Replace ReInterHand shared-v1.0 with the verified shared-v1.1 release (candidate PR) (#8)
8f94cdb 29 days ago
30.3 kB
#!/usr/bin/env python3
"""Acceptance checks for a shared-v1.0 -> shared-v1.1 migration (CONTRACT 6).
Compares the migrated tree against the source it came from, so every claim is
checked against real bytes rather than re-asserted from the migrator:
1. every component/attribute and its dictionary keys resolve; no stray files, no
   missing keys, no unregistered fields; merged numeric components leave no
   standalone ``*_valid.pt``; media/JSONL validity files survive untouched
2. every numeric field is identical to its source in dtype, shape, element-wise
   values, NaN positions and bool validity, compared as a unified contiguous byte
   view with no tolerance
3. ``num_frames``, sampling, hand slots, split, camera, view type, provenance and
   ``valid_policy`` are unchanged -- the metainfo diff must equal the allowed set
4. media and other preserved files match their source digests; the release
   manifest is regenerated because merged ``.pt`` digests necessarily changed
5. the new reader resolves RGB, camera and every annotation present on
   representative complete sequences
The report distinguishes "repackaging is numerically identical" from "human
visual acceptance"; this tool produces evidence for the former only.
"""

from __future__ import annotations

import argparse
import json
import sys
from pathlib import Path

import torch

sys.path.insert(0, str(Path(__file__).resolve().parent))

from warehouse import (  # noqa: E402
    BundleReader,
    Registry,
    same_tensor,
    sha256_file,
)

#: this package lives in <dataset>/shared-v1.1/script/reinterhand_v1_1, so the
#: release and the dataset root beside it need no arguments.
RELEASE_ROOT = Path(__file__).resolve().parent.parent.parent
DATASET_ROOT = RELEASE_ROOT.parent
#: Where ``--manifest`` writes.  The release manifest is a publication artifact and
#: lives *outside* the release tree: a manifest inside it could not hash its own final
#: bytes.  ``reinterhand_common.MANIFEST_PATH`` holds the same path -- keep the two in
#: step; validate.py is what the documented chain rebuilds the shipping manifest with.
MANIFEST_PATH = DATASET_ROOT / "tmp" / "reinterhand-tools" / "PUBLISH_MANIFEST.json"
#: The closing chain's CONTRACT 6 evidence: ``verify.py --complete`` writes it here (step
#: 2) and ``FULL_MIGRATION_SUMMARY.json`` cites it for the whole release
#: (``all_numeric_bytes_exact_evidence``), so it sits beside the manifest, outside the
#: release tree, and only the acceptance pass may write it (see the refusal in ``main``).
ACCEPTANCE_REPORT = MANIFEST_PATH.parent / "acceptance.json"

V10_SCHEMA_VERSION = "hand-warehouse/1.0"
EXPECTED_VIEWS = 6750
#: metainfo keys the migration is allowed to rewrite at the top level.  Everything
#: else is a claim the source made and must survive verbatim -- ReInterHand's release
#: status flags (``upload_authorized``/``full_output_human_review``/
#: ``full_conversion_complete``/``conversion_scope``), the omitted-component reasons,
#: the split bookkeeping.  The version keys change by definition, and ``components``
#: changes wherever a numeric field joins a bundle.
ALLOWED_TOP_LEVEL_CHANGES = {"schema_version", "release_id", "components"}
#: the one key that is neither "bumped by definition" nor "preserved verbatim": v1.0's
#: converter stamped ``lossless-video-2026-09-18``'s slot with a label of its own
#: (``lossless-video-2026-09-25``), which ``doc/schema.json`` declares nowhere, so
#: migration points it at the revision the registry itself declares -- the value every
#: published v1.1 sibling view carries.  It is handled on its own rather than by adding
#: it to the set above, because the rewrite is allowed to one target value only, and a
#: view that never declared the key must not acquire one.
STORAGE_REVISION = "storage_revision"
#: component fields the migration adds when a numeric field joins a bundle
BUNDLE_FIELDS = {"data_key", "valid_key", "storage", "valid_storage"}
#: component fields that must survive any migration untouched
PRESERVED_FIELDS = ("provenance", "source_locator", "valid_policy", "axes", "valid_axes")


class Failure(Exception):
    pass


def check(condition, message):
    if not condition:
        raise Failure(message)


def flatten(value, prefix=""):
    """Flatten a JSON object into dotted leaf paths."""
    if isinstance(value, dict):
        out = {}
        for key, item in value.items():
            out.update(flatten(item, f"{prefix}{key}."))
        return out
    if isinstance(value, list) and value and not isinstance(value[0], (dict, list)):
        return {prefix[:-1]: json.dumps(value)}
    return {prefix[:-1]: json.dumps(value)}


def is_stray_validity(registry: Registry, rel: str) -> bool:
    """True when a standalone ``*_valid.pt`` belongs to a field that was merged.
    Catches the failure CONTRACT 6.1 names explicitly: a numeric component's
    validity left behind as its own file because its name was not recognised.
    """
    if not rel.endswith("_valid.pt"):
        return False
    return registry.resolve(rel[: -len("_valid.pt")] + ".pt") is not None


def compare_metainfo(source: dict, target: dict, sequence: str, camera: str, registry: Registry):
    """Assert the metainfo diff is exactly the storage-layout change set."""
    where = f"{sequence}/{camera}"
    check(source["schema_version"] == V10_SCHEMA_VERSION,
          f"{where}: source is not {V10_SCHEMA_VERSION}")
    for key in ("num_frames", "sampling", "hand_slots", "split", "camera",
                "view_type", "dataset_id", "sequence_id", "camera_id", "sample_type"):
        check(source.get(key) == target.get(key),
              f"{where}: {key} changed ({source.get(key)!r} -> {target.get(key)!r})")
    check(target["schema_version"] == "hand-warehouse/1.1", f"{where}: schema_version not bumped")
    check(target["release_id"] == "shared-v1.1", f"{where}: release_id not bumped")

    #: every remaining top level key, compared to this view's own source rather than
    #: to a constant: these are per-view claims made by the source, so "unchanged"
    #: has to mean unchanged against *this view's* v1.0 bytes, not equal to one
    #: expected value.  ReInterHand's status flags happen to be uniform today (all
    #: 6,750 views carry ``upload_authorized=false``, ``full_output_human_review=
    #: "pending"``, ``full_conversion_complete=true``, ``conversion_scope="full"``);
    #: comparing per-view is what keeps a future re-render with a different status
    #: from being silently normalised into agreement.
    for key in sorted(set(source) | set(target)):
        if key in ALLOWED_TOP_LEVEL_CHANGES:
            continue
        check(key in source and key in target,
              f"{where}: top level key {key!r} exists on only one side")
        if key == STORAGE_REVISION:
            #: one target value only, read off the registry rather than restated here.
            check(target[key] == registry.schema[STORAGE_REVISION],
                  f"{where}: {key} must be the registry's own "
                  f"({target[key]!r} != {registry.schema[STORAGE_REVISION]!r})")
            continue
        check(source[key] == target[key],
              f"{where}: {key} changed ({source[key]!r} -> {target[key]!r})")

    source_components = source["components"]
    target_components = target["components"]
    check(set(source_components) == set(target_components),
          f"{where}: component set changed")

    for name, before in source_components.items():
        after = target_components[name]
        for key in PRESERVED_FIELDS:
            check(before.get(key) == after.get(key),
                  f"{where}/{name}: {key} changed")
        check(before.get("schema_id") == after.get("schema_id"),
              f"{where}/{name}: schema_id changed")
        check(before.get("available") == after.get("available"),
              f"{where}/{name}: availability changed")
        # attributes (e.g. visibility) follow the same rules as components
        before_attributes = before.get("attributes") or {}
        after_attributes = after.get("attributes") or {}
        check(set(before_attributes) == set(after_attributes),
              f"{where}/{name}: attribute set changed")
        for attribute, before_attribute in before_attributes.items():
            after_attribute = after_attributes[attribute]
            for key in ("axes", "dtype", "visibility_type", "meaning", "provenance",
                        "source_locator"):
                check(before_attribute.get(key) == after_attribute.get(key),
                      f"{where}/{name}.{attribute}: {key} changed")
            if "data_key" in after_attribute:
                check(before_attribute.get("valid") is True,
                      f"{where}/{name}.{attribute}: bundled without a declared all-known scalar")
                check("valid" not in after_attribute,
                      f"{where}/{name}.{attribute}: scalar valid must give way to valid_key")
                check(after_attribute["data_ref"] == after_attribute["valid_ref"],
                      f"{where}/{name}.{attribute}: data_ref and valid_ref must share a file")


def verify_distortion(registry, v10_view, v11_view, reader, name, source_component,
                      target_component, where, consumed):
    """CONTRACT 5/3: the separate distortion vector joins the camera's own bundle.
    v1.0 declares it on the component that owns the calibration, with no validity
    of its own: the camera model pairs the distortion vector with the intrinsic
    validity, so the two fields must carry the same bool values.  Returns the
    field record, or ``None`` when the component declares no distortion.
    """
    ref = source_component.get("distortion_ref")
    if not ref:
        return None
    resolved = registry.resolve(ref)
    if resolved is None:
        check(target_component.get("distortion_ref") == ref,
              f"{where}/{name}: distortion_ref changed for a preserved file")
        check((v10_view / ref).read_bytes() == (v11_view / ref).read_bytes(),
              f"{where}/{name}: preserved distortion {ref} differs")
        consumed.add(ref)
        return None
    bundle_rel, keys, _ = resolved
    check(target_component.get("distortion_ref") == bundle_rel
          and target_component.get("distortion_valid_ref") == bundle_rel,
          f"{where}/{name}: distortion refs must point at the camera bundle")
    check(target_component.get("distortion_key") == keys + ["data"],
          f"{where}/{name}: distortion_key mismatch")
    check(target_component.get("distortion_valid_key") == keys + ["valid"],
          f"{where}/{name}: distortion_valid_key mismatch")

    data = reader.get(v11_view, bundle_rel, keys + ["data"])
    valid = reader.get(v11_view, bundle_rel, keys + ["valid"])
    check(valid.dtype == torch.bool, f"{where}/{name}: distortion valid is not bool")
    check(same_tensor(torch.load(v10_view / ref, map_location="cpu", weights_only=True), data),
          f"{where}/{name}: distortion is not byte-identical")

    valid_ref = source_component.get("distortion_valid_ref")
    if valid_ref:
        check(same_tensor(torch.load(v10_view / valid_ref, map_location="cpu",
                                     weights_only=True), valid),
              f"{where}/{name}: distortion valid is not byte-identical")
        consumed.add(valid_ref)
        origin = f"source tensor {valid_ref}"
    else:
        # CONTRACT 3: with no validity of its own, the distortion vector either
        # shares the owning component's validity or expands a declared all-known
        # scalar.  Both are checked against the source rather than re-asserted.
        owner_valid_ref = source_component.get("valid_ref")
        paired = (torch.load(v10_view / owner_valid_ref, map_location="cpu", weights_only=True)
                  if owner_valid_ref else None)
        shared = (paired is not None and list(paired.shape) == list(valid.shape)
                  and same_tensor(paired, valid))
        check(shared or bool(valid.all()),
              f"{where}/{name}: distortion validity matches neither the paired camera "
              f"validity nor a declared all-known scalar")
        origin = ("shared with intrinsic validity (CONTRACT 3)" if shared
                  else "expanded declared all-known scalar (CONTRACT 3)")
    consumed.add(ref)
    return {
        "owner": f"{name}.distortion", "entry": f"{bundle_rel}:{'/'.join(keys)}",
        "dtype": str(data.dtype), "shape": list(data.shape),
        "valid_shape": list(valid.shape),
        "valid_true": int(valid.sum()), "valid_false": int((~valid).sum()),
        "nan_count": int(torch.isnan(data).sum()) if data.is_floating_point() else None,
        "valid_origin": origin, "byte_exact": True,
    }


def same_bytes(left: Path, right: Path) -> tuple:
    """``(same, how)`` for two paths: the same file, or the same bytes.
    Migration hardlinks media, so in the common case these are one inode and the
    answer is structural.  Re-reading a file to compare it with itself proves
    nothing and costs the run, and on a published tree that read is most of the
    acceptance time.  ReInterHand's *local* tree has no media at all (the ``*.mkv``
    are deliberately outside the migration scope), so here the identity shortcut
    almost never fires and nearly every shared file is settled by content -- cheap,
    because the whole tree is ~2.7 GiB.  When the inodes differ -- a tree that was
    *copied* rather than linked -- the bytes are hashed in streams, so nothing is
    assumed equal merely because the migration usually links.
    ``how`` travels into the report, so a reader can see which files were settled by
    identity and which by content instead of having to trust that either ran.
    """
    left_stat, right_stat = left.stat(), right.stat()
    if (left_stat.st_dev, left_stat.st_ino) == (right_stat.st_dev, right_stat.st_ino):
        return True, "same inode"
    if left_stat.st_size != right_stat.st_size:
        return False, "size differs"
    return sha256_file(left) == sha256_file(right), "sha256"


def same_preserved(left: Path, right: Path) -> tuple:
    """``(same, how)`` for a preserved file that may be *declared absent*.
    ReInterHand's local tree carries no ``*.mkv``: the media are 1.08 TB and were
    deliberately left out of the migration scope, to be copied server-side at
    publication time.  A media ``data_ref`` is therefore missing from **both**
    trees, and asking whether the two copies match has no answer -- ``same_bytes``
    raises ``FileNotFoundError`` on the first one.
    Absent on both sides is not a difference, it is the declared state: the claim
    being verified is that the migration left the media exactly as absent as it
    found them.  Absent on *one* side only is a real failure and is reported as
    such, because that would mean the migration dropped a file it should have
    preserved (or invented one).
    """
    if not left.exists() and not right.exists():
        return True, "declared absent (absent from both trees)"
    if left.exists() != right.exists():
        return False, "present on only one side"
    return same_bytes(left, right)


def verify_view(registry, v10_view, v11_view, reader):
    """Run CONTRACT 6 checks for one view; return its evidence record."""
    where = f"{v10_view.parent.name}/{v10_view.name}"
    source = json.loads((v10_view / "metainfo.json").read_text(encoding="utf-8"))
    target = json.loads((v11_view / "metainfo.json").read_text(encoding="utf-8"))
    compare_metainfo(source, target, v10_view.parent.name, v10_view.name, registry)

    consumed, fields, media = set(), [], []
    for name, component in source["components"].items():
        if not component.get("available", True):
            continue
        for owner, spec, kind in [(name, component, "component")] + [
            (f"{name}.{attribute}", attribute_spec, "attribute")
            for attribute, attribute_spec in (component.get("attributes") or {}).items()
        ]:
            ref = spec.get("data_ref")
            if not ref:
                continue
            resolved = registry.resolve(ref)
            after = (target["components"][name] if kind == "component"
                     else target["components"][name]["attributes"][owner.split(".", 1)[1]])
            if resolved is None:
                # preserved verbatim: same reference, same bytes
                check(after.get("data_ref") == ref, f"{where}/{owner}: data_ref changed for a media file")
                check("data_key" not in after, f"{where}/{owner}: media must not carry data_key")
                same, how = same_preserved(v10_view / ref, v11_view / ref)
                check(same, f"{where}/{owner}: preserved file {ref} differs")
                valid_ref = spec.get("valid_ref")
                valid_how = None
                if valid_ref:
                    check(after.get("valid_ref") == valid_ref,
                          f"{where}/{owner}: media valid_ref changed")
                    check("valid_key" not in after, f"{where}/{owner}: media must not carry valid_key")
                    same, valid_how = same_preserved(v10_view / valid_ref, v11_view / valid_ref)
                    check(same, f"{where}/{owner}: preserved validity {valid_ref} differs")
                media.append({"owner": owner, "data_ref": ref, "valid_ref": valid_ref,
                              "verified": how, "valid_verified": valid_how})
                continue

            bundle_rel, keys, entry = resolved
            check(after["data_ref"] == bundle_rel and after["valid_ref"] == bundle_rel,
                  f"{where}/{owner}: bundled refs must be the same file")
            check(after["data_key"] == keys + ["data"], f"{where}/{owner}: data_key mismatch")
            check(after["valid_key"] == keys + ["valid"], f"{where}/{owner}: valid_key mismatch")
            for key in ("storage", "valid_storage"):
                check(after[key]["profile"] == "pt/tensor_bundle/v1",
                      f"{where}/{owner}: {key} profile must be pt/tensor_bundle/v1")

            data, valid = reader.field(v11_view, after)
            check(valid.dtype == torch.bool, f"{where}/{owner}: valid is not bool")
            if kind == "attribute":
                check(after.get("valid_axes") == list(after["axes"])[:valid.ndim],
                      f"{where}/{owner}: valid_axes must be axes[:valid.ndim]")
                check(after.get("valid_ref") == after.get("data_ref"),
                      f"{where}/{owner}: visibility must share one bundle file")
            source_data = torch.load(v10_view / ref, map_location="cpu", weights_only=True)
            check(same_tensor(source_data, data), f"{where}/{owner}: data is not byte-identical")
            source_valid_ref = spec.get("valid_ref")
            if source_valid_ref:
                source_valid = torch.load(v10_view / source_valid_ref,
                                          map_location="cpu", weights_only=True)
                check(same_tensor(source_valid, valid), f"{where}/{owner}: valid is not byte-identical")
                origin = f"source tensor {source_valid_ref}"
                consumed.add(source_valid_ref)
            else:
                check(spec.get("valid") is True,
                      f"{where}/{owner}: no valid_ref and no declared all-known scalar")
                check(bool(valid.all()),
                      f"{where}/{owner}: expanded all-known validity must be all true")
                origin = "expanded declared all-known scalar"
            consumed.add(ref)
            fields.append({
                "owner": owner, "entry": f"{bundle_rel}:{'/'.join(keys)}",
                "dtype": str(data.dtype), "shape": list(data.shape),
                "valid_shape": list(valid.shape),
                "valid_true": int(valid.sum()), "valid_false": int((~valid).sum()),
                "nan_count": int(torch.isnan(data).sum()) if data.is_floating_point() else None,
                "valid_origin": origin, "byte_exact": True,
            })
        distortion = verify_distortion(registry, v10_view, v11_view, reader, name,
                                       component, target["components"][name], where, consumed)
        if distortion is not None:
            fields.append(distortion)

    # 1: no stray files, nothing lost
    source_files = {p.relative_to(v10_view).as_posix()
                    for p in v10_view.rglob("*") if p.is_file()}
    target_files = {p.relative_to(v11_view).as_posix()
                    for p in v11_view.rglob("*") if p.is_file()}
    expected = (source_files - consumed - {"metainfo.json"}) | {"metainfo.json"} | {
        f["entry"].split(":")[0] for f in fields}
    check(target_files == expected,
          f"{where}: file set mismatch; unexpected={sorted(target_files - expected)}, "
          f"missing={sorted(expected - target_files)}")
    strays = sorted(f for f in target_files if is_stray_validity(registry, f))
    check(not strays, f"{where}: merged numeric fields left standalone validity files: {strays}")

    return {
        "sequence_id": v10_view.parent.name, "camera_id": v10_view.name,
        "num_frames": source["num_frames"],
        "source_files": len(source_files), "target_files": len(target_files),
        "merged_fields": len(fields), "preserved_files": len(media),
        "media_verified_by": {how: sum(1 for m in media if m["verified"] == how)
                              for how in sorted({m["verified"] for m in media})},
        "fields": fields, "media": media,
    }


def main(argv=None):
    parser = argparse.ArgumentParser(description=__doc__.splitlines()[0])
    parser.add_argument("--v10", type=Path, default=DATASET_ROOT / "shared-v1.0",
                        help="shared-v1.0 release directory")
    parser.add_argument("--v11", type=Path, default=RELEASE_ROOT,
                        help="shared-v1.1 release directory")
    parser.add_argument("--schema", type=Path, default=RELEASE_ROOT / "doc" / "schema.json",
                        help="v1.1 schema.json")
    parser.add_argument("--manifest", type=Path, default=MANIFEST_PATH,
                        help=f"write the v1.1 release manifest here (default: "
                             f"{MANIFEST_PATH}, outside the release tree)")
    parser.add_argument("--report", type=Path, help="write the acceptance report here")
    parser.add_argument("--sample", type=int, default=None,
                        help="verify at most N views (default: all migrated)")
    parser.add_argument("--views", nargs="*", default=None,
                        help="restrict to these sequence ids")
    parser.add_argument("--complete", action="store_true",
                        help="fail when a source view has no migrated counterpart")
    parser.add_argument("--no-manifest", action="store_true",
                        help="skip regenerating the release manifest (CONTRACT 6.4). It "
                             "hashes every file of the tree regardless of --views, so an "
                             "incremental check while the pull is still running pays "
                             "minutes for a manifest the final run overwrites anyway. "
                             "The accepting run passes this as well: the manifest has to "
                             "be the last thing built, because it must cover this "
                             "release's own FULL_MIGRATION_SUMMARY.json and its rebuilt "
                             "doc/*.md, which are written after the acceptance -- "
                             "validate.py regenerates it at the end, so a manifest built "
                             "here would be overwritten unread. CONTRACT 6.4 asks that the "
                             "manifest be regenerated and correct, not that this tool be "
                             "the one to write it.")
    args = parser.parse_args(argv)

    if args.sample is not None and args.sample < 1:
        parser.error("--sample must be positive")
    if args.views is not None and not args.views:
        parser.error("--views requires at least one sequence")
    acceptance_pass = args.complete and args.views is None and args.sample is None
    if args.report is not None:
        if args.report.resolve() in {MANIFEST_PATH.resolve(), args.manifest.resolve()}:
            parser.error("--report must not overwrite a publication manifest")
        if not acceptance_pass and args.report.resolve() == ACCEPTANCE_REPORT.resolve():
            parser.error("the acceptance report requires --complete without --views/--sample")

    registry = Registry.load(args.schema)
    reader = BundleReader()
    source_data, target_data = args.v10 / "data", args.v11 / "data"
    if not source_data.is_dir() or not target_data.is_dir():
        parser.error("both source and target data directories must exist")
    source_views = sorted(p.parent for p in source_data.glob("*/*/metainfo.json"))
    if not source_views:
        parser.error("source contains no views; no acceptance evidence was written")
    if acceptance_pass:
        source_set = {p.relative_to(source_data) for p in source_views}
        target_set = {p.parent.relative_to(target_data)
                      for p in target_data.glob("*/*/metainfo.json")}
        if len(source_set) != EXPECTED_VIEWS or source_set != target_set:
            parser.error(f"full acceptance requires {EXPECTED_VIEWS} source views and "
                         "the exact same target view set")
    migrated, absent = [], []
    for view in source_views:
        relative = view.relative_to(args.v10 / "data")
        (migrated if (target_data / relative / "metainfo.json").is_file() else absent).append(view)
    if args.views:
        wanted = set(args.views)
        known = {view.parent.name for view in source_views}
        if not wanted <= known:
            parser.error(f"no such source sequence: {sorted(wanted - known)}")
        migrated = [v for v in migrated if v.parent.name in wanted]
        absent = [v for v in absent if v.parent.name in wanted]
    if args.sample:
        migrated = migrated[:args.sample]

    report = {"source": str(args.v10.resolve()), "target": str(args.v11.resolve()),
              "schema_sha256": sha256_file(args.schema), "views": [], "failures": [],
              "views_in_source": len(source_views), "views_migrated": len(migrated),
              "views_absent_from_target": [v.relative_to(args.v10 / "data").as_posix()
                                           for v in absent]}
    if args.complete:
        for view in absent:
            rel = view.relative_to(args.v10 / "data").as_posix()
            report["failures"].append({"view": rel, "error": "view was not migrated"})
            print(f"FAIL {rel}: view was not migrated", file=sys.stderr)
    for view in migrated:
        relative = view.relative_to(args.v10 / "data")
        target_view = target_data / relative
        try:
            report["views"].append(verify_view(registry, view, target_view, reader))
        except Failure as error:
            report["failures"].append({"view": relative.as_posix(), "error": str(error)})
            print(f"FAIL {relative}: {error}", file=sys.stderr)

    # 4: regenerated release manifest
    if args.no_manifest or not acceptance_pass:
        #: reported as skipped, not as "0 files" -- an empty manifest in the report
        #: would read like a tree with nothing in it
        report["manifest"] = {"skipped": "incremental check; not an acceptance run"}
    else:
        manifest = {}
        #: The manifest normally lives outside the release (see MANIFEST_PATH), so it
        #: is not part of the tree being hashed and there is nothing to exclude.  Only
        #: a ``--manifest`` pointed *into* the release has to be skipped: it cannot hash
        #: itself without recording the digest it is about to have.
        try:
            manifest_rel = args.manifest.relative_to(args.v11).as_posix()
        except ValueError:
            manifest_rel = None
        for path in sorted(args.v11.rglob("*")):
            rel = path.relative_to(args.v11).as_posix()
            if "__pycache__" in path.parts or rel == manifest_rel:
                continue
            if path.is_file():
                manifest[rel] = {"size": path.stat().st_size, "sha256": sha256_file(path)}
        report["manifest"] = {"files": len(manifest),
                              "bytes": sum(v["size"] for v in manifest.values()),
                              "excludes": [x for x in (manifest_rel, "__pycache__/") if x]}
        if args.manifest:
            args.manifest.parent.mkdir(parents=True, exist_ok=True)
            args.manifest.write_text(
                json.dumps({"release": "shared-v1.1",
                            "location": str(args.manifest),
                            "in_release_tree": manifest_rel is not None,
                            "files": manifest},
                           ensure_ascii=False, indent=2) + "\n", encoding="utf-8")

    report["passed"] = not report["failures"]
    report["scope"] = ("repackaging numerically identical; this is not human visual acceptance")
    if args.report:
        args.report.parent.mkdir(parents=True, exist_ok=True)
        args.report.write_text(json.dumps(report, ensure_ascii=False, indent=2) + "\n",
                               encoding="utf-8")

    total_fields = sum(v["merged_fields"] for v in report["views"])
    media_modes = {}
    for view in report["views"]:
        for mode, count in view["media_verified_by"].items():
            media_modes[mode] = media_modes.get(mode, 0) + count
    print(f"views={len(report['views'])} merged_fields={total_fields} "
          f"failures={len(report['failures'])} "
          + (f"manifest_files={report['manifest']['files']}"
             if "files" in report["manifest"] else "manifest=skipped")
          + f" media_verified_by={media_modes}")
    print("ACCEPTANCE_PASS" if report["passed"] else "ACCEPTANCE_FAIL")
    return 0 if report["passed"] else 1


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