DocForge2/tools/milestone3_benchmark.py

"""Milestone 3 projection, worker, fragment, and publication benchmark gates."""

from __future__ import annotations

import argparse
import gc
import hashlib
import json
import math
import platform
import resource
import statistics
import subprocess
import sys
import tempfile
import time
import tracemalloc
from collections.abc import Callable, Mapping, Sequence
from dataclasses import dataclass, replace
from pathlib import Path
from typing import cast

from milestone0_baseline import write_synthetic_project

from docforge.graph_projection import (
    GraphViewRequestV1,
    build_graph_projection_package,
    build_graph_view_plan,
)
from docforge.graph_rendering import GraphRenderService
from docforge.manual_projection import (
    build_manual_projection_package,
    build_manual_render_plan,
)
from docforge.models import ProjectSnapshot
from docforge.project import Project
from docforge.projection_contract import (
    MAX_PACKAGE_BYTES,
    MAX_PLAN_BYTES,
    MAX_RECEIPT_BYTES,
    GraphViewPlanV1,
    ManualRenderPlanV1,
    ProjectionPackageV1,
    ProjectionRenderResult,
    canonical_projection_bytes,
    projection_hash,
)
from docforge.projection_fragments import (
    FragmentKey,
    FragmentRecord,
    ProjectionFragmentCache,
    fragment_semantic_hash,
)
from docforge.projection_worker import (
    MAX_WORKER_ARTIFACT_BYTES,
    render_projection_in_worker,
)
from docforge.render_contract import GenericHtmlRenderer, PreparedRender
from docforge.rendering import RenderService
from docforge.telemetry import COUNTER_NAMES, request
from docforge_renderers.graph import PortableGraphHtmlRenderer
from docforge_renderers.manual import ManualHtmlRenderer

ROOT = Path(__file__).resolve().parents[1]
FULL_NODE_COUNT = 1_000
SMOKE_NODE_COUNT = 25
DEFAULT_FULL_SAMPLES = 3
DEFAULT_SMOKE_SAMPLES = 1
MAX_STATUS_RESPONSE_BYTES = 256_000
MAX_TRACED_PEAK_BYTES = 256 * 1024 * 1024
MAX_CHILD_PEAK_BYTES = 256 * 1024 * 1024

ZERO_WORK_COUNTERS = tuple(
    counter for counter in COUNTER_NAMES if counter != "source_generation_checks"
)


@dataclass(frozen=True)
class _ProjectionSample:
    plan_id: str
    package_id: str
    artifact: bytes
    receipt: Mapping[str, object]
    plan_bytes: int
    package_bytes: int
    child_peak_memory_bytes: int | None


@dataclass(frozen=True)
class _StatusSample:
    response: Mapping[str, object]
    counters: Mapping[str, object]


@dataclass(frozen=True)
class _FragmentSweep:
    fragment_count: int
    aggregate_content_bytes: int
    ordered_record_hash: str


def _parser() -> argparse.ArgumentParser:
    parser = argparse.ArgumentParser(
        description="Gate DocForge2 Milestone 3 projection behavior on a disposable project."
    )
    parser.add_argument("--mode", choices=("smoke", "full"), default="full")
    parser.add_argument("--nodes", type=int)
    parser.add_argument("--samples", type=int)
    parser.add_argument("--output", type=Path)
    return parser


def encode_report(value: object) -> str:
    """Serialize one report deterministically for files, CI logs, and comparisons."""

    return json.dumps(value, sort_keys=True, indent=2, ensure_ascii=False) + "\n"


def _git(arguments: list[str]) -> str:
    return subprocess.run(
        ["git", *arguments],
        cwd=ROOT,
        check=True,
        capture_output=True,
        text=True,
    ).stdout.strip()


def _compact_size(value: object) -> int:
    return len(canonical_projection_bytes(value))


def _sha256(content: bytes) -> str:
    return hashlib.sha256(content).hexdigest()


def _prepare_fixture(root: Path, node_count: int) -> None:
    write_synthetic_project(root, node_count)
    descriptor = root / ".docforge" / "project.toml"
    maximum_edges = node_count - 1
    maximum_work = max(100, node_count * 4)
    original = descriptor.read_text(encoding="utf-8")
    original = original.replace(
        "max_render_bytes = 20000000",
        "max_render_bytes = 4000000",
    )
    descriptor.write_text(
        original
        + f"""

[graph_render]
output_root = ".docforge/portable-graph"

[[graph_render.views]]
id = "architecture"
renderer = "portable_graph_html"
output = "architecture.html"
title = "Synthetic architecture"
query = "synthetic measurement"
initial_mode = "web"
depth = 1
max_nodes = {node_count}
max_edges = {maximum_edges}
max_work = {maximum_work}
families = ["guide"]
relations = ["depends_on"]
authorities = []
statuses = ["active"]
tags = []
include_logic = false
""",
        encoding="utf-8",
    )


def _manual_package(
    snapshot: ProjectSnapshot,
) -> tuple[ManualRenderPlanV1, ProjectionPackageV1, str]:
    descriptor = snapshot.descriptor
    render = descriptor.render
    if render is None:
        raise RuntimeError("Milestone 3 fixture has no manual render configuration")
    view = render.views[0]
    renderer_version = GenericHtmlRenderer(incremental=False).renderer_version
    plan = build_manual_render_plan(snapshot, view, changeset_hash=None)
    package = build_manual_projection_package(
        plan,
        view.template_path.read_bytes(),
        renderer_id=ManualHtmlRenderer.renderer_id,
        renderer_version=renderer_version,
        max_output_bytes=descriptor.limits.max_render_bytes,
    )
    return plan, package, renderer_version


def _graph_package(
    snapshot: ProjectSnapshot,
    node_count: int,
) -> tuple[GraphViewPlanV1, ProjectionPackageV1]:
    descriptor = snapshot.descriptor
    plan = build_graph_view_plan(
        snapshot,
        GraphViewRequestV1(
            view_id="architecture",
            title="Synthetic architecture",
            query="synthetic measurement",
            initial_mode="web",
            depth=1,
            max_nodes=node_count,
            max_edges=node_count - 1,
            max_work=max(100, node_count * 4),
            families=("guide",),
            relations=("depends_on",),
            statuses=("active",),
        ),
        False,
    )
    package = build_graph_projection_package(
        plan,
        renderer_id=PortableGraphHtmlRenderer.renderer_id,
        renderer_version=PortableGraphHtmlRenderer.renderer_version,
        max_output_bytes=descriptor.limits.max_render_bytes,
    )
    return plan, package


def _projection_sample(
    plan: ManualRenderPlanV1 | GraphViewPlanV1,
    package: ProjectionPackageV1,
    result: ProjectionRenderResult,
) -> _ProjectionSample:
    if len(result.artifacts) != 1:
        raise RuntimeError("Projection benchmark expected exactly one artifact")
    artifact = result.artifacts[0].content
    peak = result.receipt.document.get("peak_memory_bytes")
    if peak is not None and (type(peak) is not int or peak < 1):
        raise RuntimeError("Detached worker did not report valid peak memory")
    return _ProjectionSample(
        plan_id=plan.plan_id,
        package_id=package.package_id,
        artifact=artifact,
        receipt=result.receipt.as_dict(),
        plan_bytes=_compact_size(plan.as_dict()),
        package_bytes=_compact_size(package.as_dict()),
        child_peak_memory_bytes=peak,
    )


def _projection_summary(sample: _ProjectionSample) -> dict[str, object]:
    return {
        "plan_id": sample.plan_id,
        "package_id": sample.package_id,
        "artifact_sha256": _sha256(sample.artifact),
        "artifact_bytes": len(sample.artifact),
        "plan_bytes": sample.plan_bytes,
        "package_bytes": sample.package_bytes,
    }


def _measure(
    operation: Callable[[], object],
    *,
    samples: int,
    p95_limit_ms: float,
    response_limit_bytes: int,
    summary: Callable[[object], Mapping[str, object]],
    response_size: Callable[[object], int],
    warmups: int = 0,
) -> tuple[dict[str, object], object]:
    for _ in range(warmups):
        operation()
    durations: list[float] = []
    traced_peaks: list[int] = []
    response_sizes: list[int] = []
    stable_summary: Mapping[str, object] | None = None
    last: object = None
    for _ in range(samples):
        gc.collect()
        tracemalloc.start()
        started = time.perf_counter_ns()
        try:
            value = operation()
            elapsed_ms = (time.perf_counter_ns() - started) / 1_000_000
            _, traced_peak = tracemalloc.get_traced_memory()
        finally:
            tracemalloc.stop()
        current_summary = dict(summary(value))
        if stable_summary is None:
            stable_summary = current_summary
        elif current_summary != stable_summary:
            raise RuntimeError("Milestone 3 operation changed deterministic result across samples")
        current_response_size = response_size(value)
        if current_response_size > response_limit_bytes:
            raise RuntimeError(
                "Milestone 3 response exceeded its fixed benchmark boundary: "
                f"{current_response_size} > {response_limit_bytes}"
            )
        if traced_peak > MAX_TRACED_PEAK_BYTES:
            raise RuntimeError(
                "Milestone 3 operation exceeded its traced-memory boundary: "
                f"{traced_peak} > {MAX_TRACED_PEAK_BYTES}"
            )
        durations.append(elapsed_ms)
        traced_peaks.append(traced_peak)
        response_sizes.append(current_response_size)
        last = value
    ordered = sorted(durations)
    p95_index = max(0, math.ceil(len(ordered) * 0.95) - 1)
    p95 = ordered[p95_index]
    if p95 > p95_limit_ms:
        raise RuntimeError(f"Milestone 3 operation p95 {p95:.3f} ms exceeds {p95_limit_ms:.3f} ms")
    assert stable_summary is not None
    return (
        {
            "samples": samples,
            "median_ms": round(statistics.median(ordered), 3),
            "p95_ms": round(p95, 3),
            "min_ms": round(ordered[0], 3),
            "max_ms": round(ordered[-1], 3),
            "p95_limit_ms": p95_limit_ms,
            "maximum_response_bytes": max(response_sizes),
            "response_limit_bytes": response_limit_bytes,
            "maximum_traced_peak_bytes": max(traced_peaks),
            "traced_peak_limit_bytes": MAX_TRACED_PEAK_BYTES,
            "stable_result": dict(stable_summary),
        },
        last,
    )


def _projection_measurement(
    operation: Callable[[], _ProjectionSample],
    *,
    samples: int,
    p95_limit_ms: float,
) -> tuple[dict[str, object], _ProjectionSample]:
    child_peaks: list[int] = []

    def observed_operation() -> _ProjectionSample:
        sample = operation()
        if sample.child_peak_memory_bytes is not None:
            if sample.child_peak_memory_bytes > MAX_CHILD_PEAK_BYTES:
                raise RuntimeError(
                    "Detached worker peak memory "
                    f"{sample.child_peak_memory_bytes} exceeds "
                    f"{MAX_CHILD_PEAK_BYTES} bytes"
                )
            child_peaks.append(sample.child_peak_memory_bytes)
        return sample

    measurement, value = _measure(
        observed_operation,
        samples=samples,
        p95_limit_ms=p95_limit_ms,
        response_limit_bytes=MAX_RECEIPT_BYTES,
        summary=lambda item: _projection_summary(cast(_ProjectionSample, item)),
        response_size=lambda item: _compact_size(cast(_ProjectionSample, item).receipt),
    )
    sample = cast(_ProjectionSample, value)
    if sample.plan_bytes > MAX_PLAN_BYTES or sample.package_bytes > MAX_PACKAGE_BYTES:
        raise RuntimeError("Projection plan or package exceeded its protocol boundary")
    if child_peaks:
        measurement["maximum_child_peak_bytes"] = max(child_peaks)
        measurement["child_peak_limit_bytes"] = MAX_CHILD_PEAK_BYTES
    return measurement, sample


def _profiled_status(operation: Callable[[], Mapping[str, object]]) -> _StatusSample:
    with request("benchmark.m3", enabled=True) as collector:
        response = operation()
    if collector is None:
        raise RuntimeError("Milestone 3 telemetry collector was not created")
    counters = collector.as_dict(outcome="ok")["counters"]
    typed_counters = cast(Mapping[str, object], counters)
    for counter in ZERO_WORK_COUNTERS:
        if typed_counters[counter] != 0:
            raise RuntimeError(f"Receipt-only status performed forbidden work: {counter}")
    if typed_counters["source_generation_checks"] != 2:
        raise RuntimeError("Receipt-only status did not perform its two race-safe source checks")
    if response.get("status") != "ok" or response.get("state") != "current":
        raise RuntimeError("Receipt-only status did not report a current publication")
    return _StatusSample(response=response, counters=typed_counters)


def _status_summary(value: object) -> Mapping[str, object]:
    sample = cast(_StatusSample, value)
    return {
        "response": dict(sample.response),
        "counters": dict(sample.counters),
    }


def _fragment_summary(value: object) -> Mapping[str, object]:
    sample = cast(_FragmentSweep, value)
    return {
        "fragment_count": sample.fragment_count,
        "aggregate_content_bytes": sample.aggregate_content_bytes,
        "ordered_record_hash": sample.ordered_record_hash,
    }


def _fragment_sweep(records: Sequence[FragmentRecord]) -> _FragmentSweep:
    return _FragmentSweep(
        fragment_count=len(records),
        aggregate_content_bytes=sum(record.byte_count for record in records),
        ordered_record_hash=projection_hash([record.record_id for record in records]),
    )


def _prepared_summary(value: object) -> Mapping[str, object]:
    prepared = cast(PreparedRender, value)
    return {
        "render_identity": prepared.render_identity,
        "output_sha256": prepared.output_hash,
        "output_bytes": len(prepared.output),
    }


def _prepared_response_size(value: object) -> int:
    prepared = cast(PreparedRender, value)
    return _compact_size(prepared.projection_receipt)


def _benchmark(root: Path, node_count: int, samples: int) -> dict[str, object]:
    project = Project.open(root)
    snapshot = project.load()
    if len(snapshot.nodes) != node_count or len(snapshot.edges) != node_count - 1:
        raise RuntimeError("Milestone 3 fixture does not have full synthetic coverage")

    operations: dict[str, object] = {}

    def manual_full() -> _ProjectionSample:
        plan, package, renderer_version = _manual_package(snapshot)
        result = ManualHtmlRenderer(renderer_version).render(package)
        return _projection_sample(plan, package, result)

    operations["manual_full_render"], manual_result = _projection_measurement(
        manual_full,
        samples=samples,
        p95_limit_ms=15_000,
    )

    def graph_full() -> _ProjectionSample:
        plan, package = _graph_package(snapshot, node_count)
        result = PortableGraphHtmlRenderer().render(package)
        return _projection_sample(plan, package, result)

    operations["portable_graph_full_render"], graph_result = _projection_measurement(
        graph_full,
        samples=samples,
        p95_limit_ms=10_000,
    )

    manual_plan, manual_package, manual_renderer_version = _manual_package(snapshot)
    graph_plan, graph_package = _graph_package(snapshot, node_count)
    graph_diagnostics = cast(Mapping[str, object], graph_plan.document["diagnostics"])
    manual_pages = cast(list[dict[str, object]], manual_plan.document["pages"])
    if (
        len(manual_pages) != node_count
        or graph_diagnostics["returned_nodes"] != node_count
        or graph_diagnostics["returned_edges"] != node_count - 1
    ):
        raise RuntimeError("Projection plans did not retain every synthetic node and edge")

    manual_worker_measurement, manual_worker = _projection_measurement(
        lambda: _projection_sample(
            manual_plan,
            manual_package,
            render_projection_in_worker(manual_package),
        ),
        samples=samples,
        p95_limit_ms=20_000,
    )
    operations["manual_detached_worker"] = manual_worker_measurement
    manual_worker_peak = cast(
        int,
        manual_worker_measurement["maximum_child_peak_bytes"],
    )

    graph_worker_measurement, graph_worker = _projection_measurement(
        lambda: _projection_sample(
            graph_plan,
            graph_package,
            render_projection_in_worker(graph_package),
        ),
        samples=samples,
        p95_limit_ms=20_000,
    )
    operations["portable_graph_detached_worker"] = graph_worker_measurement
    graph_worker_peak = cast(
        int,
        graph_worker_measurement["maximum_child_peak_bytes"],
    )
    if (
        manual_worker.artifact != manual_result.artifact
        or graph_worker.artifact != graph_result.artifact
    ):
        raise RuntimeError("Detached worker output is not byte-equivalent to in-process output")

    manual_renderer = ManualHtmlRenderer(manual_renderer_version)
    fragment_records = [
        FragmentRecord.create(
            FragmentKey.create(
                projection_kind="manual",
                renderer_id=ManualHtmlRenderer.renderer_id,
                renderer_version=manual_renderer_version,
                component_version=GenericHtmlRenderer.page_component_version,
                semantic_input_hash=fragment_semantic_hash(page),
            ),
            manual_renderer.render_page_fragment(page).encode("utf-8"),
        )
        for page in manual_pages
    ]
    fragment_cache = ProjectionFragmentCache(
        root,
        root / ".docforge" / "milestone3-fragment-cache",
    )

    def fragment_misses() -> _FragmentSweep:
        if any(fragment_cache.get(record.key) is not None for record in fragment_records):
            raise RuntimeError("Cold fragment-cache lookup unexpectedly hit")
        return _fragment_sweep(fragment_records)

    operations["fragment_cache_miss_sweep"], _ = _measure(
        fragment_misses,
        samples=samples,
        p95_limit_ms=5_000,
        response_limit_bytes=32_768,
        summary=_fragment_summary,
        response_size=lambda item: _compact_size(_fragment_summary(item)),
    )

    def fragment_puts() -> _FragmentSweep:
        published: list[FragmentRecord] = []
        for record in fragment_records:
            stored = fragment_cache.put(record.key, record.content)
            if stored != record:
                raise RuntimeError("Fragment cache did not publish an exact record")
            assert stored is not None
            published.append(stored)
        return _fragment_sweep(published)

    operations["fragment_cache_put_sweep"], _ = _measure(
        fragment_puts,
        samples=1,
        p95_limit_ms=10_000,
        response_limit_bytes=32_768,
        summary=_fragment_summary,
        response_size=lambda item: _compact_size(_fragment_summary(item)),
    )

    def fragment_hits() -> _FragmentSweep:
        loaded: list[FragmentRecord] = []
        for expected in fragment_records:
            record = fragment_cache.get(expected.key)
            if record != expected:
                raise RuntimeError("Fragment cache hit was not byte-exact")
            assert record is not None
            loaded.append(record)
        return _fragment_sweep(loaded)

    operations["fragment_cache_hit_sweep"], _ = _measure(
        fragment_hits,
        samples=samples,
        p95_limit_ms=5_000,
        response_limit_bytes=32_768,
        summary=_fragment_summary,
        response_size=lambda item: _compact_size(_fragment_summary(item)),
    )

    render_config = snapshot.descriptor.render
    assert render_config is not None
    incremental_package = build_manual_projection_package(
        manual_plan,
        render_config.views[0].template_path.read_bytes(),
        renderer_id=ManualHtmlRenderer.renderer_id,
        renderer_version=manual_renderer_version,
        max_output_bytes=snapshot.descriptor.limits.max_render_bytes,
        fragment_records=[record.as_dict() for record in fragment_records],
    )

    def fragment_equivalence() -> _ProjectionSample:
        sample = _projection_sample(
            manual_plan,
            incremental_package,
            ManualHtmlRenderer(manual_renderer_version).render(incremental_package),
        )
        if sample.artifact != manual_result.artifact:
            raise RuntimeError("Fragment-assisted manual output is not byte-equivalent")
        return sample

    operations["fragment_assisted_equivalence"], fragment_result = _projection_measurement(
        fragment_equivalence,
        samples=samples,
        p95_limit_ms=15_000,
    )

    render_config = snapshot.descriptor.render
    assert render_config is not None
    manual_view = render_config.views[0]
    template_bytes = manual_view.template_path.read_bytes()
    production_renderer = GenericHtmlRenderer()
    operations["manual_incremental_cold"], cold_prepared = _measure(
        lambda: production_renderer.prepare(
            snapshot,
            manual_view,
            template_bytes,
            changeset_hash=None,
        ),
        samples=1,
        p95_limit_ms=20_000,
        response_limit_bytes=MAX_RECEIPT_BYTES,
        summary=_prepared_summary,
        response_size=_prepared_response_size,
    )
    operations["manual_incremental_warm"], warm_prepared = _measure(
        lambda: production_renderer.prepare(
            snapshot,
            manual_view,
            template_bytes,
            changeset_hash=None,
        ),
        samples=samples,
        p95_limit_ms=20_000,
        response_limit_bytes=MAX_RECEIPT_BYTES,
        summary=_prepared_summary,
        response_size=_prepared_response_size,
    )
    operations["manual_forced_full"], forced_prepared = _measure(
        lambda: GenericHtmlRenderer(incremental=False).prepare(
            snapshot,
            manual_view,
            template_bytes,
            changeset_hash=None,
        ),
        samples=samples,
        p95_limit_ms=20_000,
        response_limit_bytes=MAX_RECEIPT_BYTES,
        summary=_prepared_summary,
        response_size=_prepared_response_size,
    )
    production_values = tuple(
        cast(PreparedRender, value) for value in (cold_prepared, warm_prepared, forced_prepared)
    )
    if len({value.output for value in production_values}) != 1:
        raise RuntimeError("Production cold, warm, and forced-full manual output differs")

    changed_node = replace(
        snapshot.nodes[0],
        content=snapshot.nodes[0].content + "\nChanged projection content.\n",
        content_hash=_sha256(
            (snapshot.nodes[0].content + "\nChanged projection content.\n").encode()
        ),
    )
    added_node = replace(
        snapshot.nodes[-1],
        node_id="guide.synthetic-added",
        title="Synthetic added",
        source_path="docs/content/synthetic-added.md",
        content_hash=_sha256(b"synthetic-added"),
    )
    variants = {
        "change": replace(
            snapshot,
            nodes=(changed_node, *snapshot.nodes[1:]),
            source_hash=_sha256(b"manual-change-variant"),
        ),
        "add": replace(
            snapshot,
            nodes=(*snapshot.nodes, added_node),
            source_hash=_sha256(b"manual-add-variant"),
        ),
        "delete": replace(
            snapshot,
            nodes=snapshot.nodes[:-1],
            edges=tuple(
                edge
                for edge in snapshot.edges
                if edge.source_id != snapshot.nodes[-1].node_id
                and edge.target_id != snapshot.nodes[-1].node_id
            ),
            source_hash=_sha256(b"manual-delete-variant"),
        ),
        "reorder": replace(
            snapshot,
            nodes=tuple(reversed(snapshot.nodes)),
            source_hash=_sha256(b"manual-reorder-variant"),
        ),
    }
    variant_equivalence: dict[str, bool] = {}
    for name, variant in variants.items():
        incremental = GenericHtmlRenderer().prepare(
            variant,
            manual_view,
            template_bytes,
            changeset_hash=None,
        )
        full = GenericHtmlRenderer(incremental=False).prepare(
            variant,
            manual_view,
            template_bytes,
            changeset_hash=None,
        )
        variant_equivalence[name] = incremental.output == full.output
    if not all(variant_equivalence.values()):
        raise RuntimeError("Production incremental mutation output differs from forced full")

    manual_service = RenderService(project)
    graph_service = GraphRenderService(project)
    manual_service.render("manual")
    graph_service.render("architecture")
    operations["manual_status_no_work"], manual_status = _measure(
        lambda: _profiled_status(lambda: manual_service.status("manual")),
        samples=samples,
        p95_limit_ms=500,
        response_limit_bytes=MAX_STATUS_RESPONSE_BYTES,
        summary=_status_summary,
        response_size=lambda item: _compact_size(cast(_StatusSample, item).response),
    )
    operations["portable_graph_status_no_work"], graph_status = _measure(
        lambda: _profiled_status(lambda: graph_service.status("architecture")),
        samples=samples,
        p95_limit_ms=500,
        response_limit_bytes=MAX_STATUS_RESPONSE_BYTES,
        summary=_status_summary,
        response_size=lambda item: _compact_size(cast(_StatusSample, item).response),
    )

    return {
        "fixture": {
            "kind": "synthetic_generic_projection",
            "node_count": node_count,
            "edge_count": node_count - 1,
            "manual_page_count": len(manual_pages),
            "portable_graph_node_count": graph_diagnostics["returned_nodes"],
            "portable_graph_edge_count": graph_diagnostics["returned_edges"],
            "full_coverage": (
                len(manual_pages) == node_count
                and graph_diagnostics["returned_nodes"] == node_count
                and graph_diagnostics["returned_edges"] == node_count - 1
            ),
        },
        "operations": operations,
        "equivalence": {
            "manual_in_process_vs_detached": manual_worker.artifact == manual_result.artifact,
            "portable_graph_in_process_vs_detached": graph_worker.artifact == graph_result.artifact,
            "manual_full_vs_fragment_assisted": fragment_result.artifact == manual_result.artifact,
            "manual_production_cold_warm_full": len({value.output for value in production_values})
            == 1,
            "manual_production_variants": variant_equivalence,
        },
        "sizes": {
            "manual": {
                **_projection_summary(manual_result),
                "receipt_bytes": _compact_size(manual_result.receipt),
            },
            "portable_graph": {
                **_projection_summary(graph_result),
                "receipt_bytes": _compact_size(graph_result.receipt),
            },
            "fragment_assisted_manual": {
                **_projection_summary(fragment_result),
                "receipt_bytes": _compact_size(fragment_result.receipt),
                "fragment_count": len(fragment_records),
                "aggregate_fragment_content_bytes": sum(
                    record.byte_count for record in fragment_records
                ),
            },
            "manual_status_response_bytes": _compact_size(
                cast(_StatusSample, manual_status).response
            ),
            "portable_graph_status_response_bytes": _compact_size(
                cast(_StatusSample, graph_status).response
            ),
        },
        "memory": {
            "process_peak_rss_kib": int(resource.getrusage(resource.RUSAGE_SELF).ru_maxrss),
            "manual_worker_peak_bytes": manual_worker_peak,
            "portable_graph_worker_peak_bytes": graph_worker_peak,
        },
    }


def main() -> int:
    arguments = _parser().parse_args()
    default_nodes = FULL_NODE_COUNT if arguments.mode == "full" else SMOKE_NODE_COUNT
    default_samples = DEFAULT_FULL_SAMPLES if arguments.mode == "full" else DEFAULT_SMOKE_SAMPLES
    node_count = default_nodes if arguments.nodes is None else arguments.nodes
    samples = default_samples if arguments.samples is None else arguments.samples
    if not 2 <= node_count <= FULL_NODE_COUNT:
        raise SystemExit("--nodes must be between 2 and 1000")
    if arguments.mode == "full" and node_count != FULL_NODE_COUNT:
        raise SystemExit("--mode full requires exactly 1000 nodes")
    if samples < 1:
        raise SystemExit("--samples must be positive")

    with tempfile.TemporaryDirectory(prefix="docforge-milestone3-") as directory:
        benchmark_root = Path(directory).resolve()
        _prepare_fixture(benchmark_root, node_count)
        measurement = _benchmark(benchmark_root, node_count, samples)

    status = _git(["status", "--porcelain"])
    report: dict[str, object] = {
        "schema_version": 1,
        "benchmark": "docforge2_milestone3",
        "mode": arguments.mode,
        "source": {
            "revision": _git(["rev-parse", "HEAD"]),
            "dirty": bool(status),
        },
        "environment": {
            "platform": platform.platform(),
            "machine": platform.machine(),
            "python": platform.python_version(),
            "implementation": platform.python_implementation(),
        },
        "method": {
            "clock": "time.perf_counter_ns",
            "in_process_peak_memory": "tracemalloc per measured invocation",
            "detached_peak_memory": "worker receipt resource peak RSS",
            "process_peak_memory": "resource.getrusage(RUSAGE_SELF).ru_maxrss",
            "response_size": "UTF-8 bytes of canonical compact sorted JSON",
            "samples": samples,
            "full_mode_node_requirement": FULL_NODE_COUNT,
            "determinism": (
                "stable semantic summaries must match across samples; report JSON uses sorted keys"
            ),
            "maximum_worker_artifact_bytes": MAX_WORKER_ARTIFACT_BYTES,
        },
        **measurement,
    }
    encoded = encode_report(report)
    if arguments.output is not None:
        output = arguments.output.resolve()
        output.parent.mkdir(parents=True, exist_ok=True)
        output.write_text(encoded, encoding="utf-8")
    sys.stdout.write(encoded)
    return 0


if __name__ == "__main__":
    raise SystemExit(main())