harness/localize_sft_data.py

#!/usr/bin/env python3
from __future__ import annotations

import argparse
from collections import Counter, deque
import concurrent.futures
from typing import TYPE_CHECKING, Callable, Iterable, Mapping, Sequence
from dataclasses import dataclass, field
from datetime import datetime, timezone
import hashlib
import json
import multiprocessing
import os
from pathlib import Path
import queue
import socket
import sqlite3
import struct
import subprocess
import sys
import tempfile
import threading
import time
import traceback
import uuid

if TYPE_CHECKING:
    from localize_sft_core import TaskRecord

try:
    from tqdm.auto import tqdm
except ImportError:  # pragma: no cover - fallback for minimal environments
    class tqdm:  # type: ignore[no-redef]
        def __init__(
            self,
            *,
            total: int = 0,
            desc: str = "",
            unit: str = "",
            dynamic_ncols: bool = False,
        ) -> None:
            self.total = total
            self.desc = desc
            self.unit = unit
            self.current = 0
            self.postfix = ""
            self._print()

        def update(self, n: int = 1) -> None:
            self.current += n
            self._print()

        def set_postfix_str(self, text: str) -> None:
            self.postfix = text
            self._print()

        def close(self) -> None:
            print(file=sys.stderr, flush=True)

        def _print(self) -> None:
            suffix = f" {self.postfix}" if self.postfix else ""
            total = self.total if self.total > 0 else "?"
            print(
                f"\r{self.desc}: {self.current}/{total} {self.unit}{suffix}",
                end="",
                file=sys.stderr,
                flush=True,
            )

try:
    from localizer_path import add_repo_paths
except ModuleNotFoundError:
    from localizer.localizer_path import add_repo_paths

add_repo_paths()

from localize_sft_core import (  # noqa: E402
    BifrostChunkExtractor,
    BuildDbStats,
    CHUNKLESS_FILE_SUMMARY_CODEUNIT,
    DEFAULT_CLONES_DIR,
    DEFAULT_COMMITS_ROOT,
    DEFAULT_EMBEDDINGS_DIR,
    DEFAULT_EXTRACT_WORKERS,
    DEFAULT_NEGATIVES_PER_ROW,
    DEFAULT_SCAN_TOP_K,
    DEFAULT_TASKS_DIR,
    DEFAULT_TASK_LIMIT_PER_LANGUAGE,
    DEFAULT_TASK_LIMIT_PER_REPO,
    GRANITE_MODEL,
    HARD_NEGATIVES,
    MAX_SEQ_LENGTH,
    compare_eval_reports,
    configure_cuda_visibility,
    assert_query_side_compatible,
    active_revisions_for_records,
    active_revisions_for_repo_records,
    build_chunkless_file_chunks_with_client,
    build_chunkless_passage_text,
    doctor_checks,
    ensure_cache_compatible,
    evaluate_records,
    export_training_examples,
    filter_records_by_base_revision_window,
    group_tasks_by_repo,
    has_chunkless_augment_marker,
    indexed_revisions,
    init_embeddings_db,
    insert_chunk_rows,
    iter_missing_vector_work_pages,
    make_chunkless_file_chunk,
    materialize_worktree,
    read_cache_manifest,
    make_bifrost_chunk_extractor,
    make_embedder_with_manifest,
    mark_chunkless_augment_done,
    mine_ready_negatives,
    planned_vector_pipeline_pages,
    planned_remaining_unique_revisions,
    print_doctor,
    prune_orphan_chunks_for_repo,
    read_selection,
    selection_content_hash,
    selection_sidecar_payload,
    set_selection_hash_metadata,
    short_selection_hash,
    reconcile_helper_path,
    refresh_repo_positives_worker,
    revision_chunk_index,
    revision_chunk_paths_set,
    load_revision_vector_matrix,
    requested_repo_filter,
    select_tasks,
    stream_recovered_repos,
    tracked_source_files,
    verify_selection_hash_against_dbs,
    vector_cache_dir,
    write_missing_vectors,
    write_eval_report,
    write_selection,
    build_repo_db,
)


def export_train_repo_worker(
    repo: str,
    records: Sequence[object],
    db_dir: Path,
    negative_files: Sequence[Path],
    output: Path,
    split: str,
    negatives_per_row: int,
    max_positives_per_task: int,
    selection_hash: str | None,
) -> dict[str, object]:
    stats = export_training_examples(
        records,  # type: ignore[arg-type]
        db_dir=db_dir,
        negative_files=negative_files,
        output=output,
        split=split,
        negatives_per_row=negatives_per_row,
        max_positives_per_task=max_positives_per_task,
    )
    return {
        "repo": repo,
        "output": str(output),
        "selection_hash": selection_hash,
        "rows": stats.rows,
        "tasks_exported": stats.tasks_exported,
        "skipped_no_positive": stats.skipped_no_positive,
        "skipped_too_few_negatives": stats.skipped_too_few_negatives,
        "positives_per_task_max": stats.positives_per_task_max,
        "mined_negatives_per_task_max": stats.mined_negatives_per_task_max,
        "target_files_per_task_max": stats.target_files_per_task_max,
        "positive_chunks_per_target_file_max": stats.positive_chunks_per_target_file_max,
        "exported_gold_file_groups_per_task_max": stats.exported_gold_file_groups_per_task_max,
        "missing_gold_file_groups_per_task_max": stats.missing_gold_file_groups_per_task_max,
        "task_weight_exported_sum_min": stats.task_weight_exported_sum_min,
        "fraction_tasks_with_partial_positive_coverage": stats.fraction_tasks_with_partial_positive_coverage,
        "positive_slot_old_hunk": stats.positive_slot_old_hunk,
        "positive_slot_class_summary_fallback": stats.positive_slot_class_summary_fallback,
        "positive_slot_file_summary_fallback": stats.positive_slot_file_summary_fallback,
        "valid_positive_slots_1": stats.valid_positive_slots_1,
        "valid_positive_slots_2": stats.valid_positive_slots_2,
        "valid_positive_slots_3": stats.valid_positive_slots_3,
        "valid_positive_slots_4": stats.valid_positive_slots_4,
        "valid_positive_negative_pairs": stats.valid_positive_negative_pairs,
        "rows_with_3_plus_old_hunks": stats.rows_with_3_plus_old_hunks,
        "rows_with_4_old_hunks": stats.rows_with_4_old_hunks,
        "fallback_only_rows": stats.fallback_only_rows,
    }


def concatenate_jsonl_shards(shard_paths: Sequence[Path], output: Path) -> None:
    output.parent.mkdir(parents=True, exist_ok=True)
    with output.open("wb") as out:
        for shard_path in shard_paths:
            if not shard_path.exists():
                continue
            with shard_path.open("rb") as src:
                while True:
                    chunk = src.read(1024 * 1024)
                    if not chunk:
                        break
                    out.write(chunk)


def aggregate_parallel_export_stats(rows: Sequence[dict[str, object]], output: Path) -> dict[str, object]:
    int_fields = [
        "rows",
        "tasks_exported",
        "skipped_no_positive",
        "skipped_too_few_negatives",
        "positive_slot_old_hunk",
        "positive_slot_class_summary_fallback",
        "positive_slot_file_summary_fallback",
        "valid_positive_slots_1",
        "valid_positive_slots_2",
        "valid_positive_slots_3",
        "valid_positive_slots_4",
        "valid_positive_negative_pairs",
        "rows_with_3_plus_old_hunks",
        "rows_with_4_old_hunks",
        "fallback_only_rows",
    ]
    max_fields = [
        "positives_per_task_max",
        "mined_negatives_per_task_max",
        "target_files_per_task_max",
        "positive_chunks_per_target_file_max",
        "exported_gold_file_groups_per_task_max",
        "missing_gold_file_groups_per_task_max",
    ]
    summary: dict[str, object] = {"output": str(output), "repo_workers": len(rows)}
    for field_name in int_fields:
        summary[field_name] = sum(int(row.get(field_name, 0)) for row in rows)
    for field_name in max_fields:
        summary[field_name] = max((int(row.get(field_name, 0)) for row in rows), default=0)
    mins = [
        float(row["task_weight_exported_sum_min"])
        for row in rows
        if float(row.get("task_weight_exported_sum_min", 0.0)) > 0.0
    ]
    summary["task_weight_exported_sum_min"] = min(mins) if mins else 0.0
    exported = int(summary["tasks_exported"])
    if exported:
        partial_tasks = sum(
            float(row.get("fraction_tasks_with_partial_positive_coverage", 0.0))
            * int(row.get("tasks_exported", 0))
            for row in rows
        )
        summary["fraction_tasks_with_partial_positive_coverage"] = partial_tasks / exported
    else:
        summary["fraction_tasks_with_partial_positive_coverage"] = 0.0
    summary["note"] = "parallel export aggregates exact counts/maxima; percentile fields are omitted"
    return summary


def _warn_stderr(message: str) -> None:
    print(message, file=sys.stderr)


def write_selection_artifacts(records: Sequence[object], legacy_output: Path) -> tuple[Path, Path, dict[str, object]]:
    typed_records = list(records)
    selection_hash = selection_content_hash(typed_records)  # type: ignore[arg-type]
    short_hash = short_selection_hash(selection_hash)
    created_at = datetime.now(timezone.utc).replace(microsecond=0).isoformat().replace("+00:00", "Z")
    dated_name = f"selection-{created_at[:10].replace('-', '')}-{short_hash}.jsonl"
    canonical_path = legacy_output.parent / dated_name
    sidecar_path = legacy_output.parent / f"{canonical_path.stem}.meta.json"
    write_selection(typed_records, canonical_path)  # type: ignore[arg-type]
    sidecar = selection_sidecar_payload(typed_records, created_at=created_at)  # type: ignore[arg-type]
    sidecar_path.write_text(json.dumps(sidecar, indent=2, sort_keys=True), encoding="utf-8")
    write_selection(typed_records, legacy_output)  # type: ignore[arg-type]
    return canonical_path, sidecar_path, sidecar


@dataclass(frozen=True)
class GpuWorkerSpec:
    gpu: str
    repo_workers: int
    batch_size: int | None = None


def parse_single_gpu_worker(value: str) -> GpuWorkerSpec:
    parts = value.split(":")
    if len(parts) not in (1, 2) or not parts[0]:
        raise argparse.ArgumentTypeError("expected GPU[:BATCH_SIZE]")
    try:
        batch_size = int(parts[1]) if len(parts) == 2 else None
    except ValueError as exc:
        raise argparse.ArgumentTypeError("batch size must be an integer") from exc
    if batch_size is not None and batch_size < 1:
        raise argparse.ArgumentTypeError("batch size must be >= 1")
    return GpuWorkerSpec(parts[0], 1, batch_size)


@dataclass(frozen=True)
class NativeGpuWorkerSpec:
    gpu: str
    batch_size: int | None = None


@dataclass(frozen=True)
class NativeWorkerRuntime:
    key: str
    spec: NativeGpuWorkerSpec


def parse_native_gpu_worker(value: str) -> NativeGpuWorkerSpec:
    parts = value.split(":")
    if len(parts) not in (1, 2) or not parts[0]:
        raise argparse.ArgumentTypeError("expected GPU[:BATCH_SIZE]")
    try:
        batch_size = int(parts[1]) if len(parts) == 2 else None
    except ValueError as exc:
        raise argparse.ArgumentTypeError("batch size must be an integer") from exc
    if batch_size is not None and batch_size < 1:
        raise argparse.ArgumentTypeError("batch size must be >= 1")
    return NativeGpuWorkerSpec(parts[0], batch_size)


def parse_repo_shard(value: str) -> tuple[int, int]:
    try:
        index_text, count_text = value.split("/", 1)
        index = int(index_text)
        count = int(count_text)
    except ValueError as exc:
        raise argparse.ArgumentTypeError("--repo-shard must be INDEX/COUNT") from exc
    if count <= 0:
        raise argparse.ArgumentTypeError("--repo-shard count must be positive")
    if index < 0 or index >= count:
        raise argparse.ArgumentTypeError("--repo-shard index must satisfy 0 <= index < count")
    return index, count


def repo_shard_index(repo: str, shard_count: int) -> int:
    digest = hashlib.blake2b(repo.encode("utf-8"), digest_size=8).digest()
    return int.from_bytes(digest, byteorder="big") % shard_count


def filter_records_by_repo_shard(
    records: Sequence[TaskRecord],
    repo_shard: tuple[int, int] | None,
) -> list[TaskRecord]:
    if repo_shard is None:
        return list(records)
    shard_index, shard_count = repo_shard
    return [
        record
        for record in records
        if repo_shard_index(record.repo, shard_count) == shard_index
    ]


def _native_worker_runtimes(workers: Sequence[NativeGpuWorkerSpec]) -> list[NativeWorkerRuntime]:
    physical_counts = Counter(worker.gpu for worker in workers)
    seen: Counter[str] = Counter()
    runtimes: list[NativeWorkerRuntime] = []
    for spec in workers:
        ordinal = seen[spec.gpu]
        seen[spec.gpu] += 1
        key = spec.gpu if physical_counts[spec.gpu] == 1 else f"{spec.gpu}.{ordinal}"
        runtimes.append(NativeWorkerRuntime(key, spec))
    return runtimes


def selected_repos_from_grouped(
    grouped: dict[str, list],
    args: argparse.Namespace,
) -> list[str]:
    repos = sorted(grouped)
    requested = requested_repo_filter(args)
    if requested is not None:
        repos = [repo for repo in repos if repo in requested]
    if args.limit_repos is not None:
        repos = repos[: args.limit_repos]
    return repos


def repo_records_for_args(
    repo: str,
    grouped: dict[str, list],
    args: argparse.Namespace,
) -> list:
    records = grouped[repo]
    if getattr(args, "max_tasks_per_repo", None) is not None:
        records = records[: args.max_tasks_per_repo]
    start_base = getattr(args, "start_base_revision", None)
    end_base = getattr(args, "end_base_revision", None)
    if start_base is None and end_base is None:
        return records
    return filter_records_by_base_revision_window(
        records,
        args.clones_dir / repo,
        start_base_revision=start_base,
        end_base_revision=end_base,
    )


def _short_repo_name(repo: object) -> str:
    text = str(repo)
    if "__" in text:
        return text.rsplit("__", 1)[1]
    if "/" in text:
        return text.rsplit("/", 1)[1]
    return text


def _clone_repos_for_build(
    repos: Sequence[str],
    *,
    clones_dir: Path,
    progress: tqdm,
    clone_workers: int,
    debug: bool,
) -> tuple[list[str], int]:
    available: list[str] = []
    errors = 0
    clone_stream = stream_recovered_repos(
        repos,
        clones_dir,
        workers=clone_workers,
    )
    try:
        for repo, clone_error in clone_stream:
            progress.set_postfix_str(_short_repo_name(repo))
            progress.update(1)
            if clone_error is None:
                available.append(repo)
            else:
                errors += 1
                if debug:
                    print(json.dumps({"repo": repo, "error": clone_error}, sort_keys=True), file=sys.stderr)
                else:
                    print(f"error: {_short_repo_name(repo)} clone failed: {clone_error}", file=sys.stderr)
    finally:
        clone_stream.close()
    return available, errors


def _db_progress_event(queue: object, repo: str, revision: str, completed: int, total_tasks: int) -> None:
    put = getattr(queue, "put")
    put({"event": "revision", "repo": repo, "revision": revision, "completed": completed, "total_tasks": total_tasks})


def _db_log_event(queue: object, repo: str, message: str) -> None:
    put = getattr(queue, "put")
    put({"event": "log", "repo": repo, "message": message})


def _extract_task_progress(message: str) -> tuple[int, int] | None:
    marker = "task="
    index = message.find(marker)
    if index < 0:
        return None
    start = index + len(marker)
    end = start
    while end < len(message) and message[end] not in {" ", "\t", ","}:
        end += 1
    value = message[start:end]
    if "/" not in value:
        return None
    left, right = value.split("/", 1)
    try:
        current = int(left)
        total = int(right)
    except ValueError:
        return None
    if current < 0 or total < 0:
        return None
    return current, total


def build_db_worker_entry(
    records: Sequence,
    *,
    clones_dir: Path,
    db_dir: Path,
    selection_hash: str | None,
    bifrost_library: Path | None,
    resume: bool,
    extract_workers: int,
    progress_queue: object | None,
    bifrost_cache_dir: Path | None = None,
    is_big: bool = False,
    big_repo_gate: object | None = None,
    revision_overrides: Mapping[str, str] | None = None,
) -> BuildDbStats:
    repo = records[0].repo if records else "<empty>"
    # Memory packing: a big repo holds extract_workers heavy bifrost analyzers. The
    # cross-process gate (Semaphore) caps how many big repos run concurrently so the
    # many cheap small repos can fill repo_workers without ever stacking two giants
    # (which OOM-stalls). Small repos never touch the gate.
    gate = big_repo_gate if (is_big and big_repo_gate is not None) else None
    if gate is not None:
        gate.acquire()
    previous_bifrost_cache = os.environ.get("BIFROST_CACHE_DIR")
    if bifrost_cache_dir is not None:
        bifrost_cache_dir.mkdir(parents=True, exist_ok=True)
        os.environ["BIFROST_CACHE_DIR"] = str(bifrost_cache_dir)
    try:
        stats = build_repo_db(
            records,
            clones_dir=clones_dir,
            embeddings_dir=db_dir,
            selection_hash=selection_hash,
            chunk_extractor=make_bifrost_chunk_extractor(bifrost_library),
            progress=(
                (lambda message, current_repo=repo: _db_log_event(progress_queue, current_repo, message))
                if progress_queue is not None
                else None
            ),
            on_revision=(
                (lambda current_repo, revision, completed, total_tasks: _db_progress_event(
                    progress_queue, current_repo, revision, completed, total_tasks
                ))
                if progress_queue is not None
                else None
            ),
            resume=resume,
            extract_workers=extract_workers,
            revision_overrides=revision_overrides,
        )
        return stats
    except Exception as exc:
        error_text = traceback.format_exc()
        return BuildDbStats(
            repo=repo,
            tasks=len(records),
            revisions=0,
            chunks=0,
            positives=0,
            skipped=True,
            error=error_text or repr(exc),
        )
    finally:
        if bifrost_cache_dir is not None:
            if previous_bifrost_cache is None:
                os.environ.pop("BIFROST_CACHE_DIR", None)
            else:
                os.environ["BIFROST_CACHE_DIR"] = previous_bifrost_cache
        if gate is not None:
            gate.release()


def _load_seeded_chunkless_rows(
    seed_dir: Path | None,
    repo: str,
) -> dict[tuple[str, str], str]:
    if seed_dir is None:
        return {}
    path = seed_dir / repo / "synthetic_rows.jsonl"
    if not path.is_file():
        return {}
    seeded: dict[tuple[str, str], str] = {}
    with path.open(encoding="utf-8") as handle:
        for line_number, raw_line in enumerate(handle, start=1):
            line = raw_line.strip()
            if not line:
                continue
            row = json.loads(line)
            if not isinstance(row, dict):
                raise ValueError(f"{path}:{line_number}: expected object")
            revision = str(row.get("revision", ""))
            file_path = str(row.get("path", ""))
            text = str(row.get("text", ""))
            if revision and file_path and text:
                seeded[(revision, file_path)] = text
    return seeded


@dataclass
class _ChunklessMemoEntry:
    event: threading.Event
    text: str = ""
    ready: bool = False


class _ChunklessContentMemo:
    def __init__(self) -> None:
        self._lock = threading.Lock()
        self._entries: dict[str, _ChunklessMemoEntry] = {}
        self.memo_hits = 0
        self.bifrost_calls = 0

    def prime(self, content_hash: str, text: str) -> None:
        with self._lock:
            entry = self._entries.get(content_hash)
            if entry is None:
                ready = threading.Event()
                ready.set()
                self._entries[content_hash] = _ChunklessMemoEntry(
                    event=ready,
                    text=text,
                    ready=True,
                )
                return
            if not entry.ready:
                entry.text = text
                entry.ready = True
                entry.event.set()

    def get_or_build(self, content_hash: str, builder: Callable[[], str]) -> str:
        wait_for: threading.Event | None = None
        with self._lock:
            entry = self._entries.get(content_hash)
            if entry is not None:
                if entry.ready:
                    self.memo_hits += 1
                    return entry.text
                wait_for = entry.event
            else:
                wait_for = threading.Event()
                self._entries[content_hash] = _ChunklessMemoEntry(event=wait_for)
                self.bifrost_calls += 1
                wait_for = None
        if wait_for is not None:
            wait_for.wait()
            with self._lock:
                entry = self._entries[content_hash]
                self.memo_hits += 1
                return entry.text
        text = builder()
        with self._lock:
            entry = self._entries[content_hash]
            entry.text = text
            entry.ready = True
            entry.event.set()
        return text


def _worktree_file_hash(worktree: Path, path: str) -> str | None:
    file_path = worktree / path
    try:
        data = file_path.read_bytes()
    except OSError:
        return None
    return hashlib.sha256(data).hexdigest()


def _prime_chunkless_seed_memo(
    memo: _ChunklessContentMemo,
    worktree: Path,
    revision: str,
    seeded: dict[tuple[str, str], str],
) -> None:
    for (seed_revision, path), text in seeded.items():
        if seed_revision != revision:
            continue
        content_hash = _worktree_file_hash(worktree, path)
        if content_hash is not None:
            memo.prime(content_hash, text)


def _reconcile_augmented_vectors(
    repo_revisions: Sequence[tuple[str, Sequence[str]]],
    *,
    db_dir: Path,
    vector_cache_dirs: Sequence[Path],
    model: str,
    batch_size: int,
    max_seq_length: int,
    page_size: int,
) -> list[dict[str, object]]:
    if not repo_revisions or not vector_cache_dirs:
        return []
    summaries: list[dict[str, object]] = []
    for cache_dir in vector_cache_dirs:
        on_disk = read_cache_manifest(cache_dir)
        if on_disk is None:
            raise RuntimeError(f"no vector cache manifest at {cache_dir}")
        embed_texts, expected_manifest = make_embedder_with_manifest(
            model,
            batch_size=batch_size,
            max_seq_length=max_seq_length,
            role="passage",
        )
        ensure_cache_compatible(cache_dir, expected_manifest)  # type: ignore[arg-type]
        manifest_digest = expected_manifest.digest()  # type: ignore[union-attr]
        embed_lock = threading.Lock()

        def locked_embed_texts(texts: list[str]) -> list[list[float]]:
            with embed_lock:
                return embed_texts(texts)  # type: ignore[misc]

        def reconcile_repo(repo_revision: tuple[str, Sequence[str]]) -> tuple[str, int]:
            repo, revisions = repo_revision
            written = 0
            for page in iter_missing_vector_work_pages(
                repo,
                revisions,
                db_dir=db_dir,
                vector_cache_dir=cache_dir,
                manifest_digest=manifest_digest,
                page_size=page_size,
            ):
                written += write_missing_vectors(
                    cache_dir,
                    repo,
                    page.chunks,
                    locked_embed_texts,
                    manifest_digest=manifest_digest,
                    batch_size=batch_size,
                    page_size=page_size,
                )
            return repo, written

        repo_counts: dict[str, int] = {}
        # Each reconcile thread spawns a helper whose `GROUP BY/ORDER BY vector_key` spills a large
        # SQLite sort temp; an os.cpu_count()-sized pool on a high-core box runs dozens concurrently
        # and can fill the temp volume (saw 64-way concurrency blow past a 700G disk). Cap via
        # LOCALIZE_RECONCILE_WORKERS so concurrent spills stay bounded (pair with SQLITE_TMPDIR=/dev/shm).
        default_workers = min(max(1, os.cpu_count() or 1), len(repo_revisions))
        env_cap = os.environ.get("LOCALIZE_RECONCILE_WORKERS")
        workers = min(max(1, int(env_cap)), len(repo_revisions)) if env_cap else default_workers
        with concurrent.futures.ThreadPoolExecutor(max_workers=workers) as pool:
            futures = [pool.submit(reconcile_repo, item) for item in repo_revisions]
            for future in concurrent.futures.as_completed(futures):
                repo, written = future.result()
                repo_counts[repo] = written
        summaries.append(
            {
                "vector_cache_dir": str(cache_dir),
                "manifest_digest": manifest_digest,
                "vectors_written": sum(repo_counts.values()),
                "repo_counts": dict(sorted(repo_counts.items())),
            }
        )
    return summaries


def augment_chunkless_repo_worker(
    repo: str,
    *,
    db_dir: Path,
    clones_dir: Path,
    seed_dir: Path | None,
    seed_only: bool,
    revisions: Sequence[str] | None,
    summary_workers: int,
    vector_cache_dirs: Sequence[Path],
    model: str,
    batch_size: int,
    max_seq_length: int,
    page_size: int,
    dry_run: bool,
) -> dict[str, object]:
    repo_path = clones_dir / repo
    db_path = db_dir / repo / "embeddings.db"
    if not repo_path.is_dir():
        raise FileNotFoundError(f"missing clone for {repo}: {repo_path}")
    if not db_path.exists():
        raise FileNotFoundError(f"missing embeddings DB for {repo}: {db_path}")

    seeded = _load_seeded_chunkless_rows(seed_dir, repo)
    conn = init_embeddings_db(db_path)
    conn.execute("PRAGMA journal_mode=WAL")
    conn.execute("PRAGMA synchronous=NORMAL")
    try:
        available_revisions = indexed_revisions(conn)
        selected_revisions = available_revisions if revisions is None else [
            rev for rev in revisions if rev in available_revisions
        ]
        if seed_only:
            seeded_revisions = {revision for revision, _path in seeded}
            selected_revisions = [rev for rev in selected_revisions if rev in seeded_revisions]
        pending_revisions = [rev for rev in selected_revisions if not has_chunkless_augment_marker(conn, rev)]
        summary: dict[str, object] = {
            "repo": repo,
            "revisions_seen": len(selected_revisions),
            "revisions_processed": len(pending_revisions),
            "chunkless_candidates": 0,
            "chunkless_inserted": 0,
            "chunkless_skipped": 0,
            "seeded_rows_used": 0,
            "memo_hits": 0,
            "bifrost_calls": 0,
            "dry_run": dry_run,
            "seed_only": seed_only,
        }
        if dry_run:
            for revision in pending_revisions:
                indexed_paths = revision_chunk_paths_set(conn, revision)
                candidates = [
                    path
                    for path in tracked_source_files(repo_path, revision)
                    if path not in indexed_paths
                ]
                summary["chunkless_candidates"] = int(summary["chunkless_candidates"]) + len(candidates)
            return summary

        extractor = BifrostChunkExtractor()
        content_memo = _ChunklessContentMemo()
        for revision in pending_revisions:
            indexed_paths = revision_chunk_paths_set(conn, revision)
            candidate_paths = [
                path
                for path in tracked_source_files(repo_path, revision)
                if path not in indexed_paths
            ]
            summary["chunkless_candidates"] = int(summary["chunkless_candidates"]) + len(candidate_paths)
            synthetic_chunks = [
                make_chunkless_file_chunk(
                    repo=repo,
                    revision=revision,
                    path=path,
                    text=seeded[(revision, path)],
                )
                for path in candidate_paths
                if (revision, path) in seeded
            ]
            summary["seeded_rows_used"] = int(summary["seeded_rows_used"]) + len(synthetic_chunks)
            unseeded_paths = [path for path in candidate_paths if (revision, path) not in seeded]
            skipped_paths: list[str] = []
            if unseeded_paths and not seed_only:
                with materialize_worktree(repo_path, revision) as worktree:
                    _prime_chunkless_seed_memo(content_memo, worktree, revision, seeded)
                    existing_paths = [path for path in unseeded_paths if (worktree / path).is_file()]
                    with extractor.open_client(worktree) as summary_client:
                        def memoized_passage_builder(path: str) -> str:
                            content_hash = _worktree_file_hash(worktree, path)
                            if content_hash is None:
                                return ""
                            return content_memo.get_or_build(
                                content_hash,
                                lambda: build_chunkless_passage_text(summary_client, path),
                            )

                        built_chunks, skipped_paths = build_chunkless_file_chunks_with_client(
                            summary_client,
                            worktree,
                            repo,
                            revision,
                            existing_paths,
                            summary_workers=summary_workers,
                            passage_builder=memoized_passage_builder,
                        )
                    synthetic_chunks.extend(built_chunks)
            before = int(
                conn.execute(
                    """
                    SELECT COUNT(*)
                    FROM chunk_rows
                    WHERE revision = ?
                      AND codeunit_name = ?
                    """,
                    (revision, CHUNKLESS_FILE_SUMMARY_CODEUNIT),
                ).fetchone()[0]
            )
            conn.execute("BEGIN")
            try:
                insert_chunk_rows(conn, synthetic_chunks, or_ignore=True)
                mark_chunkless_augment_done(conn, revision)
                conn.commit()
            except Exception:
                conn.rollback()
                raise
            after = int(
                conn.execute(
                    """
                    SELECT COUNT(*)
                    FROM chunk_rows
                    WHERE revision = ?
                      AND codeunit_name = ?
                    """,
                    (revision, CHUNKLESS_FILE_SUMMARY_CODEUNIT),
                ).fetchone()[0]
            )
            summary["chunkless_inserted"] = int(summary["chunkless_inserted"]) + max(0, after - before)
            summary["chunkless_skipped"] = int(summary["chunkless_skipped"]) + len(skipped_paths)
        summary["memo_hits"] = content_memo.memo_hits
        summary["bifrost_calls"] = content_memo.bifrost_calls
        summary["selected_revisions"] = list(selected_revisions)
        return summary
    finally:
        conn.close()


def _plan_build_db_repo(
    repo: str,
    *,
    grouped: dict[str, list],
    args: argparse.Namespace,
) -> tuple[str, list, int]:
    records_for_repo = repo_records_for_args(repo, grouped, args)
    count, _remaining = planned_remaining_unique_revisions(
        records_for_repo,
        repo_path=args.clones_dir / repo,
        db_path=args.db_dir / repo / "embeddings.db",
        resume=not args.no_resume,
    )
    return repo, records_for_repo, count


def _plan_build_db_repos(
    repos: Sequence[str],
    *,
    grouped: dict[str, list],
    args: argparse.Namespace,
) -> tuple[dict[str, list], dict[str, int]]:
    repo_records: dict[str, list] = {}
    revision_counts: dict[str, int] = {}
    if not repos:
        return repo_records, revision_counts
    workers = max(1, min(len(repos), args.repo_workers))
    progress = tqdm(total=len(repos), desc="plan db", unit="repo", dynamic_ncols=True)
    try:
        with concurrent.futures.ThreadPoolExecutor(max_workers=workers) as pool:
            futures = {
                pool.submit(_plan_build_db_repo, repo, grouped=grouped, args=args): repo
                for repo in repos
            }
            for future in concurrent.futures.as_completed(futures):
                repo = futures[future]
                progress.set_postfix_str(_short_repo_name(repo))
                planned_repo, records_for_repo, count = future.result()
                if records_for_repo:
                    repo_records[planned_repo] = records_for_repo
                    revision_counts[planned_repo] = count
                progress.update(1)
    finally:
        progress.close()
    return repo_records, revision_counts


def _repo_tree_bytes(repo_path: Path) -> int:
    total = 0
    for root, dirs, files in os.walk(repo_path):
        if ".git" in dirs:
            dirs.remove(".git")
        for name in files:
            try:
                total += os.path.getsize(os.path.join(root, name))
            except OSError:
                pass
    return total


def _pack_repos_by_size(
    repos: Sequence[str],
    *,
    clones_dir: Path,
    big_bytes: int,
) -> tuple[list[str], set[str]]:
    """Order repos so memory-heavy 'big' repos are spread evenly among the many cheap
    'small' ones, and return the set of bigs. The extraction ProcessPool is FIFO, so this
    list order *is* the schedule: spacing bigs far apart keeps worker slots filled with
    small repos instead of blocking on the big-repo gate. Pairs with the Semaphore guard
    in build_db_worker_entry, which is the hard 'at most N bigs in flight' guarantee."""
    sizes: dict[str, int] = {}
    with concurrent.futures.ThreadPoolExecutor(max_workers=16) as pool:
        fut = {pool.submit(_repo_tree_bytes, clones_dir / r): r for r in repos}
        for f in concurrent.futures.as_completed(fut):
            sizes[fut[f]] = f.result()
    bigs = sorted((r for r in repos if sizes.get(r, 0) >= big_bytes), key=lambda r: -sizes[r])
    smalls = sorted((r for r in repos if sizes.get(r, 0) < big_bytes), key=lambda r: -sizes[r])
    if not bigs:
        return smalls, set()
    gap = max(1, len(smalls) // (len(bigs) + 1))
    ordered: list[str] = []
    bi = 0
    for i, repo in enumerate(smalls):
        if i % gap == 0 and bi < len(bigs):
            ordered.append(bigs[bi])
            bi += 1
        ordered.append(repo)
    ordered.extend(bigs[bi:])
    return ordered, set(bigs)


def _send_frame(sock: socket.socket, payload: dict[str, object]) -> None:
    data = json.dumps(payload, separators=(",", ":")).encode("utf-8")
    sock.sendall(struct.pack("<I", len(data)) + data)


def _recv_exact(sock: socket.socket, size: int) -> bytes:
    chunks: list[bytes] = []
    remaining = size
    while remaining:
        chunk = sock.recv(remaining)
        if not chunk:
            raise EOFError("native worker closed the socket")
        chunks.append(chunk)
        remaining -= len(chunk)
    return b"".join(chunks)


def _recv_frame(sock: socket.socket) -> dict[str, object]:
    header = _recv_exact(sock, 4)
    size = struct.unpack("<I", header)[0]
    if size == 0:
        return {}
    return json.loads(_recv_exact(sock, size).decode("utf-8"))


class NativeGpuWorker:
    def __init__(
        self,
        *,
        helper: Path,
        worker_key: str,
        gpu: NativeGpuWorkerSpec,
        vector_cache_dir: Path,
        manifest_digest: str,
        batch_size: int,
        max_seq_length: int,
        model: str,
        target_padded_tokens: int,
        target_attention_tokens: int,
        encode_dtype: str,
        attn_implementation: str,
    ) -> None:
        self.helper = helper
        self.worker_key = worker_key
        self.gpu = gpu
        self.vector_cache_dir = vector_cache_dir
        self.manifest_digest = manifest_digest
        self.batch_size = gpu.batch_size if gpu.batch_size is not None else batch_size
        self.max_seq_length = max_seq_length
        self.model = model
        self.target_padded_tokens = target_padded_tokens
        self.target_attention_tokens = target_attention_tokens
        self.encode_dtype = encode_dtype
        self.attn_implementation = attn_implementation
        self.socket_path: Path | None = None
        self.process: subprocess.Popen[str] | None = None
        self.sock: socket.socket | None = None
        self.lock = threading.Lock()
        self.reader: threading.Thread | None = None
        self.events: "queue.Queue[dict[str, object]]" = queue.Queue()

    def start(self, temp_dir: Path) -> None:
        safe_key = self.worker_key.replace("/", "_").replace(":", "_").replace(".", "_")
        self.socket_path = temp_dir / f"native-gpu-{safe_key}.sock"
        try:
            self.socket_path.unlink()
        except FileNotFoundError:
            pass
        command = [
            # Launch the worker with the SAME interpreter as the parent build
            # (e.g. .venv312, which already has torch/transformers/bifrost) rather
            # than the helper's `uv run` shebang, which re-syncs the entire fat
            # localizer env (PyQt6/scipy/CUDA) from a cold cache and blows the
            # connect deadline. Inheriting sys.executable keeps worker and parent
            # on one proven env.
            sys.executable,
            str(self.helper),
            "gpu-worker",
            "--gpu",
            self.gpu.gpu,
            "--socket",
            str(self.socket_path),
            "--batch-size",
            str(self.batch_size),
            "--max-seq-length",
            str(self.max_seq_length),
            "--model",
            self.model,
            "--target-padded-tokens",
            str(self.target_padded_tokens),
            "--target-attention-tokens",
            str(self.target_attention_tokens),
            "--encode-dtype",
            self.encode_dtype,
            "--attn-implementation",
            self.attn_implementation,
        ]
        self.process = subprocess.Popen(command, text=True)
        deadline = time.monotonic() + 30.0
        sock = socket.socket(socket.AF_UNIX, socket.SOCK_STREAM)
        while True:
            try:
                sock.connect(str(self.socket_path))
                break
            except OSError:
                if self.process.poll() is not None:
                    raise RuntimeError(f"native gpu worker {self.gpu.gpu} exited with code {self.process.returncode}")
                if time.monotonic() >= deadline:
                    self.process.terminate()
                    raise RuntimeError(f"timed out connecting to native gpu worker {self.gpu.gpu}")
                time.sleep(0.05)
        self.sock = sock
        self.reader = threading.Thread(target=self._read_events, daemon=True)
        self.reader.start()

    def _read_events(self) -> None:
        assert self.sock is not None
        while True:
            try:
                event = _recv_frame(self.sock)
            except EOFError:
                self.events.put({"event": "worker_closed", "gpu": self.gpu.gpu})
                return
            except Exception as exc:
                self.events.put({"event": "worker_error", "gpu": self.gpu.gpu, "error": repr(exc)})
                return
            if not event:
                continue
            self.events.put(event)
            if event.get("event") == "shutdown_ack":
                return

    def submit_page(self, task: "NativePageTask") -> None:
        cache_dir = vector_cache_dir(self.vector_cache_dir, task.repo, manifest_digest=self.manifest_digest)
        with self.lock:
            if self.sock is None:
                raise RuntimeError(f"native gpu worker {self.gpu.gpu} is not connected")
            _send_frame(
                self.sock,
                {
                    "event": "submit_page",
                    "repo": task.repo,
                    "revision": task.revision,
                    "page": task.page,
                    "page_count": task.page_count,
                    "batch_size": task.batch_size or self.batch_size,
                    "cache_dir": str(cache_dir),
                    "manifest_digest": self.manifest_digest,
                    "components": task.components,
                    "vectors": task.vectors,
                },
            )

    def shutdown(self) -> int:
        if self.sock is not None:
            try:
                with self.lock:
                    _send_frame(self.sock, {"event": "shutdown"})
            except OSError:
                pass
        if self.reader is not None:
            self.reader.join(timeout=5)
        if self.sock is not None:
            self.sock.close()
        if self.process is None:
            return 0
        try:
            return self.process.wait(timeout=10)
        except subprocess.TimeoutExpired:
            self.process.terminate()
            return self.process.wait(timeout=10)


@dataclass
class NativePageTask:
    repo: str
    revision: str
    page: int
    page_count: int
    components: list[dict[str, object]]
    vectors: list[dict[str, str]]
    work_units: int
    source: str = "build"
    batch_size: int = 0
    excluded_gpus: tuple[str, ...] = ()
    preferred_gpus: tuple[str, ...] = ()
    max_estimated_tokens: int = 0
    started_at: float = 0.0


@dataclass
class BatchTelemetry:
    shards: int = 0
    components: int = 0
    vectors: int = 0
    elapsed_seconds: float = 0.0
    encode_seconds: float = 0.0
    tokens: int = 0
    padded_tokens: int = 0
    attention_tokens: int = 0
    ewma_components_per_second: float = 0.0
    ewma_vectors_per_second: float = 0.0
    ewma_tokens_per_second: float = 0.0
    ewma_padded_tokens_per_second: float = 0.0


@dataclass
class GpuTelemetry:
    gpu: str
    configured_batch_size: int
    autotune_candidates: tuple[int, ...] = ()
    batch_profiles: dict[int, BatchTelemetry] = field(default_factory=dict)
    locked_batch_size: int = 0
    pages: int = 0
    shards: int = 0
    components: int = 0
    components_written: int = 0
    vectors: int = 0
    composed_written: int = 0
    elapsed_seconds: float = 0.0
    encode_seconds: float = 0.0
    read_seconds: float = 0.0
    write_seconds: float = 0.0
    compose_seconds: float = 0.0
    tokens: int = 0
    padded_tokens: int = 0
    attention_tokens: int = 0
    micro_batches: int = 0
    ewma_components_per_second: float = 0.0
    ewma_vectors_per_second: float = 0.0
    ewma_tokens_per_second: float = 0.0
    ewma_padded_tokens_per_second: float = 0.0
    current_batch_size: int = 0
    current_max_seq_length: int = 0
    first_started_at: float = 0.0
    measured_at: float = 0.0
    backoffs: int = 0

    def __post_init__(self) -> None:
        self.current_batch_size = self.configured_batch_size

    @property
    def measured(self) -> bool:
        if self.autotune_candidates:
            return self.locked_batch_size > 0
        return self.shards >= 3

    @property
    def weight(self) -> float:
        if self.ewma_padded_tokens_per_second > 0.0:
            return self.ewma_padded_tokens_per_second
        if self.ewma_tokens_per_second > 0.0:
            return self.ewma_tokens_per_second
        if self.ewma_components_per_second > 0.0:
            return self.ewma_components_per_second * 512.0
        return float(max(1, self.configured_batch_size))


def _batch_autotune_candidates(configured_batch_size: int) -> tuple[int, ...]:
    candidates = {max(1, configured_batch_size)}
    batch = max(1, configured_batch_size)
    while batch > 64:
        batch = max(64, batch // 2)
        candidates.add(batch)
    if configured_batch_size <= 64 and configured_batch_size > 16:
        candidates.add(max(16, configured_batch_size // 2))
    return tuple(sorted(candidates, reverse=True))


def _choose_autotune_batch(profile: GpuTelemetry) -> int:
    if profile.locked_batch_size > 0:
        return profile.locked_batch_size
    if not profile.autotune_candidates:
        return profile.current_batch_size
    for candidate in profile.autotune_candidates:
        if profile.batch_profiles.get(candidate, BatchTelemetry()).shards < 2:
            profile.current_batch_size = candidate
            return candidate
    best = max(
        profile.autotune_candidates,
        key=lambda candidate: profile.batch_profiles.get(candidate, BatchTelemetry()).ewma_components_per_second,
    )
    profile.locked_batch_size = best
    profile.current_batch_size = best
    return best


def _pop_next_eligible_task(
    pending: "deque[NativePageTask]",
    worker: NativeGpuWorker,
) -> NativePageTask | None:
    blocked: deque[NativePageTask] = deque()
    while pending:
        candidate = pending.popleft()
        physical_gpu = worker.gpu.gpu
        if physical_gpu in candidate.excluded_gpus:
            blocked.append(candidate)
        elif candidate.preferred_gpus and physical_gpu not in candidate.preferred_gpus:
            blocked.append(candidate)
        else:
            pending.extend(blocked)
            return candidate
    pending.extend(blocked)
    return None


def _wait_for_worker_revision(
    worker: NativeGpuWorker,
    repo: str,
    revision: str,
    *,
    timeout_seconds: float = 1800.0,
) -> dict[str, object]:
    deadline = time.monotonic() + timeout_seconds
    while True:
        timeout = max(0.0, deadline - time.monotonic())
        if timeout == 0.0:
            raise TimeoutError(f"timed out waiting for native worker {worker.gpu.gpu} on {repo} {revision}")
        if worker.process is not None and worker.process.poll() is not None:
            raise RuntimeError(f"native gpu worker {worker.gpu.gpu} exited with code {worker.process.returncode}")
        try:
            event = worker.events.get(timeout=min(timeout, 1.0))
        except queue.Empty:
            continue
        event_type = event.get("event")
        if event_type in {"ready", "shutdown_ack"}:
            continue
        if event_type in {"worker_closed", "worker_error"}:
            raise RuntimeError(str(event))
        if str(event.get("repo", "")) == repo and str(event.get("revision", "")) == revision:
            return event


def _submit_native_pages(
    native_workers: Sequence[NativeGpuWorker],
    pending: "deque[NativePageTask]",
    in_flight: dict[NativeGpuWorker, NativePageTask],
    next_worker: int,
    telemetry: dict[str, GpuTelemetry],
) -> tuple[int, int, int]:
    submitted = 0
    submitted_existing = 0
    idle_workers = [worker for worker in native_workers if worker not in in_flight]
    for worker in idle_workers:
        if not pending:
            break
        task = _pop_next_eligible_task(pending, worker)
        if task is None:
            continue
        task.started_at = time.monotonic()
        profile = telemetry[worker.worker_key]
        if profile.first_started_at == 0.0:
            profile.first_started_at = task.started_at
        task.batch_size = _choose_autotune_batch(profile)
        worker.submit_page(task)
        in_flight[worker] = task
        submitted += 1
        if task.source == "existing":
            submitted_existing += 1
    return submitted, submitted_existing, next_worker


def _is_fatal_native_worker_config_error(event: dict[str, object]) -> bool:
    message = " ".join(str(event.get(key, "")) for key in ("message", "error", "traceback")).lower()
    return (
        "flashattention2" in message
        or "flashattention 2" in message
        or "flash_attn" in message
        or "attn_implementation" in message
        or "importerror" in message
        or "package is not installed" in message
        or "no package metadata was found" in message
    )


def _handle_native_worker_event(
    native_workers: Sequence[NativeGpuWorker],
    worker: NativeGpuWorker,
    event: dict[str, object],
    pending: "deque[NativePageTask]",
    in_flight: dict[NativeGpuWorker, NativePageTask],
    telemetry: dict[str, GpuTelemetry],
    scheduler_stats: dict[str, float],
    progress: tqdm,
    *,
    errors: int,
) -> int:
    event_type = event.get("event")
    if event_type in {"ready", "shutdown_ack"}:
        return errors
    if event_type in {"worker_closed", "worker_error"}:
        raise RuntimeError(str(event))
    if event_type not in {"progress", "error"}:
        return errors

    task = in_flight.get(worker)
    if task is None:
        raise RuntimeError(f"native gpu worker {worker.gpu.gpu} returned {event_type} with no in-flight page")
    repo = str(event.get("repo", ""))
    revision = str(event.get("revision", ""))
    page = int(event.get("page", -1))
    if repo != task.repo or revision != task.revision or page != task.page:
        raise RuntimeError(
            f"native gpu worker {worker.gpu.gpu} returned out-of-order event "
            f"{repo} {revision} page={page}; expected {task.repo} {task.revision} page={task.page}"
        )
    in_flight.pop(worker, None)
    if event_type == "error":
        if _is_fatal_native_worker_config_error(event):
            raise RuntimeError(f"fatal native worker configuration error: {event}")
        excluded = tuple(sorted(set(task.excluded_gpus + (worker.gpu.gpu,))))
        preferred_gpus = task.preferred_gpus
        has_eligible_worker = any(
            candidate.gpu.gpu not in excluded
            and (not preferred_gpus or candidate.gpu.gpu in preferred_gpus)
            for candidate in native_workers
        )
        if not has_eligible_worker and preferred_gpus:
            preferred_gpus = ()
            has_eligible_worker = any(candidate.gpu.gpu not in excluded for candidate in native_workers)
        if has_eligible_worker:
            pending.append(
                NativePageTask(
                    task.repo,
                    task.revision,
                    task.page,
                    task.page_count,
                    task.components,
                    task.vectors,
                    task.work_units,
                    source=task.source,
                    excluded_gpus=excluded,
                    preferred_gpus=preferred_gpus,
                    max_estimated_tokens=task.max_estimated_tokens,
                )
            )
            progress.set_postfix_str(
                f"{_short_repo_name(task.repo)} retrying {task.revision[:12]} page={task.page} without gpu {worker.gpu.gpu}"
            )
        else:
            errors += 1
            print(json.dumps(event, sort_keys=True), file=sys.stderr)
    else:
        profile = telemetry[worker.worker_key]
        elapsed = float(event.get("elapsed_seconds", 0.0) or 0.0)
        encode_elapsed = float(event.get("encode_seconds", 0.0) or 0.0)
        read_elapsed = float(event.get("read_seconds", 0.0) or 0.0)
        write_elapsed = float(event.get("write_seconds", 0.0) or 0.0)
        compose_elapsed = float(event.get("compose_seconds", 0.0) or 0.0)
        components = int(event.get("components", 0) or 0)
        components_written = int(event.get("components_written", 0) or 0)
        vectors = int(event.get("vectors", 0) or 0)
        composed = int(event.get("composed_written", 0) or 0)
        tokens = int(float(event.get("tokens", 0.0) or 0.0))
        padded_tokens = int(float(event.get("padded_tokens", 0.0) or 0.0))
        attention_tokens = int(float(event.get("attention_tokens", 0.0) or 0.0))
        micro_batches = int(float(event.get("micro_batches", 0.0) or 0.0))
        page_count = int(event.get("page_count", task.page_count) or task.page_count)
        profile.shards += 1
        profile.pages += page_count
        profile.components += components
        profile.components_written += components_written
        profile.vectors += vectors
        profile.composed_written += composed
        profile.elapsed_seconds += elapsed
        profile.encode_seconds += encode_elapsed
        profile.read_seconds += read_elapsed
        profile.write_seconds += write_elapsed
        profile.compose_seconds += compose_elapsed
        profile.tokens += tokens
        profile.padded_tokens += padded_tokens
        profile.attention_tokens += attention_tokens
        profile.micro_batches += micro_batches
        profile.current_batch_size = int(event.get("batch_size", profile.current_batch_size) or profile.current_batch_size)
        profile.current_max_seq_length = int(
            event.get("max_seq_length", profile.current_max_seq_length) or profile.current_max_seq_length
        )
        if elapsed > 0.0:
            components_per_second = components / elapsed
            vectors_per_second = vectors / elapsed
            tokens_per_second = tokens / encode_elapsed if encode_elapsed > 0.0 else 0.0
            padded_tokens_per_second = padded_tokens / encode_elapsed if encode_elapsed > 0.0 else 0.0
            alpha = 0.35
            batch_profile = profile.batch_profiles.setdefault(profile.current_batch_size, BatchTelemetry())
            batch_profile.shards += 1
            batch_profile.components += components
            batch_profile.vectors += vectors
            batch_profile.elapsed_seconds += elapsed
            batch_profile.encode_seconds += encode_elapsed
            batch_profile.tokens += tokens
            batch_profile.padded_tokens += padded_tokens
            batch_profile.attention_tokens += attention_tokens
            if batch_profile.ewma_components_per_second <= 0.0:
                batch_profile.ewma_components_per_second = components_per_second
                batch_profile.ewma_vectors_per_second = vectors_per_second
                batch_profile.ewma_tokens_per_second = tokens_per_second
                batch_profile.ewma_padded_tokens_per_second = padded_tokens_per_second
            else:
                batch_profile.ewma_components_per_second = (
                    (1.0 - alpha) * batch_profile.ewma_components_per_second + alpha * components_per_second
                )
                batch_profile.ewma_vectors_per_second = (
                    (1.0 - alpha) * batch_profile.ewma_vectors_per_second + alpha * vectors_per_second
                )
                if tokens_per_second > 0.0:
                    batch_profile.ewma_tokens_per_second = (
                        (1.0 - alpha) * batch_profile.ewma_tokens_per_second + alpha * tokens_per_second
                    )
                if padded_tokens_per_second > 0.0:
                    batch_profile.ewma_padded_tokens_per_second = (
                        (1.0 - alpha) * batch_profile.ewma_padded_tokens_per_second
                        + alpha * padded_tokens_per_second
                    )
            if profile.ewma_components_per_second <= 0.0:
                profile.ewma_components_per_second = components_per_second
                profile.ewma_vectors_per_second = vectors_per_second
                profile.ewma_tokens_per_second = tokens_per_second
                profile.ewma_padded_tokens_per_second = padded_tokens_per_second
            else:
                profile.ewma_components_per_second = (
                    (1.0 - alpha) * profile.ewma_components_per_second + alpha * components_per_second
                )
                profile.ewma_vectors_per_second = (
                    (1.0 - alpha) * profile.ewma_vectors_per_second + alpha * vectors_per_second
                )
                if tokens_per_second > 0.0:
                    profile.ewma_tokens_per_second = (
                        (1.0 - alpha) * profile.ewma_tokens_per_second + alpha * tokens_per_second
                    )
                if padded_tokens_per_second > 0.0:
                    profile.ewma_padded_tokens_per_second = (
                        (1.0 - alpha) * profile.ewma_padded_tokens_per_second + alpha * padded_tokens_per_second
                    )
        if profile.autotune_candidates and profile.locked_batch_size <= 0:
            if all(profile.batch_profiles.get(candidate, BatchTelemetry()).shards >= 2 for candidate in profile.autotune_candidates):
                profile.locked_batch_size = max(
                    profile.autotune_candidates,
                    key=lambda candidate: profile.batch_profiles[candidate].ewma_components_per_second,
                )
                profile.current_batch_size = profile.locked_batch_size
                print(
                    json.dumps(
                        {
                            "event": "gpu_batch_autotune_settled",
                            "gpu": worker.worker_key,
                            "physical_gpu": worker.gpu.gpu,
                            "selected_batch_size": profile.locked_batch_size,
                            "candidates": {
                                str(candidate): {
                                    "components_per_second": profile.batch_profiles[candidate].ewma_components_per_second,
                                    "shards": profile.batch_profiles[candidate].shards,
                                }
                                for candidate in profile.autotune_candidates
                            },
                        },
                        sort_keys=True,
                    ),
                    file=sys.stderr,
                    flush=True,
                )
        if profile.measured and profile.measured_at == 0.0:
            profile.measured_at = time.monotonic()
        if all(item.measured for item in telemetry.values()):
            if scheduler_stats["settled_at"] == 0.0:
                scheduler_stats["settled_at"] = time.monotonic()
                scheduler_stats["post_settle_started_at"] = scheduler_stats["settled_at"]
                _save_gpu_profile_cache(
                    scheduler_stats["profile_cache_path"],  # type: ignore[arg-type]
                    str(scheduler_stats["profile_cache_key"]),
                    telemetry,
                )
                print(
                    json.dumps(
                        {
                            "event": "gpu_autotune_settled",
                            "settle_seconds": scheduler_stats["settled_at"] - scheduler_stats["run_started_at"],
                            "gpus": {
                                gpu: {
                                    "batch_size": item.current_batch_size,
                                    "components_per_second": item.ewma_components_per_second,
                                    "padded_tokens_per_second": item.ewma_padded_tokens_per_second,
                                    "tokens_per_second": item.ewma_tokens_per_second,
                                    "vectors_per_second": item.ewma_vectors_per_second,
                                }
                                for gpu, item in sorted(telemetry.items())
                            },
                        },
                        sort_keys=True,
                    ),
                    file=sys.stderr,
                    flush=True,
                )
            else:
                scheduler_stats["post_settle_pages"] += float(page_count)
                scheduler_stats["post_settle_components"] += float(components)
                scheduler_stats["post_settle_vectors"] += float(vectors)
                scheduler_stats["post_settle_composed"] += float(composed)
        progress.update(page_count)
        measured = sum(1 for item in telemetry.values() if item.measured)
        total = len(telemetry)
        gpu_summary = " ".join(
            f"g{gpu}:{item.ewma_components_per_second:.1f}c/s@b{item.current_batch_size}"
            for gpu, item in sorted(telemetry.items())
        )
        progress.set_postfix_str(
            f"{event.get('message', _short_repo_name(task.repo))} measured={measured}/{total} {gpu_summary}"
        )
    return errors


def _drain_native_worker_events(
    native_workers: Sequence[NativeGpuWorker],
    pending: "deque[NativePageTask]",
    in_flight: dict[NativeGpuWorker, NativePageTask],
    telemetry: dict[str, GpuTelemetry],
    scheduler_stats: dict[str, float],
    progress: tqdm,
    *,
    errors: int,
) -> int:
    for worker in native_workers:
        while True:
            try:
                event = worker.events.get_nowait()
            except queue.Empty:
                break
            errors = _handle_native_worker_event(
                native_workers,
                worker,
                event,
                pending,
                in_flight,
                telemetry,
                scheduler_stats,
                progress,
                errors=errors,
            )
    return errors


def _wait_for_native_worker_event(
    native_workers: Sequence[NativeGpuWorker],
    pending: "deque[NativePageTask]",
    in_flight: dict[NativeGpuWorker, NativePageTask],
    telemetry: dict[str, GpuTelemetry],
    scheduler_stats: dict[str, float],
    progress: tqdm,
    *,
    errors: int,
) -> int:
    while True:
        for worker, task in list(in_flight.items()):
            if worker.process is not None and worker.process.poll() is not None:
                raise RuntimeError(f"native gpu worker {worker.gpu.gpu} exited with code {worker.process.returncode}")
        previous_in_flight = len(in_flight)
        errors = _drain_native_worker_events(
            native_workers,
            pending,
            in_flight,
            telemetry,
            scheduler_stats,
            progress,
            errors=errors,
        )
        if len(in_flight) < previous_in_flight:
            return errors
        time.sleep(0.05)


def _build_embedding_manifest(args: argparse.Namespace) -> str:
    _, passage_manifest = make_embedder_with_manifest(
        args.model,
        device="cpu",
        batch_size=args.batch_size,
        max_seq_length=args.max_seq_length,
        role="passage",
    )
    ensure_cache_compatible(
        args.vector_cache_dir,
        passage_manifest,
        force_new_cache=args.force_new_cache,
        recreate_manifest=args.recreate_manifest,
    )
    manifest_digest = passage_manifest.digest()
    native_runtime_contract = {
        "native_attn_implementation": args.native_attn_implementation,
        "native_encode_dtype": args.native_encode_dtype,
    }
    if native_runtime_contract == {
        "native_attn_implementation": "default",
        "native_encode_dtype": "float32",
    }:
        return manifest_digest
    payload = json.dumps(
        {
            "base_manifest_digest": manifest_digest,
            **native_runtime_contract,
        },
        sort_keys=True,
    )
    return hashlib.sha256(payload.encode("utf-8")).hexdigest()[:32]


def _native_helper_path() -> Path:
    override = os.environ.get("LOCALIZE_NATIVE_HELPER")
    if override:
        return Path(override)
    return Path(__file__).resolve().parent / "tools" / "native_localizer_helper" / "native_localizer_helper.py"


def _native_page_size(workers: Sequence[NativeGpuWorkerSpec], default_batch_size: int) -> int:
    largest_batch = max((worker.batch_size or default_batch_size) for worker in workers)
    return max(256, min(2048, largest_batch * 4))


def _gpu_profile_cache_path(args: argparse.Namespace) -> Path:
    return args.vector_cache_dir / "gpu-autotune-profiles.json"


def _gpu_profile_cache_key(args: argparse.Namespace, manifest_digest: str) -> str:
    workers = args.gpu_worker or [NativeGpuWorkerSpec("0", args.batch_size)]
    return json.dumps(
        {
            "autotune_version": 3,
            "manifest_digest": manifest_digest,
            "max_seq_length": args.max_seq_length,
            "model": args.model,
            "native_attn_implementation": args.native_attn_implementation,
            "native_encode_dtype": args.native_encode_dtype,
            "native_fixed_batches": args.native_fixed_batches,
            "native_long_page_gpu": args.native_long_page_gpu,
            "native_long_token_threshold": args.native_long_token_threshold,
            "native_planner_max_components": args.native_planner_max_components,
            "native_planner_target_tokens": args.native_planner_target_tokens,
            "native_shard_pages": args.native_shard_pages,
            "native_target_attention_tokens": args.native_target_attention_tokens,
            "native_target_padded_tokens": args.native_target_padded_tokens,
            "workers": [
                {
                    "key": runtime.key,
                    "gpu": runtime.spec.gpu,
                    "batch_size": runtime.spec.batch_size or args.batch_size,
                    "candidates": (
                        []
                        if args.native_fixed_batches
                        else _batch_autotune_candidates(runtime.spec.batch_size or args.batch_size)
                    ),
                }
                for runtime in _native_worker_runtimes(workers)
            ],
        },
        sort_keys=True,
    )


def _load_gpu_profile_cache(path: Path, key: str) -> dict[str, dict[str, float]]:
    try:
        data = json.loads(path.read_text())
    except (FileNotFoundError, json.JSONDecodeError):
        return {}
    profiles = data.get(key, {})
    if not isinstance(profiles, dict):
        return {}
    loaded: dict[str, dict[str, float]] = {}
    for gpu, profile in profiles.items():
        if not isinstance(profile, dict):
            continue
        try:
            loaded[str(gpu)] = {
                "batch_size": float(profile["batch_size"]),
                "components_per_second": float(profile["components_per_second"]),
                "padded_tokens_per_second": float(profile.get("padded_tokens_per_second", 0.0)),
                "tokens_per_second": float(profile.get("tokens_per_second", 0.0)),
                "vectors_per_second": float(profile.get("vectors_per_second", 0.0)),
            }
        except (KeyError, TypeError, ValueError):
            continue
    return loaded


def _save_gpu_profile_cache(path: Path, key: str, telemetry: dict[str, GpuTelemetry]) -> None:
    try:
        data = json.loads(path.read_text())
    except (FileNotFoundError, json.JSONDecodeError):
        data = {}
    data[key] = {
        gpu: {
            "batch_size": profile.current_batch_size,
            "components_per_second": profile.ewma_components_per_second,
            "padded_tokens_per_second": profile.ewma_padded_tokens_per_second,
            "tokens_per_second": profile.ewma_tokens_per_second,
            "vectors_per_second": profile.ewma_vectors_per_second,
        }
        for gpu, profile in sorted(telemetry.items())
        if profile.measured and profile.ewma_components_per_second > 0.0
    }
    path.parent.mkdir(parents=True, exist_ok=True)
    temp_path = path.with_name(f"{path.name}.{os.getpid()}.{uuid.uuid4().hex}.tmp")
    temp_path.write_text(json.dumps(data, indent=2, sort_keys=True) + "\n")
    temp_path.replace(path)


def _estimated_planner_tokens(text: str, max_seq_length: int) -> int:
    return min(max_seq_length, max(1, (len(text) + 2) // 3 + 16))


def _planner_bucket_limits(
    page_size: int,
    *,
    target_tokens: int = 65_536,
    max_items: int | None = None,
) -> tuple[int, int]:
    if max_items is None:
        max_items = min(256, max(1, page_size))
    return max(32, int(max_items)), max(1, int(target_tokens))


def _bucket_shard_tasks(
    *,
    repo: str,
    revision: str,
    source: str,
    next_page: int,
    components: dict[str, dict[str, object]],
    vectors: list[dict[str, str]],
    page_size: int,
    planner_target_tokens: int = 65_536,
    planner_max_components: int | None = None,
    long_token_threshold: int = 0,
    long_page_gpus: Sequence[str] = (),
) -> tuple[list[NativePageTask], int]:
    if not components:
        if not vectors:
            return [], next_page
        return [
            NativePageTask(
                repo=repo,
                revision=revision,
                page=next_page,
                page_count=1,
                components=[],
                vectors=vectors,
                work_units=max(1, len(vectors) * 32),
                source=source,
            )
        ], next_page + 1

    max_items, target_tokens = _planner_bucket_limits(
        page_size,
        target_tokens=planner_target_tokens,
        max_items=planner_max_components,
    )
    ordered_components = sorted(
        components.values(),
        key=lambda item: (int(item["estimated_tokens"]), str(item["key"])),
        reverse=True,
    )
    buckets: list[list[dict[str, object]]] = []
    current: list[dict[str, object]] = []
    current_tokens = 0
    for component in ordered_components:
        tokens = int(component["estimated_tokens"])
        if current and (len(current) >= max_items or current_tokens + tokens > target_tokens):
            buckets.append(current)
            current = []
            current_tokens = 0
        current.append(component)
        current_tokens += tokens
    if current:
        buckets.append(current)
    if not buckets:
        buckets.append([])

    key_to_bucket: dict[str, int] = {}
    for bucket_index, bucket in enumerate(buckets):
        for component in bucket:
            key_to_bucket[str(component["key"])] = bucket_index

    bucket_vectors: list[list[dict[str, str]]] = [[] for _ in buckets]
    for vector in vectors:
        referenced = [
            key
            for key in (vector.get("component_key", ""), vector.get("parent_component_key", ""))
            if key in components
        ]
        if referenced:
            bucket_index = max(
                (key_to_bucket[key] for key in referenced),
                key=lambda index: sum(
                    int(components[key]["estimated_tokens"])
                    for key in referenced
                    if key_to_bucket.get(key) == index
                ),
            )
        else:
            bucket_index = 0
        bucket_vectors[bucket_index].append(vector)

    tasks: list[NativePageTask] = []
    for bucket_index, bucket in enumerate(buckets):
        component_keys = {str(component["key"]) for component in bucket}
        for vector in bucket_vectors[bucket_index]:
            for key in (vector.get("component_key", ""), vector.get("parent_component_key", "")):
                if key in components:
                    component_keys.add(key)
        task_components = sorted(
            (components[key] for key in component_keys),
            key=lambda item: (int(item["estimated_tokens"]), str(item["key"])),
            reverse=True,
        )
        task_vectors = bucket_vectors[bucket_index]
        if not task_components and not task_vectors:
            continue
        work_tokens = sum(int(item["estimated_tokens"]) for item in task_components)
        work_units = max(1, work_tokens + len(task_vectors) * 32)
        max_estimated_tokens = max((int(item["estimated_tokens"]) for item in task_components), default=0)
        preferred_gpus = (
            tuple(str(gpu) for gpu in long_page_gpus)
            if long_token_threshold > 0 and max_estimated_tokens >= long_token_threshold
            else ()
        )
        tasks.append(
            NativePageTask(
                repo=repo,
                revision=revision,
                page=next_page,
                page_count=1,
                components=task_components,
                vectors=task_vectors,
                work_units=work_units,
                source=source,
                preferred_gpus=preferred_gpus,
                max_estimated_tokens=max_estimated_tokens,
            )
        )
        next_page += 1
    return tasks, next_page


def _repo_page_tasks(
    repo: str,
    revisions: Sequence[str],
    *,
    args: argparse.Namespace,
    manifest_digest: str,
    page_size: int,
    source: str = "build",
) -> Iterable[NativePageTask]:
    revision_label = revisions[0] if len(revisions) == 1 else "reconcile"
    shard_pages = max(1, int(getattr(args, "native_shard_pages", 1) or 1))
    next_task_page = 0
    shard_count = 0
    shard_components: dict[str, dict[str, object]] = {}
    shard_vectors: list[dict[str, str]] = []

    def flush_shard() -> list[NativePageTask]:
        nonlocal next_task_page, shard_count, shard_components, shard_vectors
        if shard_count == 0:
            return []
        tasks, next_task_page = _bucket_shard_tasks(
            repo=repo,
            revision=revision_label,
            source=source,
            next_page=next_task_page,
            components=shard_components,
            vectors=shard_vectors,
            page_size=page_size,
            planner_target_tokens=args.native_planner_target_tokens,
            planner_max_components=args.native_planner_max_components,
            long_token_threshold=args.native_long_token_threshold,
            long_page_gpus=args.native_long_page_gpu,
        )
        shard_count = 0
        shard_components = {}
        shard_vectors = []
        return tasks

    for page, work in enumerate(
        planned_vector_pipeline_pages(
            repo,
            revisions,
            db_dir=args.db_dir,
            vector_cache_dir=args.vector_cache_dir,
            manifest_digest=manifest_digest,
            page_size=page_size,
            dedupe_components_across_pages=False,
        )
    ):
        for item in work.components:
            if item.key not in shard_components:
                shard_components[item.key] = {
                    "key": item.key,
                    "text": item.text,
                    "estimated_tokens": _estimated_planner_tokens(item.text, args.max_seq_length),
                }
        shard_vectors.extend(
            {
                "vector_key": item.vector_key,
                "component_key": item.component_key,
                "parent_component_key": item.parent_component_key,
            }
            for item in work.vectors
        )
        shard_count += 1
        if shard_count < shard_pages:
            continue
        for task in flush_shard():
            yield task
    for task in flush_shard():
        yield task


def _collect_repo_page_tasks(
    repo: str,
    revisions: Sequence[str],
    *,
    args: argparse.Namespace,
    manifest_digest: str,
    page_size: int,
    source: str,
) -> list[NativePageTask]:
    return list(
        _repo_page_tasks(
            repo,
            revisions,
            args=args,
            manifest_digest=manifest_digest,
            page_size=page_size,
            source=source,
        )
    )


def _existing_vector_revisions(
    repo: str,
    *,
    grouped: dict[str, list],
    args: argparse.Namespace,
) -> tuple[str, list[str]]:
    return (
        repo,
        active_revisions_for_repo_records(
            repo,
            repo_records_for_args(repo, grouped, args),
            db_dir=args.db_dir,
        ),
    )


def _build_vector_index_worker(
    task: tuple[str, list[str]],
    *,
    db_dir: Path,
    vector_cache_dir: Path,
    manifest_digest: str,
    load_workers: int,
) -> dict[str, object]:
    repo, revisions = task
    db_path = db_dir / repo / "embeddings.db"
    built = 0
    skipped_empty = 0
    try:
        conn = sqlite3.connect(db_path, timeout=30)
        try:
            for revision in revisions:
                entries = revision_chunk_index(conn, revision)
                if not entries:
                    skipped_empty += 1
                    continue
                load_revision_vector_matrix(
                    vector_cache_dir,
                    repo,
                    revision,
                    entries,
                    manifest_digest=manifest_digest,
                    workers=load_workers,
                )
                built += 1
        finally:
            conn.close()
        return {"repo": repo, "built": built, "skipped_empty": skipped_empty, "error": None}
    except Exception:
        return {
            "repo": repo,
            "built": built,
            "skipped_empty": skipped_empty,
            "error": traceback.format_exc(),
        }


def _repo_revisions_for_index(
    repo: str,
    repo_records: Sequence,
    *,
    db_dir: Path,
) -> tuple[str, list[str]]:
    db_path = db_dir / repo / "embeddings.db"
    if not db_path.exists():
        return repo, []
    conn = sqlite3.connect(db_path, timeout=30)
    try:
        return repo, active_revisions_for_records(conn, repo_records)
    finally:
        conn.close()


def _augment_chunkless_repo_entry(
    repo: str,
    *,
    args: argparse.Namespace,
) -> dict[str, object]:
    try:
        return augment_chunkless_repo_worker(
            repo,
            db_dir=args.db_dir,
            clones_dir=args.clones_dir,
            seed_dir=args.seed_dir,
            seed_only=args.seed_only,
            revisions=None,
            summary_workers=args.summary_workers,
            vector_cache_dirs=args.vector_cache_dir,
            model=args.model,
            batch_size=args.batch_size,
            max_seq_length=args.max_seq_length,
            page_size=args.page_size,
            dry_run=args.dry_run,
        )
    except Exception:
        return {"repo": repo, "error": traceback.format_exc()}


def run_augment_chunkless(args: argparse.Namespace) -> int:
    if args.vector_cache_dir:
        helper = reconcile_helper_path()
        if not helper.is_file():
            raise FileNotFoundError(
                f"missing reconcile helper: {helper} (set LOCALIZE_RECONCILE_HELPER to override)"
            )
    repos = sorted(
        path.name
        for path in args.db_dir.iterdir()
        if path.is_dir() and (path / "embeddings.db").is_file()
    )
    requested = requested_repo_filter(args)
    if requested is not None:
        repos = [repo for repo in repos if repo in requested]
    if args.limit_repos is not None:
        repos = repos[: args.limit_repos]
    if not repos:
        print(json.dumps({"repos": 0, "chunkless_candidates": 0, "chunkless_inserted": 0}, sort_keys=True))
        return 0

    progress = tqdm(total=len(repos), desc="augment-chunkless", unit="repo", dynamic_ncols=True)
    results: list[dict[str, object]] = []
    errors = 0
    with concurrent.futures.ProcessPoolExecutor(max_workers=args.repo_workers) as pool:
        futures = {
            pool.submit(_augment_chunkless_repo_entry, repo, args=args): repo
            for repo in repos
        }
        for future in concurrent.futures.as_completed(futures):
            repo = futures[future]
            progress.set_postfix_str(repo)
            result = future.result()
            results.append(result)
            if result.get("error"):
                errors += 1
                print(json.dumps(result, sort_keys=True), file=sys.stderr)
            progress.update(1)
    progress.close()
    repo_revisions = [
        (str(result.get("repo", "")), list(result.get("selected_revisions") or []))
        for result in results
        if result.get("repo") and not result.get("error")
    ]
    vector_caches = []
    if not args.dry_run:
        vector_caches = _reconcile_augmented_vectors(
            repo_revisions,
            db_dir=args.db_dir,
            vector_cache_dirs=args.vector_cache_dir,
            model=args.model,
            batch_size=args.batch_size,
            max_seq_length=args.max_seq_length,
            page_size=args.page_size,
        )
    for result in results:
        result.pop("selected_revisions", None)
    summary = {
        "repos": len(repos),
        "errors": errors,
        "dry_run": args.dry_run,
        "chunkless_candidates": sum(int(result.get("chunkless_candidates") or 0) for result in results),
        "chunkless_inserted": sum(int(result.get("chunkless_inserted") or 0) for result in results),
        "chunkless_skipped": sum(int(result.get("chunkless_skipped") or 0) for result in results),
        "seeded_rows_used": sum(int(result.get("seeded_rows_used") or 0) for result in results),
        "memo_hits": sum(int(result.get("memo_hits") or 0) for result in results),
        "bifrost_calls": sum(int(result.get("bifrost_calls") or 0) for result in results),
        "vector_caches": vector_caches,
        "repos_detail": sorted(results, key=lambda row: str(row.get("repo", ""))),
    }
    print(json.dumps(summary, sort_keys=True))
    return 1 if errors else 0


def run_build_vector_index(args: argparse.Namespace) -> int:
    manifest = read_cache_manifest(args.vector_cache_dir)
    if manifest is None:
        raise RuntimeError(f"no vector cache manifest at {args.vector_cache_dir}")
    records = [record for record in read_selection(args.selection) if record.split == args.split]
    grouped = group_tasks_by_repo(records)
    repos = sorted(grouped)
    requested = requested_repo_filter(args)
    if requested is not None:
        repos = [repo for repo in repos if repo in requested]
    if args.limit_repos is not None:
        repos = repos[: args.limit_repos]
    with concurrent.futures.ThreadPoolExecutor(max_workers=min(32, max(1, args.workers))) as pool:
        revision_rows = list(
            pool.map(
                lambda repo: _repo_revisions_for_index(repo, grouped[repo], db_dir=args.db_dir),
                repos,
            )
        )
    tasks = [(repo, revisions) for repo, revisions in revision_rows if revisions]
    progress = tqdm(total=len(tasks), desc="build-vector-index", unit="repo", dynamic_ncols=True)
    errors = 0
    total_built = 0
    with concurrent.futures.ProcessPoolExecutor(max_workers=args.workers) as pool:
        futures = {
            pool.submit(
                _build_vector_index_worker,
                task,
                db_dir=args.db_dir,
                vector_cache_dir=args.vector_cache_dir,
                manifest_digest=manifest.digest(),
                load_workers=args.load_workers,
            ): task[0]
            for task in tasks
        }
        for future in concurrent.futures.as_completed(futures):
            progress.update(1)
            result = future.result()
            total_built += int(result.get("built") or 0)
            if result.get("error"):
                errors += 1
                print(json.dumps(result, sort_keys=True), file=sys.stderr)
    progress.close()
    print(
        json.dumps(
            {
                "repos": len(tasks),
                "revisions_indexed": total_built,
                "errors": errors,
                "vector_cache_dir": str(args.vector_cache_dir),
                "manifest_digest": manifest.digest(),
            },
            sort_keys=True,
        )
    )
    return 1 if errors else 0


def run_build_embeddings(args: argparse.Namespace) -> int:
    native_helper = _native_helper_path()
    if not native_helper.is_file():
        raise FileNotFoundError(
            f"missing native helper: {native_helper} "
            "(set LOCALIZE_NATIVE_HELPER to override)"
        )
    reconcile_helper = reconcile_helper_path()
    if not reconcile_helper.is_file():
        raise FileNotFoundError(
            f"missing reconcile helper: {reconcile_helper} "
            "(set LOCALIZE_RECONCILE_HELPER to override)"
        )
    manifest_digest = _build_embedding_manifest(args)

    records = read_selection(args.selection)
    selection_hash = selection_content_hash(records)
    grouped = group_tasks_by_repo(records)
    all_repos = selected_repos_from_grouped(grouped, args)

    planning = tqdm(total=len(all_repos), desc="build clones", unit="repo", dynamic_ncols=True)
    if getattr(args, "skip_clone", False):
        available_repos, errors = list(all_repos), 0
        planning.update(len(all_repos))
    else:
        available_repos, errors = _clone_repos_for_build(
            all_repos,
            clones_dir=args.clones_dir,
            progress=planning,
            clone_workers=args.clone_workers,
            debug=args.debug,
        )
    planning.close()

    repo_records, revision_counts = _plan_build_db_repos(
        available_repos,
        grouped=grouped,
        args=args,
    )
    repos = [repo for repo in available_repos if repo_records.get(repo)]
    if not repos:
        print(json.dumps({"repos": 0, "revisions": 0}, sort_keys=True))
        return 1 if errors else 0

    # Memory packing: order so big repos are spread among the cheap small ones (the
    # ProcessPool is FIFO, so order == schedule) and gate concurrent bigs so peak RSS
    # stays bounded while small repos keep repo_workers busy feeding the GPUs.
    big_bytes = max(0, int(getattr(args, "big_repo_mb", 150))) * 1024 * 1024
    big_repos: set[str] = set()
    if big_bytes > 0:
        repos, big_repos = _pack_repos_by_size(
            repos, clones_dir=args.clones_dir, big_bytes=big_bytes
        )
        print(
            json.dumps(
                {
                    "event": "repo_packing",
                    "repos": len(repos),
                    "big_repos": sorted(big_repos),
                    "big_repo_mb": args.big_repo_mb,
                    "max_big_concurrent": args.max_big_concurrent,
                },
                sort_keys=True,
            ),
            file=sys.stderr,
            flush=True,
        )

    workers = args.gpu_worker or [NativeGpuWorkerSpec("0", args.batch_size)]
    runtime_workers = _native_worker_runtimes(workers)
    page_size = _native_page_size(workers, args.batch_size)
    reconcile_progress = tqdm(total=len(repos), desc="db reconcile", unit="repo", dynamic_ncols=True)
    build_db_progress = tqdm(
        total=sum(len(repo_records[repo]) for repo in repos),
        desc="build db",
        unit="task",
        dynamic_ncols=True,
    )
    page_plan_progress = tqdm(
        total=sum(revision_counts.get(repo, 0) for repo in repos),
        desc="plan pages",
        unit="rev",
        dynamic_ncols=True,
    )
    progress = tqdm(total=0, desc="build-embeddings", unit="page", dynamic_ncols=True)
    ctx = multiprocessing.get_context("spawn")
    manager = multiprocessing.Manager()
    progress_queue = manager.Queue()
    big_repo_gate = manager.Semaphore(max(1, int(getattr(args, "max_big_concurrent", 1))))
    native_workers: list[NativeGpuWorker] = []
    futures: dict[concurrent.futures.Future[BuildDbStats], str] = {}
    submitted_pages = 0
    submitted_existing_pages = 0
    completed_repos: set[str] = set()
    build_db_seen_tasks: dict[str, int] = {}
    next_worker = 0
    pending_gpu: deque[NativePageTask] = deque()
    in_flight_gpu: dict[NativeGpuWorker, NativePageTask] = {}
    profile_cache_path = _gpu_profile_cache_path(args)
    profile_cache_key = _gpu_profile_cache_key(args, manifest_digest)
    cached_profiles = _load_gpu_profile_cache(profile_cache_path, profile_cache_key)
    telemetry: dict[str, GpuTelemetry] = {
        runtime.key: GpuTelemetry(
            runtime.key,
            runtime.spec.batch_size or args.batch_size,
            autotune_candidates=(
                ()
                if args.native_fixed_batches
                else _batch_autotune_candidates(runtime.spec.batch_size or args.batch_size)
            ),
        )
        for runtime in runtime_workers
    }
    for gpu, cached in cached_profiles.items():
        if gpu not in telemetry:
            continue
        profile = telemetry[gpu]
        profile.pages = 3
        profile.shards = 3
        profile.locked_batch_size = int(cached["batch_size"])
        profile.current_batch_size = int(cached["batch_size"])
        profile.ewma_components_per_second = float(cached["components_per_second"])
        profile.ewma_padded_tokens_per_second = float(cached.get("padded_tokens_per_second", 0.0))
        profile.ewma_tokens_per_second = float(cached["tokens_per_second"])
        profile.ewma_vectors_per_second = float(cached["vectors_per_second"])
        profile.measured_at = time.monotonic()
    cache_hit = bool(telemetry) and all(item.measured for item in telemetry.values())
    if not cache_hit:
        page_size = min(page_size, 256)
    scheduler_stats: dict[str, object] = {
        "run_started_at": time.monotonic(),
        "settled_at": 0.0,
        "post_settle_started_at": 0.0,
        "post_settle_pages": 0.0,
        "post_settle_components": 0.0,
        "post_settle_vectors": 0.0,
        "post_settle_composed": 0.0,
        "profile_cache_key": profile_cache_key,
        "profile_cache_path": profile_cache_path,
    }
    if cache_hit:
        scheduler_stats["settled_at"] = scheduler_stats["run_started_at"]
        scheduler_stats["post_settle_started_at"] = scheduler_stats["run_started_at"]
        print(
            json.dumps(
                {
                    "event": "gpu_autotune_cache_hit",
                    "gpus": sorted(cached_profiles),
                    "profile_cache_path": str(profile_cache_path),
                },
                sort_keys=True,
            ),
            file=sys.stderr,
            flush=True,
        )
    try:
        with tempfile.TemporaryDirectory() as temp:
            temp_dir = Path(temp)
            for runtime in runtime_workers:
                worker = NativeGpuWorker(
                    helper=native_helper,
                    worker_key=runtime.key,
                    gpu=runtime.spec,
                    vector_cache_dir=args.vector_cache_dir,
                    manifest_digest=manifest_digest,
                    batch_size=args.batch_size,
                    max_seq_length=args.max_seq_length,
                    model=args.model,
                    target_padded_tokens=args.native_target_padded_tokens,
                    target_attention_tokens=args.native_target_attention_tokens,
                    encode_dtype=args.native_encode_dtype,
                    attn_implementation=args.native_attn_implementation,
                )
                worker.start(temp_dir)
                native_workers.append(worker)
            gpu_queue_limit = max(1, len(native_workers) * 4)
            scanner_workers = max(1, min(len(repos), args.repo_workers))
            with (
                concurrent.futures.ThreadPoolExecutor(max_workers=scanner_workers) as scanner,
                concurrent.futures.ThreadPoolExecutor(max_workers=scanner_workers) as planner,
                concurrent.futures.ProcessPoolExecutor(
                    max_workers=args.repo_workers,
                    mp_context=ctx,
                ) as executor,
            ):
                futures_existing = {
                    scanner.submit(_existing_vector_revisions, repo, grouped=grouped, args=args): repo
                    for repo in repos
                }
                page_futures: dict[concurrent.futures.Future[list[NativePageTask]], tuple[str, int]] = {}
                for repo in repos:
                    futures[
                        executor.submit(
                            build_db_worker_entry,
                            repo_records[repo],
                            clones_dir=args.clones_dir,
                            db_dir=args.db_dir,
                            selection_hash=selection_hash,
                            bifrost_library=args.bifrost_library,
                            resume=not args.no_resume,
                            extract_workers=args.extract_workers,
                            progress_queue=progress_queue,
                            is_big=repo in big_repos,
                            big_repo_gate=big_repo_gate,
                        )
                    ] = repo

                while futures_existing or futures or page_futures or pending_gpu or in_flight_gpu:
                    made_progress = False
                    if futures_existing:
                        done_existing, _pending_existing = concurrent.futures.wait(
                            futures_existing,
                            timeout=0.0,
                            return_when=concurrent.futures.FIRST_COMPLETED,
                        )
                        for future in done_existing:
                            futures_existing.pop(future)
                            repo, revisions = future.result()
                            made_progress = True
                            reconcile_progress.update(1)
                            if revisions:
                                page_futures[
                                    planner.submit(
                                        _collect_repo_page_tasks,
                                        repo,
                                        revisions,
                                        args=args,
                                        manifest_digest=manifest_digest,
                                        page_size=page_size,
                                        source="existing",
                                    )
                                ] = (f"{_short_repo_name(repo)} existing", len(revisions))
                                reconcile_progress.set_postfix_str(f"{_short_repo_name(repo)} planned existing")
                            else:
                                reconcile_progress.set_postfix_str(f"{_short_repo_name(repo)} no existing")

                    while True:
                        try:
                            event = progress_queue.get_nowait()
                        except Exception:
                            break
                        made_progress = True
                        if isinstance(event, dict) and event.get("event") == "revision":
                            repo = str(event["repo"])
                            revision = str(event["revision"])
                            page_futures[
                                planner.submit(
                                    _collect_repo_page_tasks,
                                    repo,
                                    [revision],
                                    args=args,
                                    manifest_digest=manifest_digest,
                                    page_size=page_size,
                                    source="build",
                                )
                            ] = (f"{_short_repo_name(repo)} {revision[:12]}", 1)
                            progress.set_postfix_str(f"{_short_repo_name(repo)} queued {revision[:12]}")
                        elif isinstance(event, dict) and event.get("event") == "log":
                            repo = str(event.get("repo", ""))
                            message = str(event.get("message", ""))
                            task_progress = _extract_task_progress(message)
                            if task_progress is not None and repo:
                                current, total_tasks = task_progress
                                previous = build_db_seen_tasks.get(repo, 0)
                                if current > previous:
                                    build_db_progress.update(current - previous)
                                    build_db_seen_tasks[repo] = current
                                build_db_progress.set_postfix_str(
                                    f"{_short_repo_name(repo)} {current}/{total_tasks}"
                                )

                    if page_futures:
                        done_pages, _pending_pages = concurrent.futures.wait(
                            page_futures,
                            timeout=0.0,
                            return_when=concurrent.futures.FIRST_COMPLETED,
                        )
                    else:
                        done_pages = set()
                    for future in done_pages:
                        label, planned_revisions = page_futures.pop(future)
                        tasks = future.result()
                        pending_gpu.extend(tasks)
                        made_progress = True
                        page_plan_progress.update(planned_revisions)
                        if tasks:
                            page_plan_progress.set_postfix_str(f"{label} planned {len(tasks)} shards")
                            progress.set_postfix_str(f"{label} planned {len(tasks)} shards")
                    submitted, submitted_existing, next_worker = _submit_native_pages(
                        native_workers,
                        pending_gpu,
                        in_flight_gpu,
                        next_worker,
                        telemetry,
                    )
                    submitted_pages += submitted
                    submitted_existing_pages += submitted_existing
                    if submitted:
                        made_progress = True

                    errors = _drain_native_worker_events(
                        native_workers,
                        pending_gpu,
                        in_flight_gpu,
                        telemetry,
                        scheduler_stats,
                        progress,
                        errors=errors,
                    )
                    if futures:
                        done, _pending = concurrent.futures.wait(
                            futures,
                            timeout=0.0,
                            return_when=concurrent.futures.FIRST_COMPLETED,
                        )
                    else:
                        done = set()
                    for future in done:
                        repo = futures.pop(future)
                        stats = future.result()
                        completed_repos.add(repo)
                        made_progress = True
                        repo_total_tasks = len(repo_records.get(repo, ()))
                        previous = build_db_seen_tasks.get(repo, 0)
                        if repo_total_tasks > previous:
                            build_db_progress.update(repo_total_tasks - previous)
                            build_db_seen_tasks[repo] = repo_total_tasks
                        if stats.error is not None:
                            errors += 1
                            if args.debug:
                                print(json.dumps({"repo": repo, "error": stats.error}, sort_keys=True), file=sys.stderr)
                            else:
                                print(f"error: {_short_repo_name(repo)} build failed: {stats.error}", file=sys.stderr)

                    if len(pending_gpu) + len(in_flight_gpu) >= gpu_queue_limit and in_flight_gpu:
                        errors = _wait_for_native_worker_event(
                            native_workers,
                            pending_gpu,
                            in_flight_gpu,
                            telemetry,
                            scheduler_stats,
                            progress,
                            errors=errors,
                        )
                        made_progress = True
                    elif not made_progress:
                        if in_flight_gpu:
                            errors = _wait_for_native_worker_event(
                                native_workers,
                                pending_gpu,
                                in_flight_gpu,
                                telemetry,
                                scheduler_stats,
                                progress,
                                errors=errors,
                            )
                        else:
                            time.sleep(0.1)
                while page_futures or pending_gpu or in_flight_gpu:
                    if page_futures:
                        done_pages, _pending_pages = concurrent.futures.wait(
                            page_futures,
                            timeout=0.1,
                            return_when=concurrent.futures.FIRST_COMPLETED,
                        )
                    else:
                        done_pages = set()
                    for future in done_pages:
                        _label, planned_revisions = page_futures.pop(future)
                        page_plan_progress.update(planned_revisions)
                        pending_gpu.extend(future.result())
                    submitted, submitted_existing, next_worker = _submit_native_pages(
                        native_workers,
                        pending_gpu,
                        in_flight_gpu,
                        next_worker,
                        telemetry,
                    )
                    submitted_pages += submitted
                    submitted_existing_pages += submitted_existing
                    if in_flight_gpu:
                        errors = _wait_for_native_worker_event(
                            native_workers,
                            pending_gpu,
                            in_flight_gpu,
                            telemetry,
                            scheduler_stats,
                            progress,
                            errors=errors,
                        )
    finally:
        for worker in native_workers:
            code = worker.shutdown()
            if code != 0:
                errors += 1
                print(f"error: native gpu worker {worker.gpu.gpu} exited with code {code}", file=sys.stderr)
        reconcile_progress.close()
        build_db_progress.close()
        page_plan_progress.close()
        progress.close()
        manager.shutdown()
    print(
        json.dumps(
            {
                "gpu_autotune_settle_seconds": (
                    scheduler_stats["settled_at"] - scheduler_stats["run_started_at"]
                    if scheduler_stats["settled_at"]
                    else None
                ),
                "gpu_profiles": {
                    gpu: {
                        "batch_size": profile.current_batch_size,
                        "batch_autotune_candidates": profile.autotune_candidates,
                        "batch_autotune_profiles": {
                            str(batch_size): {
                                "components_per_second": batch_profile.ewma_components_per_second,
                                "elapsed_seconds": batch_profile.elapsed_seconds,
                                "padded_tokens_per_second": batch_profile.ewma_padded_tokens_per_second,
                                "shards": batch_profile.shards,
                            }
                            for batch_size, batch_profile in sorted(profile.batch_profiles.items())
                        },
                        "components": profile.components,
                        "components_written": profile.components_written,
                        "components_per_second": profile.ewma_components_per_second,
                        "component_duplicate_ratio": profile.components / max(1, profile.components_written),
                        "composed_written": profile.composed_written,
                        "actual_composed_vectors_per_second": profile.composed_written
                        / max(0.001, profile.elapsed_seconds),
                        "encode_seconds": profile.encode_seconds,
                        "encode_fraction": profile.encode_seconds / max(0.001, profile.elapsed_seconds),
                        "elapsed_seconds": profile.elapsed_seconds,
                        "io_fraction": (profile.read_seconds + profile.write_seconds)
                        / max(0.001, profile.elapsed_seconds),
                        "measured": profile.measured,
                        "micro_batches": profile.micro_batches,
                        "pages": profile.pages,
                        "padded_tokens": profile.padded_tokens,
                        "padding_waste_ratio": profile.padded_tokens / max(1, profile.tokens),
                        "attention_tokens": profile.attention_tokens,
                        "read_seconds": profile.read_seconds,
                        "write_seconds": profile.write_seconds,
                        "compose_seconds": profile.compose_seconds,
                        "shards": profile.shards,
                        "tokens": profile.tokens,
                        "padded_tokens_per_second": profile.ewma_padded_tokens_per_second,
                        "tokens_per_second": profile.ewma_tokens_per_second,
                        "vectors": profile.vectors,
                        "vectors_per_second": profile.ewma_vectors_per_second,
                    }
                    for gpu, profile in sorted(telemetry.items())
                },
                "post_settle_components_per_second": (
                    scheduler_stats["post_settle_components"]
                    / max(0.001, time.monotonic() - scheduler_stats["post_settle_started_at"])
                    if scheduler_stats["post_settle_started_at"]
                    else None
                ),
                "post_settle_pages": int(scheduler_stats["post_settle_pages"]),
                "post_settle_vectors_per_second": (
                    scheduler_stats["post_settle_vectors"]
                    / max(0.001, time.monotonic() - scheduler_stats["post_settle_started_at"])
                    if scheduler_stats["post_settle_started_at"]
                    else None
                ),
                "post_settle_composed_vectors_per_second": (
                    scheduler_stats["post_settle_composed"]
                    / max(0.001, time.monotonic() - scheduler_stats["post_settle_started_at"])
                    if scheduler_stats["post_settle_started_at"]
                    else None
                ),
                "repos": len(completed_repos),
                "existing_pages_submitted": submitted_existing_pages,
                "page_size": page_size,
                "native_shard_pages": args.native_shard_pages,
                "pages_submitted": submitted_pages,
            },
            sort_keys=True,
        )
    )
    return 1 if errors else 0




class _DefaultsHelpFormatter(argparse.ArgumentDefaultsHelpFormatter):
    """Show each option's default in --help, including options with no help= text.

    ArgumentDefaultsHelpFormatter only appends "(default: ...)" to options that already have a help
    string, and argparse hides the help column entirely when an option has none. Almost none of our
    options set help=, so we give those a "(default: %(default)s)" template before formatting; the
    "%(default)" token also stops the base formatter from double-appending.
    """

    def _format_action(self, action: argparse.Action) -> str:
        if (
            not action.help
            and action.option_strings
            and action.default is not argparse.SUPPRESS
            and action.dest != "help"
        ):
            action.help = "(default: %(default)s)"
        return super()._format_action(action)


class _DefaultsParser(argparse.ArgumentParser):
    """ArgumentParser that shows each option's default in --help.

    Used for the top-level parser; argparse propagates this class to every subparser created via
    add_subparsers (parser_class defaults to type(self)), so all subcommands inherit it.
    """

    def __init__(self, *args: object, **kwargs: object) -> None:
        kwargs.setdefault("formatter_class", _DefaultsHelpFormatter)
        super().__init__(*args, **kwargs)


def make_arg_parser() -> argparse.ArgumentParser:
    parser = _DefaultsParser(description="Build SFT code-localization datasets and embeddings")
    parser.add_argument("--commits-root", type=Path, default=DEFAULT_COMMITS_ROOT)
    parser.add_argument("--tasks-dir", type=Path, default=DEFAULT_TASKS_DIR)
    parser.add_argument("--clones-dir", type=Path, default=DEFAULT_CLONES_DIR)
    parser.add_argument("--embeddings-dir", type=Path, default=DEFAULT_EMBEDDINGS_DIR)
    parser.add_argument("--debug", action="store_true")
    subparsers = parser.add_subparsers(dest="command", required=True)

    def add_repo_filters(command: argparse.ArgumentParser) -> None:
        command.add_argument("--repo", action="append", dest="repos")
        command.add_argument("--repo-file", type=Path)
        command.add_argument("--limit-repos", type=int)

    def add_debug_flag(command: argparse.ArgumentParser) -> None:
        command.add_argument("--debug", action="store_true")

    doctor = subparsers.add_parser("doctor", help="Check local dependencies and CUDA readiness")
    add_debug_flag(doctor)
    doctor.add_argument("--require-gpu", action="store_true")

    select = subparsers.add_parser("select-tasks", help="Select up to 1000 primary tasks per language")
    add_debug_flag(select)
    select.add_argument("--language", action="append", dest="languages")
    select.add_argument("--limit-per-language", type=int, default=DEFAULT_TASK_LIMIT_PER_LANGUAGE)
    select.add_argument("--limit-per-repo", type=int, default=DEFAULT_TASK_LIMIT_PER_REPO)
    select.add_argument("--output", type=Path, default=DEFAULT_EMBEDDINGS_DIR / "selection.jsonl")

    init_db = subparsers.add_parser("init-db", help="Create empty per-repo SQLite databases for selected tasks")
    add_debug_flag(init_db)
    init_db.add_argument("--selection", type=Path, default=DEFAULT_EMBEDDINGS_DIR / "selection.jsonl")

    build = subparsers.add_parser("build-embeddings", help="Build per-repo DBs and vectors with native GPU workers")
    add_debug_flag(build)
    build.add_argument("--selection", type=Path, default=DEFAULT_EMBEDDINGS_DIR / "selection.jsonl")
    build.add_argument("--db-dir", type=Path, required=True)
    build.add_argument("--vector-cache-dir", type=Path, required=True)
    add_repo_filters(build)
    build.add_argument("--max-tasks-per-repo", type=int)
    build.add_argument(
        "--clone-workers",
        type=int,
        default=4,
        help="Parallel git-clone workers for recovering missing repo clones on demand",
    )
    build.add_argument(
        "--skip-clone",
        action="store_true",
        help=(
            "Skip clone recovery and treat every selected repo as available. Safe when the "
            "DBs are already complete for the selection (vector-only builds, e.g. embedding "
            "an existing corpus with a new model on a machine without clones)."
        ),
    )
    build.add_argument(
        "--extract-workers",
        type=int,
        default=DEFAULT_EXTRACT_WORKERS,
        help="Parallel Bifrost extraction clients per repo revision",
    )
    build.add_argument("--repo-workers", type=int, default=max(1, os.cpu_count() or 1))
    build.add_argument(
        "--big-repo-mb",
        type=int,
        default=150,
        help=(
            "Repos whose checked-out source (excl. .git) is >= this many MB are treated "
            "as memory-heavy 'big' repos: spread apart in the schedule and gated so at "
            "most --max-big-concurrent run at once (prevents the multi-giant OOM-stall)."
        ),
    )
    build.add_argument(
        "--max-big-concurrent",
        type=int,
        default=1,
        help="Max big repos extracting concurrently (Semaphore over the build-db pool).",
    )
    build.add_argument("--no-resume", action="store_true")
    build.add_argument("--gpu-worker", action="append", type=parse_native_gpu_worker, default=[])
    build.add_argument("--model", default=GRANITE_MODEL)
    build.add_argument("--batch-size", type=int, default=16)
    build.add_argument("--max-seq-length", type=int, default=MAX_SEQ_LENGTH)
    build.add_argument("--bifrost-library", type=Path)
    build.add_argument("--start-base-revision")
    build.add_argument("--end-base-revision")
    build.add_argument("--force-new-cache", action="store_true")
    build.add_argument("--recreate-manifest", action="store_true")
    build.add_argument(
        "--native-shard-pages",
        type=int,
        default=2,
        help="Logical vector-planning pages coalesced into each native GPU task",
    )
    build.add_argument(
        "--native-fixed-batches",
        action="store_true",
        help="Use configured GPU worker batch sizes directly instead of autotuning",
    )
    build.add_argument(
        "--native-long-token-threshold",
        type=int,
        default=0,
        help="Route tasks with a component at or above this estimated token length to preferred GPUs",
    )
    build.add_argument(
        "--native-long-page-gpu",
        action="append",
        default=[],
        help="Physical GPU id allowed to handle long-token tasks; repeat for multiple GPUs",
    )
    build.add_argument(
        "--native-planner-target-tokens",
        type=int,
        default=65_536,
        help="Scheduler-side estimated-token target per native GPU task",
    )
    build.add_argument(
        "--native-planner-max-components",
        type=int,
        default=256,
        help="Scheduler-side max components per native GPU task bucket",
    )
    build.add_argument(
        "--native-target-padded-tokens",
        type=int,
        default=65_536,
        help="Native helper microbatch budget for batch_len * max_estimated_tokens",
    )
    build.add_argument(
        "--native-target-attention-tokens",
        type=int,
        default=268_435_456,
        help="Native helper microbatch budget for batch_len * max_estimated_tokens^2",
    )
    build.add_argument(
        "--native-encode-dtype",
        choices=("float32", "float16", "bfloat16"),
        default="float32",
        help="Experimental native helper inference dtype",
    )
    build.add_argument(
        "--native-attn-implementation",
        choices=("default", "eager", "sdpa", "flash_attention_2"),
        default="default",
        help="Experimental native helper attention implementation passed to Transformers",
    )

    build_index = subparsers.add_parser(
        "build-vector-index",
        help="Pack cached per-vector files into manifest-scoped per-revision matrices",
    )
    add_debug_flag(build_index)
    build_index.add_argument("--selection", type=Path, default=DEFAULT_EMBEDDINGS_DIR / "selection.jsonl")
    build_index.add_argument("--db-dir", type=Path, required=True)
    build_index.add_argument("--vector-cache-dir", type=Path, required=True)
    build_index.add_argument("--split", default="train")
    add_repo_filters(build_index)
    build_index.add_argument("--workers", type=int, default=max(1, min(32, os.cpu_count() or 1)))
    build_index.add_argument("--load-workers", type=int, default=max(1, min(16, os.cpu_count() or 1)))

    refresh = subparsers.add_parser("refresh-positives", help="Recompute task positives from existing indexed chunks")
    add_debug_flag(refresh)
    refresh.add_argument("--selection", type=Path, default=DEFAULT_EMBEDDINGS_DIR / "selection.jsonl")
    add_repo_filters(refresh)
    refresh.add_argument("--repo-workers", type=int, default=4)
    refresh.add_argument("--task-workers", type=int, default=max(1, min(32, os.cpu_count() or 1)))

    augment = subparsers.add_parser(
        "augment-chunkless",
        help="Add synthetic file-summary rows for in-scope files that produced zero function chunks",
    )
    add_debug_flag(augment)
    add_repo_filters(augment)
    augment.add_argument("--db-dir", type=Path, required=True)
    augment.add_argument(
        "--vector-cache-dir",
        action="append",
        type=Path,
        default=[],
        help=(
            "Existing passage vector cache to reconcile once after all repo workers finish; "
            "repeat for multiple caches."
        ),
    )
    augment.add_argument(
        "--seed-dir",
        type=Path,
        help="Optional per-repo synthetic_rows.jsonl directory used to seed exact chunkless summary text.",
    )
    augment.add_argument(
        "--seed-only",
        action="store_true",
        help=(
            "Process only revisions covered by --seed-dir and insert only seeded texts with zero "
            "Bifrost calls. This still writes augment_chunkless:<revision> done markers for those "
            "revisions, so a later full pass will skip them unless you clear the metadata markers."
        ),
    )
    augment.add_argument(
        "--repo-workers",
        type=int,
        default=8,
        help="CPU repo workers for snapshot scanning and DB writes; vector reconcile runs once at the end.",
    )
    augment.add_argument(
        "--summary-workers",
        type=int,
        default=DEFAULT_EXTRACT_WORKERS,
        help="Parallel summary threads within each repo revision.",
    )
    augment.add_argument("--page-size", type=int, default=1024)
    augment.add_argument("--dry-run", action="store_true")
    augment.add_argument("--model", default=GRANITE_MODEL)
    augment.add_argument("--batch-size", type=int, default=16)
    augment.add_argument("--max-seq-length", type=int, default=MAX_SEQ_LENGTH)

    prune_chunks = subparsers.add_parser("prune-orphan-chunks", help="Delete chunks not referenced by any revision")
    add_debug_flag(prune_chunks)
    prune_chunks.add_argument("--selection", type=Path, default=DEFAULT_EMBEDDINGS_DIR / "selection.jsonl")
    add_repo_filters(prune_chunks)

    evaluate = subparsers.add_parser("eval", help="Evaluate a model on the selected test split")
    add_debug_flag(evaluate)
    evaluate.add_argument("--selection", type=Path, required=True)
    evaluate.add_argument("--db-dir", type=Path, required=True)
    evaluate.add_argument("--vector-cache-dir", type=Path, required=True)
    evaluate.add_argument("--split", default="test")
    evaluate.add_argument("--gpu-worker", type=parse_single_gpu_worker)
    evaluate.add_argument("--cuda-visible-devices", help=argparse.SUPPRESS)
    evaluate.add_argument("--model", default=GRANITE_MODEL)
    evaluate.add_argument("--batch-size", type=int, default=16)
    evaluate.add_argument("--max-seq-length", type=int, default=MAX_SEQ_LENGTH)
    evaluate.add_argument("--skip-missing-vectors", action="store_true")
    evaluate.add_argument(
        "--score-report-top-k",
        type=int,
        default=10,
        help="Include file-score diagnostics for the top K files in eval reports; use 0 to disable.",
    )
    evaluate.add_argument("--output", type=Path, default=DEFAULT_EMBEDDINGS_DIR / "baseline-eval.json")
    evaluate.add_argument("--allow-selection-mismatch", action="store_true")

    compare = subparsers.add_parser(
        "compare-evals", help="Paired flip / bootstrap comparison of base vs trained eval reports"
    )
    add_debug_flag(compare)
    compare.add_argument("--base", type=Path, required=True)
    compare.add_argument("--trained", type=Path, required=True)
    compare.add_argument("-k", type=int, default=10)
    compare.add_argument("--bootstrap-iterations", type=int, default=1000)
    compare.add_argument("--seed", type=int, default=0)
    compare.add_argument("--allow-selection-mismatch", action="store_true")

    mine_ready = subparsers.add_parser("mine-ready-negatives", help="Mine negatives for tasks whose full candidate set is vectorized")
    add_debug_flag(mine_ready)
    mine_ready.add_argument("--selection", type=Path, required=True)
    mine_ready.add_argument("--db-dir", type=Path, required=True)
    mine_ready.add_argument("--vector-cache-dir", type=Path, required=True)
    mine_ready.add_argument("--split", default="train")
    mine_ready.add_argument("--output", type=Path, default=DEFAULT_EMBEDDINGS_DIR / "hard-negatives-ready.jsonl")
    mine_ready.add_argument("--gpu-worker", type=parse_single_gpu_worker)
    mine_ready.add_argument("--cuda-visible-devices", help=argparse.SUPPRESS)
    mine_ready.add_argument("--model", default=GRANITE_MODEL)
    mine_ready.add_argument("--batch-size", type=int, default=256)
    mine_ready.add_argument("--max-seq-length", type=int, default=MAX_SEQ_LENGTH)
    mine_ready.add_argument("--max-tasks", type=int)
    mine_ready.add_argument(
        "--shard-repo-workers",
        type=int,
        default=4,
        help="Parallel repo workers within the current shard; GPU embedding stays serialized per process.",
    )
    mine_ready.add_argument("--scan-top-k", type=int, default=DEFAULT_SCAN_TOP_K)
    mine_ready.add_argument("--hard-negatives", type=int, default=HARD_NEGATIVES)
    mine_ready.add_argument(
        "--existing-negative-file",
        action="append",
        type=Path,
        default=[],
        help="Additional mined-negative files whose rows should be treated as already complete.",
    )
    mine_ready.add_argument(
        "--repo-shard",
        type=parse_repo_shard,
        help="Mine only repos whose stable hash maps to INDEX/COUNT.",
    )
    mine_ready.add_argument("--allow-selection-mismatch", action="store_true")

    export_train = subparsers.add_parser("export-train", help="Export training rows with positives and mined negatives")
    add_debug_flag(export_train)
    export_train.add_argument("--selection", type=Path, required=True)
    export_train.add_argument("--db-dir", type=Path, required=True)
    export_train.add_argument("--split", default="train")
    export_train.add_argument("--negative-file", action="append", type=Path, default=[])
    export_train.add_argument("--output", type=Path, default=DEFAULT_EMBEDDINGS_DIR / "train-pairs.jsonl")
    export_train.add_argument("--negatives-per-row", type=int, default=DEFAULT_NEGATIVES_PER_ROW)
    export_train.add_argument(
        "--repo-workers",
        type=int,
        default=1,
        help="Export repos in parallel into shard files, then concatenate.",
    )
    export_train.add_argument(
        "--max-positives-per-task",
        type=int,
        default=0,
        help="Cap row-local positive rows per task after deterministic ordering; 0 means no cap.",
    )
    export_train.add_argument("--allow-selection-mismatch", action="store_true")

    return parser


def main(argv: Sequence[str] | None = None) -> int:
    parser = make_arg_parser()
    args = parser.parse_args(argv)
    if args.command == "doctor":
        return print_doctor(doctor_checks(require_gpu=args.require_gpu))
    if args.command == "select-tasks":
        records = select_tasks(
            args.commits_root,
            args.tasks_dir,
            languages=args.languages,
            limit_per_language=args.limit_per_language,
            limit_per_repo=args.limit_per_repo,
            show_progress=True,
        )
        canonical_path, sidecar_path, sidecar = write_selection_artifacts(records, args.output)
        print(
            f"wrote {len(records)} task records to {canonical_path} "
            f"(legacy updated at {args.output}; sidecar {sidecar_path}; selection_hash={sidecar['selection_hash']['short12']})",
            file=sys.stderr,
        )
        return 0
    if args.command == "init-db":
        records = read_selection(args.selection)
        selection_hash = selection_content_hash(records)
        for repo in sorted({record.repo for record in records}):
            conn = init_embeddings_db(args.embeddings_dir / repo / "embeddings.db")
            set_selection_hash_metadata(conn, selection_hash)
            conn.close()
        return 0
    if args.command == "build-embeddings":
        if args.extract_workers < 1:
            parser.error("--extract-workers must be >= 1")
        if args.repo_workers < 1:
            parser.error("--repo-workers must be >= 1")
        if args.native_shard_pages < 1:
            parser.error("--native-shard-pages must be >= 1")
        if args.native_long_token_threshold < 0:
            parser.error("--native-long-token-threshold must be >= 0")
        if args.native_long_page_gpu and args.native_long_token_threshold == 0:
            parser.error("--native-long-page-gpu requires --native-long-token-threshold")
        if args.native_long_token_threshold > 0:
            if not args.native_long_page_gpu:
                parser.error("--native-long-token-threshold requires at least one --native-long-page-gpu")
            worker_gpus = {worker.gpu for worker in (args.gpu_worker or [NativeGpuWorkerSpec("0", args.batch_size)])}
            unknown_long_gpus = sorted(set(args.native_long_page_gpu) - worker_gpus)
            if unknown_long_gpus:
                parser.error(
                    "--native-long-page-gpu must refer to configured --gpu-worker physical GPUs: "
                    + ", ".join(unknown_long_gpus)
                )
        if args.native_planner_target_tokens < 1:
            parser.error("--native-planner-target-tokens must be >= 1")
        if args.native_planner_max_components < 1:
            parser.error("--native-planner-max-components must be >= 1")
        if args.native_target_padded_tokens < 1:
            parser.error("--native-target-padded-tokens must be >= 1")
        if args.native_target_attention_tokens < 1:
            parser.error("--native-target-attention-tokens must be >= 1")
        return run_build_embeddings(args)
    if args.command == "augment-chunkless":
        if args.repo_workers < 1:
            parser.error("--repo-workers must be >= 1")
        if args.summary_workers < 1:
            parser.error("--summary-workers must be >= 1")
        if args.page_size < 1:
            parser.error("--page-size must be >= 1")
        if not args.dry_run and not args.vector_cache_dir:
            parser.error("--vector-cache-dir is required unless --dry-run is set")
        return run_augment_chunkless(args)
    if args.command == "build-vector-index":
        if args.workers < 1:
            parser.error("--workers must be >= 1")
        if args.load_workers < 1:
            parser.error("--load-workers must be >= 1")
        return run_build_vector_index(args)
    if args.command == "refresh-positives":
        if args.repo_workers < 1:
            parser.error("--repo-workers must be >= 1")
        if args.task_workers < 1:
            parser.error("--task-workers must be >= 1")
        records = read_selection(args.selection)
        grouped = group_tasks_by_repo(records)
        repos = sorted(grouped)
        requested = requested_repo_filter(args)
        if requested is not None:
            repos = [repo for repo in repos if repo in requested]
        if args.limit_repos is not None:
            repos = repos[: args.limit_repos]
        errors = 0
        stats_rows = []
        workers = min(args.repo_workers, len(repos)) if repos else 1
        progress = tqdm(total=len(repos), desc="refresh-positives", unit="repo", dynamic_ncols=True)
        with concurrent.futures.ProcessPoolExecutor(max_workers=workers) as pool:
            futures = {
                pool.submit(
                    refresh_repo_positives_worker,
                    grouped[repo],
                    args.clones_dir,
                    args.embeddings_dir,
                    args.task_workers,
                ): repo
                for repo in repos
            }
            for future in concurrent.futures.as_completed(futures):
                progress.update(1)
                stats_rows.append(future.result())
        progress.close()
        for stats in sorted(stats_rows, key=lambda item: item.repo):
            if stats.error is not None:
                errors += 1
            print(
                json.dumps(
                    {
                        "repo": stats.repo,
                        "tasks": stats.tasks,
                        "refreshed": stats.refreshed,
                        "positives": stats.positives,
                        "skipped_missing_task": stats.skipped_missing_task,
                        "skipped_missing_chunks": stats.skipped_missing_chunks,
                        "error": stats.error,
                    },
                    sort_keys=True,
                ),
                file=sys.stderr,
            )
        return 1 if errors else 0
    if args.command == "prune-orphan-chunks":
        records = read_selection(args.selection)
        repos = sorted(group_tasks_by_repo(records))
        requested = requested_repo_filter(args)
        if requested is not None:
            repos = [repo for repo in repos if repo in requested]
        if args.limit_repos is not None:
            repos = repos[: args.limit_repos]
        total_deleted = 0
        for repo in repos:
            deleted = prune_orphan_chunks_for_repo(repo, embeddings_dir=args.embeddings_dir)
            total_deleted += deleted
            print(json.dumps({"repo": repo, "deleted": deleted}, sort_keys=True), file=sys.stderr)
        print(json.dumps({"deleted": total_deleted}, sort_keys=True))
        return 0
    if args.command == "eval":
        records = read_selection(args.selection)
        selection_hash = verify_selection_hash_against_dbs(
            records,
            db_dir=args.db_dir,
            allow_selection_mismatch=args.allow_selection_mismatch,
            warning_printer=_warn_stderr,
        )
        if args.gpu_worker and args.cuda_visible_devices:
            parser.error("--gpu-worker cannot be combined with --cuda-visible-devices")
        batch_size = (
            args.gpu_worker.batch_size
            if args.gpu_worker is not None and args.gpu_worker.batch_size is not None
            else args.batch_size
        )
        configure_cuda_visibility(
            args.gpu_worker.gpu if args.gpu_worker is not None else None,
            args.cuda_visible_devices,
        )
        embedder, query_manifest = make_embedder_with_manifest(
            args.model,
            device="cuda:0",
            batch_size=batch_size,
            max_seq_length=args.max_seq_length,
            role="query",
        )
        assert_query_side_compatible(args.vector_cache_dir, query_manifest)
        evaluations, metrics = evaluate_records(
            records,
            db_dir=args.db_dir,
            vector_cache_dir=args.vector_cache_dir,
            embed_texts=embedder,
            manifest_digest=query_manifest.digest(),
            split=args.split,
            skip_missing_vectors=args.skip_missing_vectors,
            score_report_top_k=args.score_report_top_k,
            incremental_path=args.output.with_suffix(".partial.jsonl"),
        )
        write_eval_report(args.output, evaluations, metrics, selection_hash=selection_hash)
        print(
            json.dumps(
                {
                    "selection_hash": selection_hash,
                    "total": metrics.total,
                    "all_positive_acc": metrics.all_positive_acc,
                    "any_positive_acc": metrics.any_positive_acc,
                    "mrr": metrics.mrr,
                    "output": str(args.output),
                },
                sort_keys=True,
            )
        )
        return 0
    if args.command == "compare-evals":
        report = compare_eval_reports(
            args.base,
            args.trained,
            k=args.k,
            iterations=args.bootstrap_iterations,
            seed=args.seed,
            allow_selection_mismatch=args.allow_selection_mismatch,
        )
        print(json.dumps(report, indent=2, sort_keys=True))
        return 0
    if args.command == "mine-ready-negatives":
        records = filter_records_by_repo_shard(read_selection(args.selection), args.repo_shard)
        selection_hash = verify_selection_hash_against_dbs(
            records,
            db_dir=args.db_dir,
            allow_selection_mismatch=args.allow_selection_mismatch,
            warning_printer=_warn_stderr,
        )
        if args.shard_repo_workers < 1:
            parser.error("--shard-repo-workers must be >= 1")
        if args.gpu_worker and args.cuda_visible_devices:
            parser.error("--gpu-worker cannot be combined with --cuda-visible-devices")
        batch_size = (
            args.gpu_worker.batch_size
            if args.gpu_worker is not None and args.gpu_worker.batch_size is not None
            else args.batch_size
        )
        configure_cuda_visibility(
            args.gpu_worker.gpu if args.gpu_worker is not None else None,
            args.cuda_visible_devices,
        )
        embedder, query_manifest = make_embedder_with_manifest(
            args.model,
            device="cuda:0",
            batch_size=batch_size,
            max_seq_length=args.max_seq_length,
            role="query",
        )
        assert_query_side_compatible(args.vector_cache_dir, query_manifest)
        total_tasks = sum(1 for record in records if record.split == args.split)
        progress = tqdm(total=total_tasks, desc="mine-ready-negatives", unit="task", dynamic_ncols=True)
        last_ready = 0
        progress_lock = threading.Lock()

        def on_progress(event: dict[str, object]) -> None:
            nonlocal last_ready
            with progress_lock:
                if event.get("event") == "mine_progress":
                    ready = int(event.get("ready", 0))
                    if ready > last_ready:
                        progress.update(ready - last_ready)
                        last_ready = ready
                    repo = str(event.get("repo", ""))
                    revision = str(event.get("revision", ""))[:12]
                    stage = str(event.get("stage", ""))
                    written = int(event.get("written", 0))
                    if revision:
                        progress.set_postfix_str(f"{repo} {stage} {revision} written={written}")
                    else:
                        progress.set_postfix_str(f"{repo} {stage} written={written}")
                elif event.get("event") == "mine_repo_complete":
                    print(
                        json.dumps(
                            {
                                "repo": str(event["repo"]),
                                "considered": int(event["considered"]),
                                "ready": int(event["ready"]),
                                "written": int(event["written"]),
                                "revisions_processed": int(event["revisions_processed"]),
                                "prompt_read_seconds": round(float(event["prompt_read_seconds"]), 2),
                                "query_embed_seconds": round(float(event["query_embed_seconds"]), 2),
                                "matrix_load_seconds": round(float(event["matrix_load_seconds"]), 2),
                                "gpu_score_seconds": round(float(event["gpu_score_seconds"]), 2),
                                "negative_write_seconds": round(float(event["negative_write_seconds"]), 2),
                            },
                            sort_keys=True,
                        ),
                        file=sys.stderr,
                    )
                elif event.get("event") == "mine_repo_error":
                    print(
                        json.dumps(
                            {
                                "repo": str(event["repo"]),
                                "error": str(event["error"]),
                            },
                            sort_keys=True,
                        ),
                        file=sys.stderr,
                    )
        stats = mine_ready_negatives(
            records,
            db_dir=args.db_dir,
            vector_cache_dir=args.vector_cache_dir,
            embed_texts=embedder,
            manifest_digest=query_manifest.digest(),
            output=args.output,
            split=args.split,
            max_tasks=args.max_tasks,
            scan_top_k=args.scan_top_k,
            hard_negative_count=args.hard_negatives,
            existing_negative_files=args.existing_negative_file,
            selection_hash=selection_hash,
            progress=on_progress,
            shard_repo_workers=args.shard_repo_workers,
        )
        if stats.ready > last_ready:
            progress.update(stats.ready - last_ready)
        progress.close()
        print(
            json.dumps(
                {
                    "output": str(args.output),
                    "selection_hash": selection_hash,
                    "considered": stats.considered,
                    "ready": stats.ready,
                    "written": stats.written,
                    "skipped_existing": stats.skipped_existing,
                    "skipped_no_positive": stats.skipped_no_positive,
                    "skipped_incomplete_vectors": stats.skipped_incomplete_vectors,
                    "prompt_read_seconds": round(stats.prompt_read_seconds, 2),
                    "query_embed_seconds": round(stats.query_embed_seconds, 2),
                    "matrix_load_seconds": round(stats.matrix_load_seconds, 2),
                    "gpu_score_seconds": round(stats.gpu_score_seconds, 2),
                    "negative_write_seconds": round(stats.negative_write_seconds, 2),
                    "repaired_trailing_lines": stats.repaired_trailing_lines,
                    "errors": len(stats.repo_errors),
                    "repo_errors": [
                        {"repo": repo, "error": error}
                        for repo, error in stats.repo_errors
                    ],
                    "unique_eligible_negative_files_p10": stats.unique_eligible_negative_files_p10,
                    "unique_eligible_negative_files_p50": stats.unique_eligible_negative_files_p50,
                    "unique_eligible_negative_files_p95": stats.unique_eligible_negative_files_p95,
                    "fraction_ready_with_8_plus_negative_files": stats.fraction_ready_with_8_plus_negative_files,
                },
                sort_keys=True,
            )
        )
        return 1 if stats.repo_errors else 0
    if args.command == "export-train":
        records = read_selection(args.selection)
        selection_hash = verify_selection_hash_against_dbs(
            records,
            db_dir=args.db_dir,
            allow_selection_mismatch=args.allow_selection_mismatch,
            warning_printer=_warn_stderr,
        )
        if args.repo_workers < 1:
            parser.error("--repo-workers must be >= 1")
        if args.repo_workers > 1:
            grouped = group_tasks_by_repo(records)
            repos = sorted(
                repo for repo, repo_records in grouped.items() if any(record.split == args.split for record in repo_records)
            )
            shard_dir = args.output.parent / f"{args.output.name}.shards"
            shard_dir.mkdir(parents=True, exist_ok=True)
            workers = min(args.repo_workers, len(repos)) if repos else 1
            progress = tqdm(total=len(repos), desc="export-train", unit="repo", dynamic_ncols=True)
            rows: list[dict[str, object]] = []
            with concurrent.futures.ProcessPoolExecutor(max_workers=workers) as pool:
                futures = {
                    pool.submit(
                        export_train_repo_worker,
                        repo,
                        grouped[repo],
                        args.db_dir,
                        args.negative_file,
                        shard_dir / f"{repo}.jsonl",
                        args.split,
                        args.negatives_per_row,
                        args.max_positives_per_task,
                        selection_hash,
                    ): repo
                    for repo in repos
                }
                for future in concurrent.futures.as_completed(futures):
                    repo = futures[future]
                    row = future.result()
                    rows.append(row)
                    progress.update(1)
                    progress.set_postfix_str(
                        f"{repo} rows={int(row.get('rows', 0))} tasks={int(row.get('tasks_exported', 0))}"
                    )
            progress.close()
            rows.sort(key=lambda row: str(row["repo"]))
            concatenate_jsonl_shards([Path(str(row["output"])) for row in rows], args.output)
            summary = aggregate_parallel_export_stats(rows, args.output)
            summary["selection_hash"] = selection_hash
            print(json.dumps(summary, sort_keys=True), file=sys.stderr)
            return 0
        stats = export_training_examples(
            records,
            db_dir=args.db_dir,
            negative_files=args.negative_file,
            output=args.output,
            split=args.split,
            negatives_per_row=args.negatives_per_row,
            max_positives_per_task=args.max_positives_per_task,
        )
        print(
            json.dumps(
                {
                    "output": str(args.output),
                    "selection_hash": selection_hash,
                    "rows": stats.rows,
                    "tasks_exported": stats.tasks_exported,
                    "skipped_no_positive": stats.skipped_no_positive,
                    "skipped_too_few_negatives": stats.skipped_too_few_negatives,
                    "positives_per_task_p50": stats.positives_per_task_p50,
                    "positives_per_task_p95": stats.positives_per_task_p95,
                    "positives_per_task_max": stats.positives_per_task_max,
                    "num_tasks_with_0_negatives": stats.num_tasks_with_0_negatives,
                    "num_tasks_with_1_7_negatives": stats.num_tasks_with_1_7_negatives,
                    "num_tasks_with_8_plus_negatives": stats.num_tasks_with_8_plus_negatives,
                    "mined_negatives_per_task_p50": stats.mined_negatives_per_task_p50,
                    "mined_negatives_per_task_p95": stats.mined_negatives_per_task_p95,
                    "mined_negatives_per_task_max": stats.mined_negatives_per_task_max,
                    "target_files_per_task_p50": stats.target_files_per_task_p50,
                    "target_files_per_task_p95": stats.target_files_per_task_p95,
                    "target_files_per_task_max": stats.target_files_per_task_max,
                    "positive_chunks_per_target_file_p50": stats.positive_chunks_per_target_file_p50,
                    "positive_chunks_per_target_file_p95": stats.positive_chunks_per_target_file_p95,
                    "positive_chunks_per_target_file_max": stats.positive_chunks_per_target_file_max,
                    "fraction_tasks_dropped_too_few_distinct_negative_files": (
                        stats.fraction_tasks_dropped_too_few_distinct_negative_files
                    ),
                    "true_gold_files_per_task_p50": stats.true_gold_files_per_task_p50,
                    "true_gold_files_per_task_p95": stats.true_gold_files_per_task_p95,
                    "true_gold_files_per_task_min": stats.true_gold_files_per_task_min,
                    "exported_gold_file_groups_per_task_p50": (
                        stats.exported_gold_file_groups_per_task_p50
                    ),
                    "exported_gold_file_groups_per_task_p95": (
                        stats.exported_gold_file_groups_per_task_p95
                    ),
                    "exported_gold_file_groups_per_task_max": (
                        stats.exported_gold_file_groups_per_task_max
                    ),
                    "missing_gold_file_groups_per_task_p50": stats.missing_gold_file_groups_per_task_p50,
                    "missing_gold_file_groups_per_task_p95": stats.missing_gold_file_groups_per_task_p95,
                    "missing_gold_file_groups_per_task_max": stats.missing_gold_file_groups_per_task_max,
                    "task_weight_exported_sum_p50": stats.task_weight_exported_sum_p50,
                    "task_weight_exported_sum_p95": stats.task_weight_exported_sum_p95,
                    "task_weight_exported_sum_min": stats.task_weight_exported_sum_min,
                    "fraction_tasks_with_partial_positive_coverage": (
                        stats.fraction_tasks_with_partial_positive_coverage
                    ),
                    "positive_slot_old_hunk": stats.positive_slot_old_hunk,
                    "positive_slot_class_summary_fallback": stats.positive_slot_class_summary_fallback,
                    "positive_slot_file_summary_fallback": stats.positive_slot_file_summary_fallback,
                    "valid_positive_slots_1": stats.valid_positive_slots_1,
                    "valid_positive_slots_2": stats.valid_positive_slots_2,
                    "valid_positive_slots_3": stats.valid_positive_slots_3,
                    "valid_positive_slots_4": stats.valid_positive_slots_4,
                    "valid_positive_negative_pairs": stats.valid_positive_negative_pairs,
                    "rows_with_3_plus_old_hunks": stats.rows_with_3_plus_old_hunks,
                    "rows_with_4_old_hunks": stats.rows_with_4_old_hunks,
                    "fallback_only_rows": stats.fallback_only_rows,
                },
                sort_keys=True,
            ),
            file=sys.stderr,
        )
        return 0
    raise AssertionError(f"unhandled command: {args.command}")


if __name__ == "__main__":
    raise SystemExit(main(sys.argv[1:]))