put(idx)
    # Shared-state locks for the parallel per-zip workers below.
    _state_lock = _threading.Lock()
    _stop_event = _threading.Event()

    def _process_zip(i: int, zip_rel: str, zip_sha) -> None:
        nonlocal cache_hits, zips_processed, profile_autodetect_done
        nonlocal profile, consecutive_unrecoverable, was_cancelled
        nonlocal use_plugin_default, issue_filed_for_registration_bug
        nonlocal issue_filing_disabled, any_preliminary_check_failed
        # Honor early abort (cancel or unrecoverable failure) — tasks
        # queued before the stop signal still get scheduled and have
        # to no-op themselves.
        if _stop_event.is_set():
            return
        # Cancel check: dashboard's Cancel button POSTs to the Space's
        # cancel_run endpoint which creates /tmp/sr-cancel/<id>. Stop
        # accepting new work so the in-flight gradio call returns
        # promptly with whatever partial results we have instead of
        # grinding through hundreds more zips.
        if submission_id and _is_cancelled(submission_id):
            with _state_lock:
                if not was_cancelled:
                    out(f"  CANCEL signal received — stopping (in-flight tasks finish)")
                    try:
                        p = cancel_path_for(submission_id)
                        if p and p.exists():
                            p.unlink()
                    except OSError:
                        pass
                    was_cancelled = True
            _stop_event.set()
            return
        _emit_unit_progress(processed=i, total=len(zip_entries),
                            current=zip_rel, started_at=started_at, state="zip",
                            results=_accumulated_progress_rows(merged_results))

        # Cache lookup. Skip when zip_sha is None (HF didn't surface a
        # blob sha — rare, defensive). force=True already cleared the
        # cache above so the read here will miss.
        cache_key = None
        if zip_sha:
            cache_key = _zip_cache_key(zip_sha, profile, val_ver, found_sha)
            cached = _read_zip_cache(dataset, cache_key) if not force else None
            if cached:
                merged_results.extend(cached.get("results", []))
                merged_layout.extend(cached.get("layout_findings") or [])
                if cached.get("preliminary_check_failed"):
                    any_preliminary_check_failed = True
                cache_hits += 1
                out(f"  [{i+1}/{len(zip_entries)}] cache hit: {zip_rel} "
                    f"({len(cached.get('results', []))} asset(s))")
                _emit_unit_progress(processed=i + 1, total=len(zip_entries),
                                    current=zip_rel,
                                    started_at=started_at, state="zip",
                                    results=_accumulated_progress_rows(merged_results))
                return

        per_zip = work / f"zip_{i:04d}"
        per_zip.mkdir(parents=True, exist_ok=True)
        if is_flat:
            # Flat dataset: caller already materialized the contents in
            # flat_target. Treat it as the sole pre-extracted "unit".
            # No download, no zip extraction step — just point the
            # validator at the snapshot dir directly.
            extract_dir = flat_target
        else:
            extract_dir = per_zip / "extracted"
        out_dir = per_zip / "out"
        out_dir.mkdir(parents=True, exist_ok=True)

        zip_results: list[dict] = []
        zip_layout: list[dict] = []
        try:
            if not is_flat:
                out(f"  [{i+1}/{len(zip_entries)}] download: {zip_rel}")
                downloaded_str = hf_hub_download(
                    repo_id=dataset, repo_type="dataset",
                    filename=zip_rel, local_dir=str(per_zip),
                    token=token,
                )
                downloaded = Path(downloaded_str)
                if not downloaded.is_file():
                    out(f"    ! download produced no file at {downloaded}")
                    return

                out(f"  [{i+1}/{len(zip_entries)}] extract")
                extract_dir.mkdir(parents=True, exist_ok=True)
                with zipfile.ZipFile(downloaded) as zf:
                    zf.extractall(extract_dir)
                try: downloaded.unlink()
                except OSError: pass

            # Profile auto-detect — only when the caller explicitly
            # asked for "Auto" (the dashboard's default). Any specific
            # profile bypasses detection entirely: operator override
            # is respected as the source of truth.
            if not profile_autodetect_done and profile.lower() == "auto":
                sample = _find_representative_usd(extract_dir)
                if sample is not None:
                    detected = _detect_profile_from_usd(sample, out)
                    if detected:
                        out(f"  profile auto-detect: 'Auto' → '{detected}' "
                            f"(sampled {sample.relative_to(extract_dir)})")
                        profile = detected
                        if zip_sha:
                            cache_key = _zip_cache_key(zip_sha, profile, val_ver, found_sha)
                    else:
                        # Detection failed — fall back to a permissive default.
                        out(f"  profile auto-detect: could not classify; falling back to Prop-Robotics-Neutral")
                        profile = "Prop-Robotics-Neutral"
                        if zip_sha:
                            cache_key = _zip_cache_key(zip_sha, profile, val_ver, found_sha)
                profile_autodetect_done = True

            def _run_validator(use_plugin: bool) -> int:
                # Check out a daemon from the pool, send the request,
                # read response, return daemon to pool. The pool's
                # blocking get() naturally limits concurrent validates
                # to len(daemon_pool). Falls back to one-shot
                # subprocess.call when no daemons are available.
                if daemon_pool:
                    daemon_idx = available_daemons.get()
                    try:
                        proc = daemon_pool[daemon_idx]
                        if proc.poll() is not None:
                            # This daemon died — skip pool entry and fall through.
                            raise RuntimeError(f"daemon[{daemon_idx}] dead")
                        req = {
                            "target": str(extract_dir),
                            "output": str(out_dir),
                            "profile": profile,
                            "version": version,
                            "use_kit": False,
                        }
                        # Progress-file ownership:
                        # - Zip path (many units, 1 asset each after
                        #   scene-root reduction): streaming loop owns
                        #   the progress file, emits "k of N zips". Do
                        #   NOT pass progress_file to the daemon — it
                        #   would overwrite the zip counter with "1 of 1".
                        # - Flat path (1 unit with many assets): unit
                        #   counter is useless ("0/1" / "1/1"). Hand the
                        #   progress file to the daemon so the validator
                        #   emits per-asset progress. Streaming loop
                        #   skips its own writes (see is_flat checks).
                        if is_flat and progress_file is not None:
                            req["progress_file"] = str(progress_file)
                        with daemon_locks[daemon_idx]:
                            proc.stdin.write(json.dumps(req) + "\n")
                            proc.stdin.flush()
                            rc = 99
                            with log_file.open("w", encoding="utf-8") as logf:
                                for line in proc.stdout:
                                    line = line.rstrip("\n")
                                    if line.startswith("__DAEMON_RESPONSE__"):
                                        try:
                                            payload = json.loads(line[len("__DAEMON_RESPONSE__"):].strip())
                                            rc = int(payload.get("rc", 99))
                                        except Exception:
                                            pass
                                        break
                                    logf.write(line + "\n")
                        return rc
                    except Exception as e:
                        out(f"    daemon[{daemon_idx}] failed ({type(e).__name__}: {e}); falling back to subprocess")
                    finally:
                        available_daemons.put(daemon_idx)
                # Fallback: one-shot subprocess (original path).
                cmd = [
                    sys.executable, str(VALIDATOR), str(extract_dir),
                    "--profile", profile, "--version", version,
                    "--output", str(out_dir), "--no-use-kit",
                    "--workers", workers,
                ]
                if use_plugin:
                    cmd.append("--use-plugin")
                with log_file.open("wb") as logf:
                    return subprocess.call(cmd, stdout=logf, stderr=subprocess.STDOUT)

            log_file = out_dir / "validator.log"
            out(f"  [{i+1}/{len(zip_entries)}] validate"
                + (" (--use-plugin)" if use_plugin_default else ""))
            rc = _run_validator(use_plugin=use_plugin_default)
            results_path = out_dir / "results.json"

            # Deterministic recovery: if the default loader produced
            # the "profile not registered" signature, retry the same
            # zip with --use-plugin. If that works, promote it to the
            # default for every remaining zip.
            if (not results_path.is_file() and not use_plugin_default
                    and _is_profile_registration_failure(log_file)):
                out(f"    detected loader-registration failure; retrying with --use-plugin")
                if not issue_filed_for_registration_bug and not issue_filing_disabled:
                    try:
                        _file_registration_issue(dataset, profile, val_ver, found_sha,
                                                  log_file, out)
                        issue_filed_for_registration_bug = True
                    except Exception as e:
                        if "404" in str(e):
                            out(f"    issue filing 404'd; disabling for the rest of this run")
                            issue_filing_disabled = True
                rc = _run_validator(use_plugin=True)
                if results_path.is_file():
                    out(f"    --use-plugin recovered; switching default for remaining zips")
                    use_plugin_default = True

            # Unrecoverable: SimReadyPlugin entry point not installed.
            # Both loader paths go through the same plugin discovery,
            # so retrying won't help. Track consecutive failures and
            # abort the loop after N to avoid wasting compute.
            if not results_path.is_file() and _is_unrecoverable_plugin_miss(log_file):
                consecutive_unrecoverable += 1
                if consecutive_unrecoverable >= UNRECOVERABLE_ABORT_AT:
                    out(f"  ABORTING: {consecutive_unrecoverable} consecutive failures with "
                        f"'SimReadyPlugin not discovered' — the foundation entry point isn't "
                        f"installed on this Space, validator cannot proceed regardless of "
                        f"how many zips we try")
                    shutil.rmtree(per_zip, ignore_errors=True)
                    # Signal all other in-flight tasks to stop early.
                    _stop_event.set()
                    return
            elif results_path.is_file():
                consecutive_unrecoverable = 0
                zips_processed += 1

            if results_path.is_file():
                try:
                    rj = json.loads(results_path.read_text(encoding="utf-8"))
                except json.JSONDecodeError:
                    rj = {}
                for asset in rj.get("results", []):
                    asset_rel = (asset.get("rel_path") or "").lstrip("./")
                    if not is_flat:
                        asset["rel_path"] = f"{zip_rel}/{asset_rel}".replace("//", "/")
                    else:
                        asset["rel_path"] = asset_rel
                zip_results = rj.get("results", [])
                zip_layout = rj.get("layout_findings") or []
                zip_preliminary = rj.get("preliminary_findings") or []
                merged_results.extend(zip_results)
                merged_layout.extend(zip_layout)
                merged_preliminary.extend(zip_preliminary)
                if rj.get("preliminary_check_failed"):
                    any_preliminary_check_failed = True
                out(f"    {len(zip_results)} asset(s); rc={rc}")
                # Emit a progress write so the dashboard sees the
                # updated zip-count + per-asset rows immediately
                # (next poll picks them up). Without this the chip
                # only updates on the NEXT zip's "zip" state write.
                _emit_unit_progress(processed=i + 1, total=len(zip_entries),
                                    current=zip_rel,
                                    started_at=started_at, state="zip",
                                    results=_accumulated_progress_rows(merged_results))
                # Write per-zip cache entry on successful validation. We
                # cache even when rc!=0 IF results.json was produced —
                # the validator may exit 1 to signal failures-present
                # while still having emitted a valid report.
                if cache_key and rj:
                    _write_zip_cache(dataset, cache_key, {
                        "schema_version": 1,
                        "zip_rel": zip_rel,
                        "zip_sha": zip_sha,
                        "results": zip_results,
                        "layout_findings": zip_layout,
                        "preliminary_check_failed": bool(rj.get("preliminary_check_failed")),
                        "validator_version": val_ver,
                        "foundation_sha": found_sha,
                        "profile": profile,
                        "cached_at": _now(),
                    })
            else:
                # Diagnostic: dump the validator's own log tail into
                # the Space log so we can see WHY the zip failed.
                # Without this we just see "rc=N" lines forever and
                # have no idea what the validator was complaining about.
                tail_lines: list[str] = []
                if log_file.is_file():
                    try:
                        text = log_file.read_text(encoding="utf-8", errors="replace")
                        tail_lines = text.splitlines()[-20:]
                    except OSError:
                        pass
                # Also list what's actually in the extracted tree to
                # diagnose "extracted but no USDs found" cases — common
                # if the zip has USDs nested deeper than discover_assets
                # walks, or uses an extension we don't recognize.
                tree_sample: list[str] = []
                try:
                    files = sorted(p for p in extract_dir.rglob("*") if p.is_file())
                    for p in files[:8]:
                        rel = p.relative_to(extract_dir)
                        tree_sample.append(f"      {rel}")
                    if len(files) > 8:
                        tree_sample.append(f"      ... and {len(files) - 8} more")
                except OSError:
                    pass
                out(f"    ! no results.json (rc={rc})")
                if tree_sample:
                    out(f"    extracted tree ({sum(1 for _ in extract_dir.rglob('*') if _.is_file())} files):")
                    for line in tree_sample:
                        out(line)
                if tail_lines:
                    out(f"    validator log tail:")
                    for line in tail_lines:
                        out(f"      {line[:240]}")
        except Exception as e:
            out(f"  ! [{i+1}/{len(zip_entries)}] {type(e).__name__}: {e}")
        finally:
            shutil.rmtree(per_zip, ignore_errors=True)

    # Dispatch all zips to the daemon pool via a thread pool. Concurrency
    # is bounded by max_workers (= number of live daemons). Each thread
    # runs _process_zip for one zip end-to-end (download + extract +
    # validate). Cancel signal causes pending tasks to no-op via the
    # _stop_event check at function entry.
    import concurrent.futures as _futures
    n_parallel = max(1, len(daemon_pool))
    out(f"  dispatching {len(zip_entries)} zip(s) across {n_parallel} parallel worker(s)")
    with _futures.ThreadPoolExecutor(max_workers=n_parallel) as _ex:
        _all_futures = [
            _ex.submit(_process_zip, i, zr, zs)
            for i, (zr, zs) in enumerate(zip_entries)
        ]
        for _fut in _futures.as_completed(_all_futures):
            try:
                _fut.result()
            except Exception as _e:
                out(f"  ! task crashed: {type(_e).__name__}: {_e}")

    # Teardown daemon pool. Close stdins so daemons exit cleanly;
    # short wait then kill to bound shutdown time.
    for proc in daemon_pool:
        try:
            proc.stdin.close()
        except Exception:
            pass
    for proc in daemon_pool:
        try:
            proc.wait(timeout=10)
        except subprocess.TimeoutExpired:
            proc.kill()
        except Exception:
            pass

    _emit_unit_progress(processed=len(zip_entries), total=len(zip_entries),
                        current=None, started_at=started_at, state="done",
                        results=_accumulated_progress_rows(merged_results))
    out(f"  zip-streaming done: {cache_hits} cached, "
        f"{zips_processed} freshly validated"
        + (f", CANCELLED after {zips_processed + cache_hits} of {len(zip_entries)}" if was_cancelled else ""))
    return {
        "schema_version": 1,
        "results": merged_results,
        "layout_findings": merged_layout,
        "preliminary_findings": merged_preliminary,
        "preliminary_check_failed": any_preliminary_check_failed,
        "profile_coverage": {},
        "streaming_zips": len(zip_entries),
        "streaming_cache_hits": cache_hits,
        "streaming_processed": zips_processed + cache_hits,
        "cancelled": was_cancelled,
    }


def _summarize(results_json: dict) -> tuple[str, str]:
    """Return (status, one-line summary)."""
    # Preliminary-check failures short-circuit the normal
    # "M/N assets passed" framing — the dataset didn't get to USD
    # validation because filesystem-only foundation checks already
    # flagged issues. The summary names the phase so the operator
    # knows what to do (forward the report to the partner; address
    # these before re-validating to surface deeper USD findings).
    if results_json.get("preliminary_check_failed"):
        # Count actual issues by summing across results — robust to
        # whichever sidecar field the validator populated.
        violations = sum(len(r.get("issues") or [])
                          for r in (results_json.get("results") or []))
        if violations == 0:
            # Fall back to the sidecar list when results is empty
            # (shouldn't happen, defensive).
            violations = len(results_json.get("preliminary_findings")
                             or results_json.get("layout_findings") or [])
        files_affected = len(results_json.get("results") or [])
        # Per-code breakdown for the chip text — the partner-facing
        # summary is more useful when it names the failing rules.
        code_counts: dict[str, int] = {}
        for r in (results_json.get("results") or []):
            for iss in (r.get("issues") or []):
                c = iss.get("code") or "UNKNOWN"
                code_counts[c] = code_counts.get(c, 0) + 1
        top_codes = sorted(code_counts.items(), key=lambda kv: -kv[1])[:3]
        codes_text = ", ".join(f"{c} ×{n}" for c, n in top_codes) if top_codes else "0 issues"
        return "fail", (f"PRELIMINARY CHECK FAILED — {codes_text} "
                        f"({files_affected} file(s) affected). Address these "
                        f"before deeper validation runs.")
    counts = {"error": 0, "failure": 0, "warning": 0}
    total = len(results_json.get("results", []))
    failed = 0
    for asset in results_json.get("results", []):
        if not asset.get("passed"):
            failed += 1
        for issue in asset.get("issues", []):
            sev = (issue.get("severity") or "").lower()
            if sev in counts:
                counts[sev] += 1
    if counts["error"] or counts["failure"]:
        status = "fail"
    elif counts["warning"]:
        status = "warn"
    elif total > 0:
        status = "pass"
    else:
        status = "warn"
    parts = [f"{total - failed}/{total} assets passed"]
    parts += [f"{k}={v}" for k, v in counts.items() if v]
    coverage = results_json.get("profile_coverage") or {}
    if coverage.get("missing"):
        parts.append(f"coverage {coverage.get('loaded')}/{coverage.get('declared')} features")
    return status, " · ".join(parts)


def _open_verdict_pr(
    api: HfApi, dataset: str, results_path: Path, report_dir: Path,
    profile: str, version: str, status: str, summary: str,
) -> str | None:
    """Upload `validation/results.json` + `validation/report/` to the dataset
    as a PR. Returns the discussion URL.

    Why PR rather than commit-to-main: the dataset owner reviews the
    verdict like any other change. The HF Hub PR flow is exactly the
    surface the production end-state assumes — see PRD §3.
    """
    import io

    pr_branch = f"simready-validate/{profile}-v{version}-{_now().replace(':', '-')}"
    body_md = (
        f"### SimReady validation\n\n"
        f"- **Profile**: `{profile}` v{version}\n"
        f"- **Status**: **{status.upper()}**\n"
        f"- **Summary**: {summary}\n"
        f"- **Generated**: {_now()}\n\n"
        f"Run by the SimReady Validator HF Space. The full HTML report "
        f"is in `validation/report/index.html`; machine-readable "
        f"results in `validation/results.json`.\n"
    )

    # Stage everything that should land in the dataset under a single
    # tree we can iterate. `validation/results.json` plus the entire
    # `validation/report/` directory.
    additions: list[tuple[str, bytes]] = []
    additions.append(("validation/results.json", results_path.read_bytes()))
    for path in report_dir.rglob("*"):
        if path.is_file():
            rel = path.relative_to(report_dir.parent)  # keep `report/...`
            additions.append((f"validation/{rel.as_posix()}", path.read_bytes()))

    from huggingface_hub import CommitOperationAdd
    operations = [
        CommitOperationAdd(path_in_repo=p, path_or_fileobj=io.BytesIO(b))
        for p, b in additions
    ]
    commit = api.create_commit(
        repo_id=dataset, repo_type="dataset",
        operations=operations,
        commit_message=f"simready-validate: {profile} v{version} → {status}",
        create_pr=True,
    )
    # `create_pr=True` returns the PR's revision; the discussion URL is
    # derivable from it. HfApi exposes the field but its key name has
    # varied across versions — fall back gracefully.
    return getattr(commit, "pr_url", None) or getattr(commit, "discussion_url", None)


PROGRESS_DIR = Path("/tmp/sr-progress")
# Cancel signal directory. The streaming-zip loop checks for the
# existence of /tmp/sr-cancel/<submission_id> between zips; presence
# means abort. Set by the Space's cancel_run gradio endpoint (called
# from the dashboard when the operator clicks Cancel) — the GH Action
# cancel alone doesn't stop the in-flight gradio call server-side.
CANCEL_DIR = Path("/tmp/sr-cancel")


def cancel_path_for(submission_id: str) -> Path | None:
    if not submission_id:
        return None
    safe = "".join(c if c.isalnum() or c in "-_." else "_" for c in submission_id)
    return CANCEL_DIR / safe


def _is_cancelled(submission_id: str) -> bool:
    p = cancel_path_for(submission_id)
    return bool(p and p.is_file())

# Persistent volume mounted on the Space — survives container restarts.
# See space_info().runtime.raw["volumes"]: nvidia/simready-validator-storage
# is mounted at /data. We keep results.json + the summary keyed by the
# four-tuple that determines "would this run produce the same answer?"
# When the next call matches all four, we serve the cached result
# instead of paying ~5 min for an identical re-run.
CACHE_DIR = Path("/data/sr-cache")


def _cache_key(dataset_head: str, profile: str, validator_version: str,
               foundation_sha: str) -> str:
    """Stable key over every input that determines the verdict. Runner
    wrapper-code changes are intentionally NOT in the key — they don't
    change what assets passed/failed, only how the result is shaped on
    its way out. Shift+Click is the operator's escape valve when they
    actually want a fresh re-run."""
    import hashlib
    blob = f"{dataset_head}|{profile}|{validator_version}|{foundation_sha}"
    return hashlib.sha256(blob.encode("utf-8")).hexdigest()[:16]


def _cache_path_for(dataset: str, key: str) -> Path:
    """One file per dataset+key. Dataset name in the path so an operator
    can browse the cache by partner."""
    safe = "".join(c if c.isalnum() or c in "-_." else "_" for c in dataset)
    return CACHE_DIR / safe / f"{key}.json"


def _foundation_sha() -> str:
    """Pinned commit of NVIDIA/simready-foundation that the Space was
    built against. Set by the Dockerfile (ENV SIMREADY_FOUNDATIONS_COMMIT)."""
    return os.environ.get("SIMREADY_FOUNDATIONS_COMMIT", "unpinned")


# Path to the spec-sync state file. The state file declares which
# foundation commit our hardcoded validator rules were last aligned
# against. The Space's checkout includes the file at this relative
# path (tools/spec_sync/state.json in the repo).
_SPEC_SYNC_STATE_FILE = Path(__file__).resolve().parents[2] / "tools" / "spec_sync" / "state.json"


def _check_foundation_spec_drift() -> dict:
    """Check whether NVIDIA/simready-foundation has new commits to its
    spec dir since we last synced our hardcoded rules.

    Cheap event-driven drift detection (one GitHub API call per run,
    soft-fails on errors). The validator surfaces drift in results.json
    so the dashboard can warn operators that hardcoded rules may be
    stale relative to the source of truth. Auto-update is the
    follow-up step (spec-sync workflow opens a PR to refresh rules).

    Returns a dict the dashboard renders; never raises.
    """
    out: dict = {"checked": False}
    try:
        if not _SPEC_SYNC_STATE_FILE.is_file():
            out["reason"] = "state file missing"
            return out
        state = json.loads(_SPEC_SYNC_STATE_FILE.read_text(encoding="utf-8"))
        last_sync_sha = state.get("foundation_commit_sha") or ""
        last_sync_at = state.get("synced_at") or ""
        watched_path = state.get("watched_path") or "nv_core/sr_specs/docs"
        repo = state.get("foundation_repo") or "NVIDIA/simready-foundation"
    except Exception as e:
        out["reason"] = f"could not read state: {type(e).__name__}: {e}"
        return out
    try:
        import urllib.request
        url = (f"https://api.github.com/repos/{repo}/commits"
               f"?path={watched_path}&per_page=1")
        req = urllib.request.Request(url)
        req.add_header("Accept", "application/vnd.github.v3+json")
        token = os.environ.get("GITHUB_TOKEN") or os.environ.get("GH_VALIDATOR_TOKEN")
        if token:
            req.add_header("Authorization", f"Bearer {token}")
        with urllib.request.urlopen(req, timeout=8) as resp:
            data = json.loads(resp.read())
    except Exception as e:
        out["reason"] = f"github api: {type(e).__name__}: {e}"
        return out
    if not isinstance(data, list) or not data:
        out["reason"] = "no commit data"
        return out
    current = data[0] or {}
    current_sha = current.get("sha") or ""
    current_at = (current.get("commit") or {}).get("committer", {}).get("date") or ""
    drifted = bool(current_sha) and current_sha != last_sync_sha
    return {
        "checked": True,
        "drifted": drifted,
        "current_sha": current_sha,
        "current_at": current_at,
        "last_sync_sha": last_sync_sha,
        "last_sync_at": last_sync_at,
        "repo": repo,
        "watched_path": watched_path,
    }


def _validator_version() -> str:
    """Version of the simready-validate package that ships in this Space."""
    try:
        import importlib.metadata as md
        return md.version("simready-validate")
    except Exception:
        return "unknown"


def _read_cache(dataset: str, key: str) -> dict | None:
    """Read + sanity-check a cached dataset-level entry. Returns None
    for stale / broken entries so the caller falls through to a real
    re-run instead of replaying garbage.

    Stale signatures we reject:
      - results_json.results == [] with status pass/warn/fail
        (impossible from a correct run — validator emits status=error
        when it can't find any USDs, never pass/warn/fail with zero).
        Detects the broken-pre-streaming era where zips were excluded
        and the cached payload looks "successful" with zero work done.
    """
    p = _cache_path_for(dataset, key)
    if not p.is_file():
        return None
    try:
        payload = json.loads(p.read_text(encoding="utf-8"))
    except (OSError, json.JSONDecodeError):
        return None
    rj = payload.get("results_json") or {}
    results = rj.get("results") or []
    status = payload.get("status") or ""
    if not results and status in ("pass", "warn", "fail"):
        # Suspicious — looks like a stale entry from a code path that
        # didn't actually validate anything. Treat as miss.
        return None
    return payload


def _write_cache(dataset: str, key: str, payload: dict) -> None:
    p = _cache_path_for(dataset, key)
    try:
        p.parent.mkdir(parents=True, exist_ok=True)
        # Atomic via temp+rename so concurrent reads can't see a half file.
        tmp = p.with_suffix(p.suffix + ".tmp")
        tmp.write_text(json.dumps(payload), encoding="utf-8")
        os.replace(tmp, p)
    except OSError:
        # Cache is advisory — never block a real validation on disk hiccups.
        pass



def progress_path_for(submission_id: str) -> Path:
    """Where the validator writes per-asset progress for this submission.
    Read by the Space's get_progress endpoint to feed the dashboard's
    fill-up progress bar. Empty submission_id → None (caller skips)."""
    if not submission_id:
        return None  # type: ignore[return-value]
    safe = "".join(c if c.isalnum() or c in "-_." else "_" for c in submission_id)
    return PROGRESS_DIR / f"{safe}.json"


def _finalize_run(*, dataset: str, profile: str, version: str,
                  results_json: dict, status: str, summary: str,
                  out_dir: Path, api: HfApi, token: str | None,
                  open_pr: bool, results_path: Path, out,
                  dataset_head: str | None = None) -> RunResult:
    """Shared tail-end of run(): file issues, optionally open PR on
    dataset, persist report, write cache, return RunResult."""
    try:
        from github_issues import ensure_internal_issues
        ensure_internal_issues(results_json, dataset=dataset, profile=profile, log_fn=out)
    except Exception as e:
        out(f"  ! issue-filing skipped: {type(e).__name__}: {e}")

    pr_url = None
    if open_pr:
        if not token:
            out("  ! HF_TOKEN missing; cannot open PR")
        else:
            try:
                pr_url = _open_verdict_pr(
                    api=api, dataset=dataset,
                    results_path=results_path, report_dir=out_dir,
                    profile=profile, version=version,
                    status=status, summary=summary,
                )
                out(f"  PR opened: {pr_url}")
            except Exception as e:
                out(f"  ! PR creation failed: {type(e).__name__}: {e}")

    persisted = Path("/tmp") / f"hfsp-report-{dataset.replace('/', '_')}"
    if persisted.exists():
        shutil.rmtree(persisted)
    shutil.copytree(out_dir, persisted)

    # Skip dataset-level cache write for incomplete runs. Two cases:
    #  - Cancelled mid-streaming (operator clicked Cancel)
    #  - Unrecoverable plugin-miss abort
    # Either way the merged_results don't represent the dataset — they're
    # the partial output up to where we bailed. Caching them would replay
    # the partial verdict on the next click. Per-zip cache entries from
    # the zips we DID process are still kept (they're keyed on zip_sha,
    # not dataset HEAD, and represent real validation of those zips).
    is_cancelled = bool(results_json.get("cancelled"))
    is_partial = (
        results_json.get("streaming_zips") is not None
        and results_json.get("streaming_processed", 0) < results_json["streaming_zips"]
    )
    if is_cancelled or is_partial:
        out(f"  skipping dataset-level cache write "
            f"({'cancelled' if is_cancelled else 'partial: ' + str(results_json.get('streaming_processed')) + '/' + str(results_json.get('streaming_zips'))})")
    else:
        try:
            # Reuse the pre-resolved HEAD when run() already fetched it.
            # Falls back to a fresh API call only if the caller didn't
            # pass one (e.g. legacy call sites).
            head = dataset_head if dataset_head is not None else api.repo_info(dataset, repo_type="dataset").sha
            key = _cache_key(head, profile, _validator_version(), _foundation_sha())
            _write_cache(dataset, key, {
                "schema_version": 1,
                "dataset": dataset, "dataset_head": head,
                "profile": profile, "validator_version": _validator_version(),
                "foundation_sha": _foundation_sha(),
                "status": status, "summary": summary,
                "results_json": results_json,
                "report_path": str(persisted),
                "cached_at": _now(),
            })
            out(f"  cached result under key={key}")
        except Exception as e:
            out(f"  ! cache write failed ({type(e).__name__}: {e}); ignored")

    return RunResult(
        dataset=dataset, profile=profile, version=version,
        status=status, summary=summary,
        results_json=results_json,
        report_path=persisted, pr_url=pr_url,
    )


def run(
    dataset: str,
    profile: str = "Robot-Body-Runnable",
    version: str = "1.0.0",
    open_pr: bool = False,
    hf_token: str | None = None,
    log: Iterator[str] | None = None,
    submission_id: str = "",
    force: bool = False,
    preliminary: bool = False,
) -> RunResult:
    """Validate a single HF dataset. Yields log lines via the `log` callable.

    The Space's Gradio UI passes a callable that streams lines to the
    output panel; the test harness can pass `print` directly.

    `force=True` bypasses the dataset-level cache — used by manual
    "Validate now" clicks from the dashboard so the operator gets a
    real re-run even if nothing relevant changed. Auto-triggered runs
    (PR webhooks, scheduled re-validation) leave force=False and get
    the cached result when the four-tuple matches.

    `preliminary=True` is a structure-only sweep used by the
    dashboard's Preliminary scan tab:
      - Zip-bundled datasets are scanned (skips the strict-spec
        PKG.NO-ARCHIVES pre-check). Only the first zip is processed.
      - Flat datasets are sliced to the first asset directory before
        validation, so per-asset checks run on one sample asset only.
    """
    out = log or (lambda s: print(s, flush=True))
    flags = []
    if force: flags.append("force")
    if preliminary: flags.append("preliminary")
    flag_str = f" ({', '.join(flags)})" if flags else ""
    out(f"[{_now()}] validating dataset={dataset} profile={profile} v{version}{flag_str}")

    token = hf_token or os.environ.get("HF_TOKEN") or os.environ.get("HUGGING_FACE_HUB_TOKEN")
    api = HfApi(token=token)

    # Resolve the dataset HEAD ONCE up front. Used for: (a) the
    # dataset-level cache key, (b) the per-unit cache key in the flat
    # path, (c) the streaming function's "synthetic zip sha" for the
    # flat unit. Without this, the same metadata was re-fetched from
    # HF up to 4 times per validation.
    dataset_head: str | None = None
    try:
        dataset_head = api.repo_info(dataset, repo_type="dataset").sha
    except Exception as e:
        out(f"  ! could not resolve dataset HEAD ({type(e).__name__}: {e}); cache + drift checks skipped")
    if not force and dataset_head:
        key = _cache_key(dataset_head, profile, _validator_version(), _foundation_sha())
        cached = _read_cache(dataset, key)
        if cached:
            out(f"  cache hit (key={key}, head={dataset_head[:8]}, "
                f"cached_at={cached.get('cached_at')}); returning without re-running")
            return RunResult(
                dataset=dataset, profile=profile, version=version,
                status=cached["status"], summary=cached["summary"],
                results_json=cached["results_json"],
                report_path=Path(cached.get("report_path") or "/tmp"),
                pr_url=None,
            )
        out(f"  cache miss (key={key}, head={dataset_head[:8]}); running validator")

    with tempfile.TemporaryDirectory(prefix=f"hfsp-{dataset.replace('/', '_')}-") as td:
        work = Path(td)
        out(f"  workdir: {work}")

        # Single validation path: every dataset (zip-bundled or flat)
        # goes through _validate_zip_streaming, which uses a persistent
        # daemon pool + per-unit cache + cancel signaling + live
        # progress. Flat datasets pre-materialize once via
        # snapshot_download and pass the dir as flat_target.
        prog_path = progress_path_for(submission_id) if submission_id else None
        if prog_path:
            PROGRESS_DIR.mkdir(parents=True, exist_ok=True)
            prog_path.write_text(json.dumps({
                "processed": 0, "total": 0, "current": None,
                "started_at": _now(), "updated_at": _now(),
                "state": "starting",
            }))

        # Pre-probe: ask the API which case we're in before downloading.
        flat_target: Path | None = None
        try:
            probe_zip_entries = _list_dataset_zips(api, dataset, token)
        except Exception as e:
            out(f"  ! zip probe failed ({type(e).__name__}: {e}); assuming flat")
            probe_zip_entries = []

        # STRICT PRE-CHECK at the dataset level: zips are not in any
        # spec's allowlist (foundation AA.002 lists only USD/image/
        # audio extensions; SDK packaging-spec.md describes an unpacked
        # layout). Zip-bundled datasets fail PKG.NO-ARCHIVES at the
        # dataset listing stage — we never download anything. Partner
        # must repackage as unpacked.
        #
        # Exception: preliminary scan. The dashboard's Preliminary
        # scan tab wants a structure check on a sample asset even when
        # the dataset is zip-bundled, so the strict pre-check is
        # bypassed in that mode and the first zip is streamed.
        if probe_zip_entries and preliminary:
            # Preliminary scan + zip-bundled: stream just the first zip
            # through _validate_zip_streaming and return. Skips the
            # PKG.NO-ARCHIVES strict-fail block AND the flat
            # snapshot_download path entirely — we only want one zip's
            # worth of validator work.
            probe_zip_entries = probe_zip_entries[:1]
            out(f"  preliminary mode: streaming first zip only "
                f"({probe_zip_entries[0][0]})")
            streamed = _validate_zip_streaming(
                api=api, dataset=dataset, token=token, work=work,
                profile=profile, version=version,
                progress_file=prog_path, out=out, force=force,
                submission_id=submission_id,
                flat_target=None,
                prefetched_zip_entries=probe_zip_entries,
                prefetched_dataset_head=dataset_head,
            )
            out_dir = work / "out"
            out_dir.mkdir(parents=True, exist_ok=True)
            results_path = out_dir / "results.json"
            if streamed is None:
                return RunResult(
                    dataset=dataset, profile=profile, version=version,
                    status="error",
                    summary="validator produced no result (preliminary zip path returned None)",
                    results_json={}, report_path=out_dir, pr_url=None,
                )
            results_path.write_text(json.dumps(streamed), encoding="utf-8")
            status, summary = _summarize(streamed)
            out(f"  {status.upper()}: {summary}")
            return _finalize_run(
                dataset=dataset, profile=profile, version=version,
                results_json=streamed, status=status, summary=summary,
                out_dir=out_dir, api=api, token=token, open_pr=open_pr,
                results_path=results_path, out=out,
                dataset_head=dataset_head,
            )

        if probe_zip_entries:
            out(f"  PRELIMINARY FAILURE: dataset ships {len(probe_zip_entries)} "
                f"zip archive(s); zips are not in the spec's allowlist. "
                f"Skipping download + validation entirely.")
            zip_issues = []
            for zip_rel, _zip_sha in probe_zip_entries:
                zip_issues.append({
                    "code": "PKG.NO-ARCHIVES",
                    "severity": "failure",
                    "path": zip_rel,
                    "spec_url": ("https://github.com/NVIDIA-dev/"
                                 "simready-oem-library-pm/blob/main/"
                                 "docs/sdk/packaging-spec.md"
                                 "#folder-structure"),
                    "msg": ("SimReady datasets must be delivered as "
                            "unpacked directories — neither foundation "
                            "AA.002 nor the SDK packaging spec lists "
                            ".zip as an accepted file type."),
                })
            by_path: dict[str, list[dict]] = {}
            for f in zip_issues:
                by_path.setdefault(f["path"], []).append(f)
            results = []
            for rel, issues_here in by_path.items():
                results.append({
                    "asset_path": f"{dataset}/{rel}",
                    "rel_path": rel,
                    "validation_status": "fail",
                    "profile": profile,
                    "profile_version": version,
                    "issues": issues_here,
                    "passed": False,
                })
            results_json = {
                "schema_version": 1,
                "profile": profile,
                "profile_version": version,
                "results": results,
                "preliminary_findings": zip_issues,
                "preliminary_check_failed": True,
            }
            out_dir = work / "out"
            out_dir.mkdir(parents=True, exist_ok=True)
            results_path = out_dir / "results.json"
            results_path.write_text(json.dumps(results_json, indent=2),
                                    encoding="utf-8")
            status, summary = _summarize(results_json)
            out(f"  {status.upper()}: {summary}")
            return _finalize_run(
                dataset=dataset, profile=profile, version=version,
                results_json=results_json, status=status, summary=summary,
                out_dir=out_dir, api=api, token=token, open_pr=open_pr,
                results_path=results_path, out=out,
                dataset_head=dataset_head,
            )

        # No zips: standard flat-dataset path. Materialize via
        # snapshot_download, then hand off to _validate_zip_streaming
        # (which treats the unpacked dir as a single "unit" and runs
        # the daemon-pool + per-unit cache + cancel + progress code).
        local = work / "raw"
        local.mkdir(parents=True, exist_ok=True)
        out(f"  $ snapshot_download {dataset} ignore_patterns={list(HF_DOWNLOAD_EXCLUDES)}")
        snapshot_download(
            repo_id=dataset,
            repo_type="dataset",
            local_dir=str(local),
            ignore_patterns=list(HF_DOWNLOAD_EXCLUDES),
            token=token,
        )
        flat_target = _wrap_layout_for_validator(local, work)
        out(f"  validator target: {flat_target}")

        # Preliminary scan + flat dataset: slice flat_target down to its
        # first asset directory (one level deep, contains at least one
        # .usd/.usda/.usdc file) so per-asset validation only runs on
        # one sample asset. Preliminary structure checks (PKG.01, .06,
        # AA.002) still surface from that single asset's vantage; for a
        # fuller sweep the operator promotes the partner out of the
        # Preliminary scan tab.
        if preliminary:
            try:
                _USD_EXTS = (".usd", ".usda", ".usdc")
                first_asset_dir = None
                for child in sorted(flat_target.iterdir()):
                    if not child.is_dir():
                        continue
                    if any(p.suffix.lower() in _USD_EXTS for p in child.rglob("*")
                           if p.is_file()):
                        first_asset_dir = child
                        break
                if first_asset_dir is not None:
                    slim = work / "preliminary-sample"
                    slim.mkdir(parents=True, exist_ok=True)
                    target = slim / first_asset_dir.name
                    if not target.exists():
                        import shutil
                        shutil.copytree(first_asset_dir, target, symlinks=True)
                    flat_target = slim
                    out(f"  preliminary mode: sliced flat target to "
                        f"first asset dir '{first_asset_dir.name}'")
                else:
                    out("  preliminary mode: no asset directory found to "
                        "slice; running on full flat target")
            except Exception as e:
                out(f"  ! preliminary slice failed ({type(e).__name__}: "
                    f"{e}); running on full flat target")

        streamed = _validate_zip_streaming(
            api=api, dataset=dataset, token=token, work=work,
            profile=profile, version=version,
            progress_file=prog_path, out=out, force=force,
            submission_id=submission_id,
            flat_target=flat_target,
            prefetched_zip_entries=probe_zip_entries,
            prefetched_dataset_head=dataset_head,
        )

        out_dir = work / "out"
        out_dir.mkdir(parents=True, exist_ok=True)
        results_path = out_dir / "results.json"
        if streamed is None:
            # Should not happen — either zip path or flat path always
            # returns a dict. Defensive bail-out.
            return RunResult(
                dataset=dataset, profile=profile, version=version,
                status="error",
                summary="validator produced no result (unified streaming path returned None)",
                results_json={}, report_path=out_dir, pr_url=None,
            )
        # Event-driven foundation spec drift check. One GitHub API
        # call per run; soft-fails so network hiccups don't block
        # validation. The dashboard renders a notice if drifted=true.
        drift = _check_foundation_spec_drift()
        if drift.get("checked"):
            if drift.get("drifted"):
                out(f"  ⚠ spec drift: foundation HEAD={drift.get('current_sha', '')[:8]} "
                    f"@ {drift.get('current_at', '')}, last synced "
                    f"{drift.get('last_sync_sha', '')[:8]} @ {drift.get('last_sync_at', '')} "
                    f"— hardcoded rules may be stale; run spec-sync to refresh")
            else:
                out(f"  spec sync: in sync with foundation HEAD "
                    f"{drift.get('current_sha', '')[:8]}")
        else:
            out(f"  spec sync: skipped ({drift.get('reason', 'unknown')})")
        streamed["spec_drift"] = drift

        results_path.write_text(json.dumps(streamed), encoding="utf-8")
        results_json = streamed
        status, summary = _summarize(results_json)
        out(f"  {status.upper()}: {summary}")
        return _finalize_run(
            dataset=dataset, profile=profile, version=version,
            results_json=results_json, status=status, summary=summary,
            out_dir=out_dir, api=api, token=token, open_pr=open_pr,
            results_path=results_path, out=out,
            dataset_head=dataset_head,
        )