diff --git a/README.md b/README.md
index 5abe4d0..0d25d10 100644
--- a/README.md
+++ b/README.md
@@ -228,6 +228,30 @@ work_item/scripts/python -m pytest -m phase_f -q
 
 Phases A–E remain green in the full run above.
 
+## Media and metadata hardening (US07-03)
+
+Every pixel the application reads goes through `photo_pipeline/imaging.py`: the
+declared dimensions are checked before anything is decoded, Pillow's
+decompression-bomb warning is treated as a refusal, JPEG decodes near the requested
+size, and each decoder failure becomes one of two typed errors. A damaged file is a
+per-item error with a persisted code, never a failed scan or a dead worker.
+
+Every metadata stage ends with an EXIF checkpoint (`services/exif_checkpoint.py`):
+snapshot, write the owned keywords, read back, prove the owned fields landed and that
+nothing else moved, refresh the file hash. A field the stage does not own that
+changed anyway makes the checkpoint `divergent` — recorded in `exif_projections`,
+shown in the review queue, never repaired behind the user's back, and not counted as
+verified, so upload stays blocked.
+
+The golden corpus that proves all of it is generated, not committed:
+`tests/fixtures/media_corpus.py` declares every format, orientation, profile,
+damage, and metadata case with its expected outcome, and the suite regenerates it
+twice to prove it does not drift.
+
+```bash
+work_item/scripts/python -m pytest tests/integration/test_media_hardening.py tests/integration/test_exif_checkpoints.py -q
+```
+
 ## Legacy CLI archive
 
 The command-line tools this application was extracted from are frozen in
diff --git a/legacy_cli_archive/donor_ledger.yaml b/legacy_cli_archive/donor_ledger.yaml
index 0044f17..b17deb9 100644
--- a/legacy_cli_archive/donor_ledger.yaml
+++ b/legacy_cli_archive/donor_ledger.yaml
@@ -480,6 +480,13 @@ rows:
       inference path needs the local model + deterministic fake.
     target: photo_pipeline/integrations/nsfw_model.py
     tests: [test_nsfwtag::test_score_images_cache_hit_skips_model]
+    delta: >
+      The donor set Pillow's process-global ImageFile.LOAD_TRUNCATED_IMAGES so a
+      partially downloaded file still scored. Here the same process also hashes
+      files and renders previews, and those must keep refusing a truncated file
+      rather than silently working on half of one; scoring opens images through
+      the bounded photo_pipeline.imaging door instead and skips the ones it cannot
+      read, leaving them unscored and visibly undecided (US07-03).
     status: characterized
 
   - id: nt-exif-keyword
diff --git a/migrations/versions/0015_exif_projections.py b/migrations/versions/0015_exif_projections.py
new file mode 100644
index 0000000..1b74aaa
--- /dev/null
+++ b/migrations/versions/0015_exif_projections.py
@@ -0,0 +1,42 @@
+"""Durable EXIF projections per asset and stage (US07-03).
+
+Revision ID: 0015_exif_projections
+Revises: 0014_restore_plans
+Create Date: 2026-08-16
+
+The concept's ``exif_projections`` table, added at the point it earns its keep: a
+checkpoint that finds a field it does not own changed must be able to say so after
+a restart. ``state`` is verified | divergent | failed, and only ``verified`` counts
+as a completed metadata stage.
+"""
+
+import sqlalchemy as sa
+from alembic import op
+
+revision = "0015_exif_projections"
+down_revision = "0014_restore_plans"
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    op.create_table(
+        "exif_projections",
+        sa.Column("asset_id", sa.String(), sa.ForeignKey("assets.id"), primary_key=True),
+        sa.Column("stage", sa.String(), primary_key=True),  # safety | analysis
+        sa.Column("id", sa.String(), nullable=False),
+        sa.Column("projection_version", sa.Integer(), nullable=False, server_default="0"),
+        # What the stage asked for: {"add": [...], "remove": [...]}.
+        sa.Column("desired_json", sa.String(), nullable=True),
+        # Fields outside the stage's ownership that did not survive the write.
+        sa.Column("divergent_fields", sa.String(), nullable=True),
+        sa.Column("result_file_sha256", sa.String(), nullable=True),
+        sa.Column("state", sa.String(), nullable=False),
+        sa.Column("error_code", sa.String(), nullable=True),
+        sa.Column("verified_at", sa.DateTime(timezone=True), nullable=True),
+        sa.Column("updated_at", sa.DateTime(timezone=True), nullable=True),
+    )
+
+
+def downgrade() -> None:
+    op.drop_table("exif_projections")
diff --git a/photo_pipeline/api/app.py b/photo_pipeline/api/app.py
index 79ac1a7..97408f0 100644
--- a/photo_pipeline/api/app.py
+++ b/photo_pipeline/api/app.py
@@ -41,6 +41,7 @@ import photo_pipeline.jobs.domain_handlers  # noqa: F401
 from photo_pipeline.config import Config
 from photo_pipeline.db import create_db_engine, create_session_factory, run_migrations
 from photo_pipeline.logging import configure_logging
+from photo_pipeline.services.thumbnails import ThumbnailService
 from photo_pipeline.services.upload_batches import UploadBatchService
 
 FRONTEND_DIR = Path(__file__).resolve().parents[2] / "frontend"
@@ -96,6 +97,9 @@ def create_app(config: Config | None = None) -> FastAPI:
         # An upload whose process died left no outcome behind; resolve it now so the
         # uploader lane is free and the uncertain batch is visible (US05-02).
         UploadBatchService(app.state.session_factory, config=config).recover()
+        # A render killed mid-write leaves its temporary beside the cache entry;
+        # remove those recognized leftovers, and only those (US07-03).
+        ThumbnailService(app.state.session_factory, config).cleanup_temp_files()
         try:
             yield
         finally:
diff --git a/photo_pipeline/imaging.py b/photo_pipeline/imaging.py
new file mode 100644
index 0000000..6db1afc
--- /dev/null
+++ b/photo_pipeline/imaging.py
@@ -0,0 +1,119 @@
+"""Bounded, defensive image decoding — one door for every pixel this app reads.
+
+A photo library contains files nobody planned for: truncated downloads, zero-byte
+placeholders, a PNG whose header claims 200000×200000, a TIFF with a broken ICC
+profile, an extension that lies about its content. None of them may take down a
+request or a worker, and none may decode more pixels than the caller allowed
+(concept §17: decoded pixels, not file size, are what exhausts memory).
+
+``open_image`` is that single door:
+
+* the declared dimensions are checked **before** a pixel is decoded;
+* Pillow's decompression-bomb *warning* is promoted to an error, because the
+  warning band (between Pillow's limit and twice it) still decodes the image;
+* every decoder failure — at open time or during the caller's decode — becomes one
+  of two typed errors, so callers map them to their own item state instead of
+  catching ``Exception``;
+* error text names no path: it reaches API responses, and the full reason goes to
+  the server log instead (US07-02).
+
+``to_srgb`` and ``draft`` are the other two bounded-decode helpers: colour-manage a
+profile-bearing image into sRGB, and let JPEG decode straight to a size near the
+requested one rather than at full resolution.
+"""
+
+from __future__ import annotations
+
+import io
+import logging
+import warnings
+from contextlib import contextmanager
+from pathlib import Path
+
+from PIL import Image, ImageCms, ImageFile, UnidentifiedImageError
+
+log = logging.getLogger(__name__)
+
+# Matches ``Config.thumbnail_max_pixels``; used where no configuration is at hand
+# (hashing runs inside discovery, which takes no config).
+DEFAULT_MAX_PIXELS = 100_000_000
+
+
+class MediaError(Exception):
+    """A file could not be turned into pixels safely."""
+
+
+class UndecodableImage(MediaError):
+    """Corrupt, truncated, empty, or not an image at all."""
+
+
+class ImageTooLarge(MediaError):
+    """More pixels than this operation is allowed to decode."""
+
+
+@contextmanager
+def open_image(path: Path | str, *, max_pixels: int = DEFAULT_MAX_PIXELS):
+    """Yield an open :class:`PIL.Image.Image`, bounded and with typed failures.
+
+    Decoder errors raised inside the ``with`` body are translated too — a truncated
+    JPEG only fails when its pixels are actually pulled, which is the caller's line,
+    not this one.
+    """
+    # Pillow's truncation tolerance is a process-global switch that any library in
+    # the process can flip (the donor CLI did). This door decides the policy for its
+    # own callers: half a file is not a picture.
+    tolerated = ImageFile.LOAD_TRUNCATED_IMAGES
+    ImageFile.LOAD_TRUNCATED_IMAGES = False
+    with warnings.catch_warnings():
+        # The warning band is not a warning for us: it means Pillow was willing to
+        # decode an image large enough to be a denial-of-service.
+        warnings.simplefilter("error", Image.DecompressionBombWarning)
+        try:
+            with Image.open(path) as image:
+                width, height = image.size
+                if width * height > max_pixels:
+                    raise ImageTooLarge(f"{width}x{height} exceeds the {max_pixels} pixel limit")
+                yield image
+        except MediaError:
+            raise
+        except (Image.DecompressionBombError, Image.DecompressionBombWarning) as error:
+            log.info("refused oversized image %s: %s", path, error)
+            raise ImageTooLarge("image exceeds the decompression-bomb limit") from error
+        except (UnidentifiedImageError, OSError, ValueError, SyntaxError, MemoryError) as error:
+            log.info("cannot decode %s: %s", path, error)
+            raise UndecodableImage(f"cannot decode image ({type(error).__name__})") from error
+        finally:
+            ImageFile.LOAD_TRUNCATED_IMAGES = tolerated
+
+
+def draft(image: Image.Image, size: int) -> None:
+    """Ask the decoder for a smaller image where the format allows it (JPEG).
+
+    This is the difference between decoding a 40-megapixel JPEG and decoding the
+    roughly 1-megapixel version a 1280px preview needs.
+    """
+    try:
+        image.draft(None, (size, size))
+    except (AttributeError, ValueError, OSError):  # not a draft-capable format
+        pass
+
+
+def to_srgb(image: Image.Image, *, mode: str) -> Image.Image:
+    """Convert into ``mode``, colour-managing through an embedded ICC profile.
+
+    Without this a wide-gamut original renders with visibly wrong colours, because
+    its numbers are interpreted as sRGB. A broken or unreadable profile is not a
+    reason to fail a preview — the plain conversion is still a correct picture.
+    """
+    profile = image.info.get("icc_profile")
+    if profile:
+        try:
+            return ImageCms.profileToProfile(
+                image,
+                ImageCms.ImageCmsProfile(io.BytesIO(profile)),
+                ImageCms.createProfile("sRGB"),
+                outputMode=mode,
+            )
+        except Exception as error:  # noqa: BLE001 - any ICC failure falls back
+            log.info("ignoring unusable ICC profile on %s: %s", getattr(image, "filename", "?"), error)
+    return image.convert(mode)
diff --git a/photo_pipeline/integrations/exiftool.py b/photo_pipeline/integrations/exiftool.py
index 92cd64e..56aae86 100644
--- a/photo_pipeline/integrations/exiftool.py
+++ b/photo_pipeline/integrations/exiftool.py
@@ -56,6 +56,30 @@ def read_keyword_sets(paths: Iterable[str]) -> dict[str, set[str]]:
     return out
 
 
+def read_all(path: str) -> dict | None:
+    """Every tag exiftool can read from ``path``, or ``None`` when it cannot answer.
+
+    This is the snapshot an EXIF checkpoint compares against: proving that a write
+    preserved the fields it does not own requires knowing all of them, not just the
+    ones being written (US07-03). ``None`` (exiftool missing, unreadable file,
+    unparsable output) is not an empty snapshot — a caller must not read it as
+    "nothing was there".
+    """
+    try:
+        result = subprocess.run(
+            ["exiftool", "-m", "-j", "-G0:1", path], capture_output=True, text=True
+        )
+    except FileNotFoundError:
+        return None
+    try:
+        records = json.loads(result.stdout or "[]")
+    except ValueError:
+        return None
+    if not records:
+        return None
+    return {k: v for k, v in records[0].items() if k != "SourceFile"}
+
+
 def apply_keywords(path: str, *, add: Iterable[str] = (), remove: Iterable[str] = ()) -> bool:
     """Idempotently add/remove keywords in Keywords + Subject; preserve all else."""
     args = ["exiftool", "-m", "-overwrite_original"]
diff --git a/photo_pipeline/integrations/nsfw_model.py b/photo_pipeline/integrations/nsfw_model.py
index b28327e..d2fbe36 100644
--- a/photo_pipeline/integrations/nsfw_model.py
+++ b/photo_pipeline/integrations/nsfw_model.py
@@ -15,6 +15,8 @@ from __future__ import annotations
 
 from pathlib import Path
 
+from photo_pipeline import imaging
+
 MODEL_ID = "AdamCodd/vit-base-nsfw-detector"
 BATCH = 16
 
@@ -54,13 +56,17 @@ class NsfwModel:
         self._ensure_loaded()
         import numpy as np
         import torch
-        from PIL import Image, ImageFile
+        from PIL import Image
 
-        ImageFile.LOAD_TRUNCATED_IMAGES = True
-
-        def preprocess(image):
-            image = image.convert("RGB").resize((self._size, self._size), Image.BILINEAR)
-            array = (np.asarray(image, dtype="float32") / 255.0 - 0.5) / 0.5
+        # The donor set ``ImageFile.LOAD_TRUNCATED_IMAGES = True`` here. That flag is
+        # process-global: in this application the same process also hashes files and
+        # renders previews, and those must keep failing loudly on a truncated file
+        # rather than quietly working on half of one (US07-03). An unreadable image
+        # is skipped instead — it stays unscored, and therefore visibly undecided.
+        def preprocess(path):
+            with imaging.open_image(path) as image:
+                small = image.convert("RGB").resize((self._size, self._size), Image.BILINEAR)
+            array = (np.asarray(small, dtype="float32") / 255.0 - 0.5) / 0.5
             return torch.from_numpy(array).permute(2, 0, 1)
 
         results: list[tuple[str, float]] = []
@@ -69,9 +75,10 @@ class NsfwModel:
             tensors, batch_paths = [], []
             for path in items[start : start + self.batch]:
                 try:
-                    tensors.append(preprocess(Image.open(path)))
+                    tensors.append(preprocess(path))
                     batch_paths.append(path)
-                except Exception:
+                except (imaging.MediaError, OSError, ValueError):
+                    # One bad file must not cost the batch its other fifteen.
                     continue
             if not tensors:
                 continue
diff --git a/photo_pipeline/models/__init__.py b/photo_pipeline/models/__init__.py
index 0beb416..a8234dd 100644
--- a/photo_pipeline/models/__init__.py
+++ b/photo_pipeline/models/__init__.py
@@ -7,6 +7,7 @@ Alembic environment relies on.
 from photo_pipeline.models.albums import AlbumProposal
 from photo_pipeline.models.archives import ArchiveLocation, ArchiveOperation, ArchivePlan
 from photo_pipeline.models.assets import Asset, AssetPath
+from photo_pipeline.models.exif import ExifProjection
 from photo_pipeline.models.duplicates import (
     DuplicateCluster,
     DuplicateMember,
@@ -28,6 +29,7 @@ __all__ = [
     "DuplicateCluster",
     "DuplicateMember",
     "DuplicateNegativeLink",
+    "ExifProjection",
     "Job",
     "JobItem",
     "JobEvent",
diff --git a/photo_pipeline/models/exif.py b/photo_pipeline/models/exif.py
new file mode 100644
index 0000000..96d9e0a
--- /dev/null
+++ b/photo_pipeline/models/exif.py
@@ -0,0 +1,33 @@
+"""The durable EXIF projection per asset and stage (concept §3, US07-03).
+
+One row per ``(asset_id, stage)``: what the stage wanted written, what the file
+looked like afterwards, and whether anything outside the stage's ownership moved.
+``state = divergent`` is the whole point of the table — it survives restarts, keeps
+the asset out of stages that require verified metadata, and gives a human something
+to look at instead of a silent repair.
+"""
+
+from __future__ import annotations
+
+from datetime import datetime
+
+from sqlalchemy import DateTime, ForeignKey, Integer, String
+from sqlalchemy.orm import Mapped, mapped_column
+
+from photo_pipeline.db import Base
+
+
+class ExifProjection(Base):
+    __tablename__ = "exif_projections"
+
+    asset_id: Mapped[str] = mapped_column(ForeignKey("assets.id"), primary_key=True)
+    stage: Mapped[str] = mapped_column(String, primary_key=True)  # safety | analysis
+    id: Mapped[str] = mapped_column(String, nullable=False)
+    projection_version: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
+    desired_json: Mapped[str | None] = mapped_column(String)
+    divergent_fields: Mapped[str | None] = mapped_column(String)  # JSON array
+    result_file_sha256: Mapped[str | None] = mapped_column(String)
+    state: Mapped[str] = mapped_column(String, nullable=False)  # verified|divergent|failed
+    error_code: Mapped[str | None] = mapped_column(String)
+    verified_at: Mapped[datetime | None] = mapped_column(DateTime(timezone=True))
+    updated_at: Mapped[datetime | None] = mapped_column(DateTime(timezone=True))
diff --git a/photo_pipeline/services/analysis.py b/photo_pipeline/services/analysis.py
index 5c4ae8f..3698d5d 100644
--- a/photo_pipeline/services/analysis.py
+++ b/photo_pipeline/services/analysis.py
@@ -26,9 +26,8 @@ from sqlalchemy import func, select
 from sqlalchemy.orm import sessionmaker
 
 from photo_pipeline import path_policy
-from photo_pipeline.integrations import exiftool
 from photo_pipeline.models import AnalysisResult, Asset, SafetyReview
-from photo_pipeline.services import hashing
+from photo_pipeline.services import exif_checkpoint
 from photo_pipeline.services.safety import SFW
 
 MODEL = "gemini-2.5-flash"
@@ -170,7 +169,7 @@ class AnalysisService:
                 tokens=int(result.get("_tokens", 0)) if isinstance(result, dict) else 0,
                 raw=json.dumps(result, ensure_ascii=False),
             )
-            _write_analysis_exif(path, result)
+            self._write_analysis_exif(asset_id, path, result)
             analyzed += 1
         return {"analyzed": analyzed, "skipped": skipped, "errors": errors}
 
@@ -195,10 +194,42 @@ class AnalysisService:
                 row.location_hint = result.get("location_hint")
                 row.approx_year = result.get("approx_year")
                 row.analyzed_at = now
-                row.exif_written_at = now
             session.add(row)
             session.commit()
 
+    def _write_analysis_exif(self, asset_id: str, path: str, result: dict) -> None:
+        """The analysis EXIF checkpoint: additive keywords, then prove the rest held.
+
+        Additive by design — safety keywords and the user's own keywords are merged
+        with, never replaced (concept §3). ``exif_written_at`` is set only when the
+        read-back verified both the new keywords and every field this stage does not
+        own; a divergent result is recorded and left for a human (US07-03).
+        ponytail: the managed ``AI:`` caption segment, once captions are owned here.
+        """
+        tags = tuple(str(tag) for tag in (result.get("tags") or []))
+        if not tags:
+            return
+        checkpoint = exif_checkpoint.run(path, add=tags)
+        exif_checkpoint.record(
+            self._session_factory,
+            asset_id=asset_id,
+            stage="analysis",
+            result=checkpoint,
+            add=tags,
+        )
+        if not checkpoint.verified:
+            return
+        with self._session_factory() as session:
+            row = session.get(AnalysisResult, asset_id)
+            if row is not None:
+                row.exif_written_at = checkpoint.verified_at
+            asset = session.get(Asset, asset_id)
+            if asset is not None and checkpoint.sha256:
+                # The bytes changed when the container was rewritten; upload must use
+                # the hash of what is actually on disk now (concept §3).
+                asset.current_sha256 = checkpoint.sha256
+            session.commit()
+
     def get(self, asset_id: str) -> dict | None:
         with self._session_factory() as session:
             row = session.get(AnalysisResult, asset_id)
@@ -211,17 +242,6 @@ def _album_hint(path: str) -> str:
     return Path(path).parent.name
 
 
-def _write_analysis_exif(path: str, result: dict) -> None:
-    """Additive analysis keywords into EXIF (Keywords/Subject), preserving safety
-    and user keywords. The donor also wrote a managed caption; only keywords are
-    written here via the shared adapter.
-    ponytail: add the managed ``AI:`` caption segment + read-back preservation check
-    when the analysis EXIF checkpoint is hardened."""
-    tags = result.get("tags") if isinstance(result, dict) else None
-    if tags:
-        exiftool.apply_keywords(path, add=[str(t) for t in tags])
-
-
 def _result_dict(row: AnalysisResult) -> dict:
     data = {field: getattr(row, field) for field in RESULT_FIELDS}
     data["tags"] = json.loads(row.tags) if row.tags else []
diff --git a/photo_pipeline/services/exif_checkpoint.py b/photo_pipeline/services/exif_checkpoint.py
new file mode 100644
index 0000000..17bf5b3
--- /dev/null
+++ b/photo_pipeline/services/exif_checkpoint.py
@@ -0,0 +1,173 @@
+"""The EXIF checkpoint every metadata stage ends with (concept §3, US07-03).
+
+A stage does not own a file's metadata; it owns a few fields in it. So writing is
+never "set these tags" — it is:
+
+    snapshot everything → write only the owned fields → read everything back
+    → prove the owned fields landed → prove nothing else moved
+    → refresh the file hash → record the projection
+
+Non-destructive here means *semantic* preservation: exiftool may rewrite the whole
+container, so the file's bytes, size, and timestamps legitimately change. What may
+not change is any field this stage does not own. When one does, the checkpoint is
+``divergent``: the result is recorded, the stage is **not** marked verified, and
+nothing is silently repaired — a later stage that needs verified metadata (upload)
+therefore stays blocked until a human looks.
+
+``failed`` is the third outcome and is deliberately distinct: exiftool missing, an
+unreadable file, or a write that did not take is not evidence that metadata is fine.
+"""
+
+from __future__ import annotations
+
+import json
+import uuid
+from dataclasses import dataclass
+from datetime import datetime, timezone
+
+from photo_pipeline.integrations import exiftool
+from photo_pipeline.models import ExifProjection
+from photo_pipeline.services import hashing
+
+# The tags the safety and analysis stages may change. Matched on the tag name, so
+# every group carries the same rule (IPTC:Keywords, XMP:XMP-dc:Subject, ...).
+OWNED_TAGS = frozenset({"Keywords", "Subject"})
+
+# Not metadata about the picture: filesystem facts, the digest that necessarily
+# moves whenever IPTC does, and the structural tags exiftool has to create the first
+# time it writes an IPTC or XMP block. Comparing these would report every write as
+# divergent and make the signal worthless.
+VOLATILE_PREFIXES = ("File:System:", "ExifTool:")
+VOLATILE_KEYS = frozenset(
+    {
+        "File:CurrentIPTCDigest",
+        "IPTC:ApplicationRecordVersion",
+        "XMP:XMP-x:XMPToolkit",
+        "XMP:XMP-xmp:MetadataDate",
+    }
+)
+
+VERIFIED = "verified"
+DIVERGENT = "divergent"
+FAILED = "failed"
+
+
+@dataclass(frozen=True)
+class CheckpointResult:
+    state: str  # verified | divergent | failed
+    changed_fields: tuple[str, ...] = ()
+    sha256: str | None = None
+    verified_at: datetime | None = None
+    reason: str | None = None
+
+    @property
+    def verified(self) -> bool:
+        return self.state == VERIFIED
+
+
+def _now() -> datetime:
+    return datetime.now(timezone.utc)
+
+
+def is_owned(key: str) -> bool:
+    return key.rsplit(":", 1)[-1] in OWNED_TAGS
+
+
+def is_volatile(key: str) -> bool:
+    return key in VOLATILE_KEYS or key.startswith(VOLATILE_PREFIXES)
+
+
+def compare(before: dict, after: dict) -> tuple[str, ...]:
+    """Fields outside this stage's ownership whose value did not survive the write.
+
+    Additions count: a tag that appears out of nowhere is as much a divergence as a
+    tag that disappeared — both mean the write did more than it was asked to.
+    """
+    keys = set(before) | set(after)
+    return tuple(
+        sorted(
+            key
+            for key in keys
+            if not is_owned(key)
+            and not is_volatile(key)
+            and before.get(key) != after.get(key)
+        )
+    )
+
+
+def owned_values(snapshot: dict) -> set[str]:
+    """Lowercased Keywords/Subject values across every group in the snapshot."""
+    values: set[str] = set()
+    for key, value in snapshot.items():
+        if not is_owned(key):
+            continue
+        items = value if isinstance(value, list) else [value]
+        values.update(str(item).strip().lower() for item in items if item is not None)
+    return values
+
+
+def run(
+    path: str,
+    *,
+    add: tuple[str, ...] = (),
+    remove: tuple[str, ...] = (),
+) -> CheckpointResult:
+    """Write the owned keywords for one asset and verify the whole file around them."""
+    before = exiftool.read_all(path)
+    if before is None:
+        return CheckpointResult(FAILED, reason="metadata_unreadable")
+
+    if not exiftool.apply_keywords(path, add=add, remove=remove):
+        return CheckpointResult(FAILED, reason="write_failed")
+
+    after = exiftool.read_all(path)
+    if after is None:
+        return CheckpointResult(FAILED, reason="readback_unreadable")
+
+    present = owned_values(after)
+    wanted = {value.strip().lower() for value in add}
+    unwanted = {value.strip().lower() for value in remove}
+    if not wanted <= present or (unwanted & present):
+        return CheckpointResult(FAILED, reason="owned_fields_not_written")
+
+    changed = compare(before, after)
+    sha256 = hashing.sha256_file(path)
+    if changed:
+        return CheckpointResult(DIVERGENT, changed_fields=changed, sha256=sha256)
+    return CheckpointResult(VERIFIED, sha256=sha256, verified_at=_now())
+
+
+def record(
+    session_factory,
+    *,
+    asset_id: str,
+    stage: str,
+    result: CheckpointResult,
+    add: tuple[str, ...] = (),
+    remove: tuple[str, ...] = (),
+) -> None:
+    """Persist the projection for ``(asset_id, stage)`` — one current row per pair.
+
+    The row is what makes divergence durable and reviewable rather than a log line
+    that scrolled away.
+    """
+    with session_factory() as session:
+        row = session.get(ExifProjection, (asset_id, stage))
+        if row is None:
+            row = ExifProjection(asset_id=asset_id, stage=stage, id=str(uuid.uuid4()))
+            session.add(row)
+        row.projection_version = (row.projection_version or 0) + 1
+        row.desired_json = json.dumps({"add": list(add), "remove": list(remove)})
+        row.divergent_fields = json.dumps(list(result.changed_fields))
+        row.result_file_sha256 = result.sha256
+        row.state = result.state
+        row.error_code = result.reason
+        row.verified_at = result.verified_at
+        row.updated_at = _now()
+        session.commit()
+
+
+def state_for(session_factory, asset_id: str, stage: str) -> str | None:
+    with session_factory() as session:
+        row = session.get(ExifProjection, (asset_id, stage))
+        return row.state if row else None
diff --git a/photo_pipeline/services/hashing.py b/photo_pipeline/services/hashing.py
index 9718683..6fb5dea 100644
--- a/photo_pipeline/services/hashing.py
+++ b/photo_pipeline/services/hashing.py
@@ -16,7 +16,9 @@ from __future__ import annotations
 import hashlib
 from pathlib import Path
 
-from PIL import Image, ImageOps
+from PIL import ImageOps
+
+from photo_pipeline import imaging
 
 PIXEL_HASH_VERSION = 1
 PHASH_VERSION = 1
@@ -41,7 +43,9 @@ def _digest_file(path: Path | str, digest) -> str:
 
 
 def pixel_sha256(path: Path | str) -> str:
-    with Image.open(path) as image:
+    # Bounded decode: a header claiming a billion pixels is refused before it is
+    # ever turned into memory (US07-03).
+    with imaging.open_image(path) as image:
         oriented = ImageOps.exif_transpose(image)
         rgb = oriented.convert("RGB")
         header = f"{PIXEL_HASH_VERSION}:{rgb.width}x{rgb.height}:".encode()
@@ -68,7 +72,9 @@ def phash(path: Path | str) -> str:
     import numpy as np
     from scipy.fftpack import dct
 
-    with Image.open(path) as image:
+    from PIL import Image
+
+    with imaging.open_image(path) as image:
         small = image.convert("L").resize((32, 32), Image.LANCZOS)
         matrix = np.asarray(small, dtype=np.float64)
     transformed = dct(dct(matrix, axis=0), axis=1)
diff --git a/photo_pipeline/services/safety.py b/photo_pipeline/services/safety.py
index 5983c45..00652da 100644
--- a/photo_pipeline/services/safety.py
+++ b/photo_pipeline/services/safety.py
@@ -108,9 +108,8 @@ from datetime import datetime, timezone
 from sqlalchemy import select
 from sqlalchemy.orm import sessionmaker
 
-from photo_pipeline.integrations import exiftool
-from photo_pipeline.models import Asset, SafetyReview
-from photo_pipeline.services import hashing
+from photo_pipeline.models import Asset, ExifProjection, SafetyReview
+from photo_pipeline.services import exif_checkpoint
 
 DECISIONS = {SFW, NSFW, "deferred"}
 
@@ -174,6 +173,14 @@ class SafetyService:
         with self._session_factory() as session:
             assets = list(session.scalars(_eligible_assets_query().order_by(Asset.current_path)))
             latest = self._latest_by_asset(session)
+            # One query, not one per asset: the reviewer needs to see a divergent
+            # checkpoint, which is neither "verified" nor a plain failure (US07-03).
+            projections = {
+                row.asset_id: row.state
+                for row in session.scalars(
+                    select(ExifProjection).where(ExifProjection.stage == "safety")
+                )
+            }
             rows = []
             for asset in assets:
                 review = latest.get(asset.id)
@@ -189,6 +196,7 @@ class SafetyService:
                         "decision": decision,
                         "suggested": classify(review.score) if review and review.score is not None else None,
                         "exif_verified": bool(review and review.exif_verified_at),
+                        "exif_state": projections.get(asset.id),
                     }
                 )
             return {"total": len(rows), "items": rows[offset : offset + limit]}
@@ -247,13 +255,24 @@ class SafetyService:
         result_sha256 = None
         if write_exif and decision in (SFW, NSFW) and path:
             ops = exif_projection(decision)
-            if exiftool.apply_keywords(path, add=ops["add"], remove=ops["remove"]):
-                # Read back: the chosen keyword present, the opposite absent.
-                keywords = exiftool.read_keyword_sets([path]).get(path, set())
-                opposite = NSFW if decision == SFW else SFW
-                if decision in keywords and opposite not in keywords:
-                    exif_verified_at = _now()
-                    result_sha256 = hashing.sha256_file(path)
+            # The full checkpoint: write the owned keyword, read the whole file back,
+            # and prove every field this stage does not own survived. A divergent
+            # result is recorded and left alone — it must not count as verified, so
+            # upload stays blocked until a human decides (US07-03).
+            result = exif_checkpoint.run(
+                path, add=tuple(ops["add"]), remove=tuple(ops["remove"])
+            )
+            exif_checkpoint.record(
+                self._session_factory,
+                asset_id=asset_id,
+                stage="safety",
+                result=result,
+                add=tuple(ops["add"]),
+                remove=tuple(ops["remove"]),
+            )
+            if result.verified:
+                exif_verified_at = result.verified_at
+                result_sha256 = result.sha256
 
         now = _now()
         with self._session_factory() as session:
diff --git a/photo_pipeline/services/thumbnails.py b/photo_pipeline/services/thumbnails.py
index fb7fce9..a2a3a29 100644
--- a/photo_pipeline/services/thumbnails.py
+++ b/photo_pipeline/services/thumbnails.py
@@ -20,15 +20,16 @@ pa-imaging).
 
 from __future__ import annotations
 
+import os
 import uuid
 from pathlib import Path
 
-from PIL import Image, ImageOps, UnidentifiedImageError
+from PIL import Image, ImageOps
 from sqlalchemy import select
 from sqlalchemy.exc import IntegrityError
 from sqlalchemy.orm import sessionmaker
 
-from photo_pipeline import path_policy
+from photo_pipeline import imaging, path_policy
 from photo_pipeline.config import Config
 from photo_pipeline.models import Asset, Thumbnail
 from photo_pipeline.services import availability
@@ -42,6 +43,9 @@ except Exception:  # pragma: no cover
     pass
 
 SIZES = (256, 512, 1280)
+# Every in-flight render writes one of these beside its destination; startup
+# cleanup recognises exactly this pattern and nothing else.
+TEMP_SUFFIX = ".tmp"
 THUMB_VERSION = 1
 THUMB_FORMAT = "webp"
 # The size kept as durable comparison evidence for archived assets (concept §9).
@@ -222,28 +226,33 @@ class ThumbnailService:
         return self._cache_dir / safe[:2] / f"{safe}.{THUMB_FORMAT}"
 
     def _render(self, source: str, size: int, cache_key: str) -> dict:
+        """Decode bounded, orient, colour-manage, resize, and write atomically.
+
+        The temporary file is removed on every failure path: a decoder that dies
+        halfway through ``save`` would otherwise leave a stray ``.tmp`` in the cache
+        forever (US07-03).
+        """
+        destination = self._cache_path(cache_key)
+        destination.parent.mkdir(parents=True, exist_ok=True)
+        tmp = destination.with_name(f".{destination.name}.{uuid.uuid4().hex}{TEMP_SUFFIX}")
         try:
-            with Image.open(source) as image:
-                width, height = image.size
-                if width * height > self._config.thumbnail_max_pixels:
-                    raise ImageTooLarge(
-                        f"{width}x{height} exceeds {self._config.thumbnail_max_pixels} px"
-                    )
+            with imaging.open_image(source, max_pixels=self._config.thumbnail_max_pixels) as image:
+                imaging.draft(image, size)  # JPEG decodes near the target size
                 oriented = ImageOps.exif_transpose(image)
                 mode = "RGBA" if _has_alpha(oriented) else "RGB"
-                converted = oriented.convert(mode)
+                converted = imaging.to_srgb(oriented, mode=mode)
                 converted.thumbnail((size, size), Image.LANCZOS)
                 out_width, out_height = converted.size
-                destination = self._cache_path(cache_key)
-                destination.parent.mkdir(parents=True, exist_ok=True)
-                tmp = destination.with_name(f".{destination.name}.{uuid.uuid4().hex}.tmp")
                 converted.save(tmp, format="WEBP", quality=82, method=4)
-        except ImageTooLarge:
+        except imaging.ImageTooLarge as error:
+            tmp.unlink(missing_ok=True)
+            raise ImageTooLarge(str(error)) from None
+        except imaging.UndecodableImage as error:
+            tmp.unlink(missing_ok=True)
+            raise UnsupportedImage(str(error)) from None
+        except Exception:
+            tmp.unlink(missing_ok=True)
             raise
-        except (UnidentifiedImageError, OSError, ValueError) as error:
-            raise UnsupportedImage(f"cannot decode {source}: {error}") from error
-
-        import os
 
         os.replace(tmp, destination)
         return {
@@ -253,6 +262,28 @@ class ThumbnailService:
             "format": THUMB_FORMAT,
         }
 
+    def cleanup_temp_files(self) -> int:
+        """Remove leftover render temporaries, and only those.
+
+        Runs at startup, where the concept is explicit: clean *recognized* stale
+        temporary files, never arbitrary ones. The pattern is this service's own —
+        a dot-prefixed name inside the managed cache directory ending in
+        ``.tmp`` — so nothing outside the cache and nothing a user put there can
+        match. Returns how many were removed.
+        """
+        if not self._cache_dir.is_dir():
+            return 0
+        removed = 0
+        for leftover in self._cache_dir.rglob(f".*{TEMP_SUFFIX}"):
+            if not leftover.is_file() or leftover.is_symlink():
+                continue
+            try:
+                leftover.unlink()
+                removed += 1
+            except OSError:  # pragma: no cover - a racing render already won
+                pass
+        return removed
+
     # ── persistence ────────────────────────────────────────────────────────────
     def _record_ready(
         self, cache_key: str, asset_id: str, size: int, rendered: dict, *, protected: bool = False
diff --git a/tests/fixtures/media_corpus.py b/tests/fixtures/media_corpus.py
new file mode 100644
index 0000000..7a7f235
--- /dev/null
+++ b/tests/fixtures/media_corpus.py
@@ -0,0 +1,325 @@
+"""The golden media corpus: every format, orientation, profile, damage, and
+metadata case this application claims to survive (US07-03, concept §18).
+
+``CASES`` is the manifest and the authority. Each entry declares a stable logical
+id (never a path), how the file is generated, and what the pipeline must do with
+it — decode it, or refuse it with one precise error code. Tests parametrize over the
+manifest, so a case that is added here without an expectation, or an expectation
+that stops holding, fails the suite rather than quietly going untested.
+
+Everything is generated, never committed: fixed pixel seeds, fixed EXIF strings, no
+clock, no network, no personal data. Regeneration is byte-stable, which
+``test_media_hardening.py`` proves by building the corpus twice and comparing
+checksums — a golden corpus that drifts is not golden.
+"""
+
+from __future__ import annotations
+
+import struct
+import zlib
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Callable
+
+import numpy as np
+from PIL import Image
+
+# Error codes the pipeline may answer with; ``None`` means "must render".
+UNSUPPORTED = "unsupported_image"
+TOO_LARGE = "image_too_large"
+
+# One fixed capture time for every metadata case: the corpus must not depend on when
+# it was generated.
+CAPTURE_TIME = "2019:07:14 10:30:00"
+
+
+@dataclass(frozen=True)
+class MediaCase:
+    id: str
+    filename: str
+    kind: str  # format | orientation | profile | damaged | metadata
+    build: Callable[[Path], None]
+    expect_error: str | None = None
+    # Declared for the metadata cases: exiftool arguments applied after the pixels
+    # are written, and the user fields that must survive every later stage.
+    exif_args: tuple[str, ...] = ()
+    preserved_fields: tuple[str, ...] = ()
+    notes: str = ""
+    tags: tuple[str, ...] = field(default_factory=tuple)
+
+
+# ── generators ────────────────────────────────────────────────────────────────
+
+
+def _pixels(width: int, height: int, seed: int, bands: int = 3) -> np.ndarray:
+    return np.random.default_rng(seed).integers(0, 256, (height, width, bands), dtype=np.uint8)
+
+
+def _save(path: Path, image: Image.Image, **kwargs) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    image.save(path, **kwargs)
+
+
+def _jpeg(width=320, height=240, seed=1, **kwargs):
+    def build(path: Path) -> None:
+        _save(path, Image.fromarray(_pixels(width, height, seed)), quality=90, **kwargs)
+
+    return build
+
+
+def _oriented_jpeg(orientation: int):
+    def build(path: Path) -> None:
+        image = Image.fromarray(_pixels(400, 200, 5))  # landscape source
+        exif = image.getexif()
+        exif[274] = orientation  # 0x0112 Orientation
+        _save(path, image, exif=exif, quality=90)
+
+    return build
+
+
+def _rotated_pixels(path: Path) -> None:
+    """The same scene rotated in the pixels instead of in a tag."""
+    _save(path, Image.fromarray(_pixels(400, 200, 5)).transpose(Image.ROTATE_90), quality=90)
+
+
+def _png(alpha: bool = False, seed: int = 2):
+    def build(path: Path) -> None:
+        if alpha:
+            _save(path, Image.fromarray(_pixels(120, 90, seed, bands=4), "RGBA"))
+        else:
+            _save(path, Image.fromarray(_pixels(120, 90, seed)))
+
+    return build
+
+
+def _webp(path: Path) -> None:
+    _save(path, Image.fromarray(_pixels(150, 100, 3)), quality=80)
+
+
+def _tiff(path: Path) -> None:
+    _save(path, Image.fromarray(_pixels(140, 110, 4)))
+
+
+def _grayscale(path: Path) -> None:
+    _save(path, Image.fromarray(_pixels(100, 100, 6)).convert("L"), quality=90)
+
+
+def _cmyk(path: Path) -> None:
+    _save(path, Image.fromarray(_pixels(100, 100, 7)).convert("CMYK"), quality=90)
+
+
+def _tiny(path: Path) -> None:
+    _save(path, Image.fromarray(_pixels(1, 1, 8)))
+
+
+def _icc_tagged(path: Path) -> None:
+    """A profile-bearing image: the colour-managed decode path must run."""
+    from PIL import ImageCms
+
+    profile = bytearray(ImageCms.ImageCmsProfile(ImageCms.createProfile("sRGB")).tobytes())
+    # An ICC header carries its creation timestamp at bytes 24..35. Left alone, the
+    # corpus would be a different corpus every time it is generated.
+    profile[24:36] = b"\x00" * 12
+    _save(path, Image.fromarray(_pixels(120, 80, 9)), icc_profile=bytes(profile), quality=90)
+
+
+def _broken_icc(path: Path) -> None:
+    """A profile that is not a profile: a picture is still a picture."""
+    _save(path, Image.fromarray(_pixels(120, 80, 10)), icc_profile=b"not-a-profile", quality=90)
+
+
+def _wide_jpeg(path: Path) -> None:
+    """Large enough that decoding it at full resolution is visible in memory."""
+    _save(path, Image.fromarray(_pixels(4000, 3000, 11)), quality=70)
+
+
+def _zero_byte(path: Path) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_bytes(b"")
+
+
+def _truncated_jpeg(path: Path) -> None:
+    image = Image.fromarray(_pixels(400, 300, 12))
+    _save(path, image, quality=90)
+    data = path.read_bytes()
+    path.write_bytes(data[: len(data) // 2])  # header intact, pixels missing
+
+
+def _corrupt_png(path: Path) -> None:
+    image = Image.fromarray(_pixels(120, 90, 13))
+    _save(path, image)
+    data = bytearray(path.read_bytes())
+    data[40:80] = b"\x00" * 40  # shred the compressed stream, keep the header
+    path.write_bytes(bytes(data))
+
+
+def _not_an_image(path: Path) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_bytes(b"This is a text file that happens to be named .jpg\n")
+
+
+def _png_declaring(width: int, height: int):
+    """A tiny, structurally valid PNG whose header claims an enormous picture.
+
+    A few hundred bytes on disk, gigapixels on paper: the pipeline must refuse it
+    from the declared dimensions, before a single pixel is allocated. Rewriting the
+    IHDR of a real PNG (rather than hand-rolling a stub) keeps the file openable, so
+    the refusal is proven to come from the size check and not from a parse failure.
+    """
+
+    def build(path: Path) -> None:
+        import io
+
+        buffer = io.BytesIO()
+        Image.fromarray(_pixels(4, 4, 15)).save(buffer, format="PNG")
+        data = bytearray(buffer.getvalue())
+        start = 8 + 4  # PNG signature, then the IHDR length field
+        struct.pack_into(">II", data, start + 4, width, height)
+        ihdr = bytes(data[start : start + 4 + 13])
+        struct.pack_into(">I", data, start + 4 + 13, zlib.crc32(ihdr))
+        path.parent.mkdir(parents=True, exist_ok=True)
+        path.write_bytes(bytes(data))
+
+    return build
+
+
+CASES: tuple[MediaCase, ...] = (
+    # ── formats ───────────────────────────────────────────────────────────────
+    MediaCase("jpeg", "formats/plain.jpg", "format", _jpeg()),
+    MediaCase("jpeg_uppercase_ext", "formats/UPPER.JPG", "format", _jpeg(seed=14)),
+    MediaCase("png", "formats/plain.png", "format", _png()),
+    MediaCase("png_alpha", "formats/alpha.png", "format", _png(alpha=True)),
+    MediaCase("webp", "formats/plain.webp", "format", _webp),
+    MediaCase("tiff", "formats/plain.tiff", "format", _tiff),
+    # ── orientation ───────────────────────────────────────────────────────────
+    *(
+        MediaCase(
+            f"orientation_{value}",
+            f"orientation/exif_{value}.jpg",
+            "orientation",
+            _oriented_jpeg(value),
+            notes="EXIF orientation must be applied before resizing",
+        )
+        for value in range(1, 9)
+    ),
+    MediaCase("rotated_pixels", "orientation/rotated.jpg", "orientation", _rotated_pixels),
+    # ── colour and size profiles ──────────────────────────────────────────────
+    MediaCase("grayscale", "profiles/gray.jpg", "profile", _grayscale),
+    MediaCase("cmyk", "profiles/cmyk.jpg", "profile", _cmyk),
+    MediaCase("tiny", "profiles/tiny.png", "profile", _tiny),
+    MediaCase("icc_tagged", "profiles/icc.jpg", "profile", _icc_tagged),
+    MediaCase(
+        "broken_icc",
+        "profiles/broken-icc.jpg",
+        "profile",
+        _broken_icc,
+        notes="an unusable ICC profile falls back to a plain conversion, never an error",
+    ),
+    MediaCase(
+        "large_jpeg",
+        "profiles/large.jpg",
+        "profile",
+        _wide_jpeg,
+        notes="12 megapixels: the decode must stay near the requested size",
+    ),
+    # ── damaged and hostile inputs ────────────────────────────────────────────
+    MediaCase("zero_byte", "damaged/empty.jpg", "damaged", _zero_byte, UNSUPPORTED),
+    MediaCase("truncated_jpeg", "damaged/truncated.jpg", "damaged", _truncated_jpeg, UNSUPPORTED),
+    MediaCase("corrupt_png", "damaged/corrupt.png", "damaged", _corrupt_png, UNSUPPORTED),
+    MediaCase("text_as_jpeg", "damaged/text.jpg", "damaged", _not_an_image, UNSUPPORTED),
+    MediaCase(
+        "bomb_header",
+        "damaged/bomb.png",
+        "damaged",
+        _png_declaring(60_000, 60_000),
+        TOO_LARGE,
+        notes="3.6 gigapixels declared in the header and nothing else",
+    ),
+    MediaCase(
+        "bomb_warning_band",
+        "damaged/bomb-warning.png",
+        "damaged",
+        _png_declaring(10_000, 10_000),
+        TOO_LARGE,
+        notes="inside Pillow's warn-only band; the warning is promoted to a refusal",
+    ),
+    # ── metadata ──────────────────────────────────────────────────────────────
+    MediaCase("no_exif", "metadata/bare.jpg", "metadata", _jpeg(seed=20)),
+    MediaCase(
+        "user_exif",
+        "metadata/user.jpg",
+        "metadata",
+        _jpeg(seed=21),
+        exif_args=(
+            "-Artist=Ada Lovelace",
+            "-Copyright=(c) Ada",
+            f"-DateTimeOriginal={CAPTURE_TIME}",
+            "-GPSLatitude=48.137",
+            "-GPSLatitudeRef=N",
+            "-Rating=4",
+            "-ImageDescription=A day out",
+        ),
+        preserved_fields=(
+            "EXIF:IFD0:Artist",
+            "EXIF:IFD0:Copyright",
+            "EXIF:ExifIFD:DateTimeOriginal",
+            "EXIF:IFD0:ImageDescription",
+            "XMP:XMP-xmp:Rating",
+        ),
+        notes="user metadata that every stage must leave exactly as it found it",
+    ),
+    MediaCase(
+        "prior_safety_keyword",
+        "metadata/prior-safety.jpg",
+        "metadata",
+        _jpeg(seed=22),
+        exif_args=("-Keywords+=nsfw", "-Subject+=nsfw", "-Artist=Ada Lovelace"),
+        preserved_fields=("EXIF:IFD0:Artist",),
+        notes="a safety decision already written by an earlier run",
+    ),
+    MediaCase(
+        "prior_analysis_keywords",
+        "metadata/prior-analysis.jpg",
+        "metadata",
+        _jpeg(seed=23),
+        exif_args=("-Keywords+=beach", "-Keywords+=sunset", "-Subject+=beach", "-Subject+=sunset"),
+        notes="analysis keywords from an earlier run; a safety write must not drop them",
+    ),
+    MediaCase(
+        "conflicting_safety_keywords",
+        "metadata/conflicting.jpg",
+        "metadata",
+        _jpeg(seed=24),
+        exif_args=("-Keywords+=sfw", "-Keywords+=nsfw", "-Subject+=sfw", "-Subject+=nsfw"),
+        notes="both safety keywords at once: mutually exclusive means one must go",
+    ),
+    MediaCase(
+        "malformed_metadata",
+        "metadata/malformed.jpg",
+        "metadata",
+        _jpeg(seed=25, exif=b"\x00\x01\x02not-a-valid-exif-block"),
+        notes="a broken EXIF block must not stop the picture from being usable",
+    ),
+)
+
+CASES_BY_ID = {case.id: case for case in CASES}
+
+
+def build_corpus(root: Path, *, ids: tuple[str, ...] | None = None) -> dict[str, Path]:
+    """Generate the corpus (or a named subset) under ``root``; return id → path."""
+    import subprocess
+
+    built: dict[str, Path] = {}
+    for case in CASES:
+        if ids is not None and case.id not in ids:
+            continue
+        path = root / case.filename
+        case.build(path)
+        if case.exif_args:
+            subprocess.run(
+                ["exiftool", "-m", "-overwrite_original", *case.exif_args, str(path)],
+                capture_output=True,
+                check=False,
+            )
+        built[case.id] = path
+    return built
diff --git a/tests/integration/test_exif_checkpoints.py b/tests/integration/test_exif_checkpoints.py
new file mode 100644
index 0000000..b1426bb
--- /dev/null
+++ b/tests/integration/test_exif_checkpoints.py
@@ -0,0 +1,297 @@
+"""US07-03: EXIF checkpoints, asserted with before/after metadata snapshots.
+
+Every test here reads the complete metadata of a real file through exiftool before
+the stage runs and again afterwards, then compares the two. That is the only way to
+prove the property the concept actually asks for: a stage owns a few fields and must
+leave literally everything else — dates, GPS, camera, artist, rating, other people's
+keywords — exactly as it found them.
+
+The other half is divergence. When something outside the stage's ownership does move,
+the checkpoint must say so, refuse to call itself verified, and change nothing back:
+a silent repair is how a library quietly loses the user's metadata.
+"""
+
+from __future__ import annotations
+
+import json
+import shutil
+import subprocess
+import uuid
+from datetime import datetime, timezone
+from types import SimpleNamespace
+
+import pytest
+
+from photo_pipeline.config import Config
+from photo_pipeline.db import create_db_engine, create_session_factory, run_migrations
+from photo_pipeline.integrations import exiftool
+from photo_pipeline.models import Asset, ExifProjection
+from photo_pipeline.services import exif_checkpoint
+from photo_pipeline.services.analysis import AnalysisService
+from photo_pipeline.services.safety import SafetyService
+from tests.fixtures.media_corpus import CASES_BY_ID, build_corpus
+
+pytestmark = pytest.mark.skipif(
+    shutil.which("exiftool") is None, reason="exiftool not installed"
+)
+
+
+@pytest.fixture
+def env(tmp_path):
+    data = tmp_path / "data"
+    data.mkdir()
+    lib = tmp_path / "lib"
+    files = build_corpus(
+        lib,
+        ids=(
+            "user_exif",
+            "prior_safety_keyword",
+            "prior_analysis_keywords",
+            "conflicting_safety_keywords",
+            "malformed_metadata",
+            "no_exif",
+        ),
+    )
+    config = Config.from_env(
+        {"PHOTO_PIPELINE_DATA_DIR": str(data), "PHOTO_PIPELINE_LIBRARY_ROOTS": str(lib)}
+    )
+    run_migrations(config.database_url)
+    engine = create_db_engine(config.database_url)
+    sf = create_session_factory(engine)
+    assets = {}
+    with sf() as session:
+        for case_id, path in files.items():
+            asset = Asset(
+                id=str(uuid.uuid4()),
+                original_path=str(path),
+                current_path=str(path),
+                discovered_at=datetime.now(timezone.utc),
+                hash_version=1,
+            )
+            session.add(asset)
+            assets[case_id] = asset.id
+        session.commit()
+    yield SimpleNamespace(config=config, lib=lib, sf=sf, files=files, assets=assets)
+    engine.dispose()
+
+
+class Provider:
+    """A vision provider whose tags are fixed, so the EXIF assertion is the test."""
+
+    def __init__(self, tags):
+        self.tags = list(tags)
+
+    def analyze(self, path, *, album_hint):
+        return {"description": "a photo", "tags": self.tags}
+
+
+def snapshot(path):
+    return exiftool.read_all(str(path))
+
+
+def stage_fields(before, after):
+    """Everything that changed except this application's owned and volatile tags."""
+    return exif_checkpoint.compare(before, after)
+
+
+# ── preservation ──────────────────────────────────────────────────────────────
+
+
+def test_a_safety_decision_preserves_every_user_field(env):
+    case = CASES_BY_ID["user_exif"]
+    path = env.files["user_exif"]
+    before = snapshot(path)
+    assert before, "the fixture must actually carry user metadata"
+
+    result = SafetyService(env.sf).decide(env.assets["user_exif"], "nsfw")
+    after = snapshot(path)
+
+    assert result["exif_verified"] is True
+    assert stage_fields(before, after) == ()
+    for field in case.preserved_fields:
+        assert after[field] == before[field], field
+    assert "nsfw" in exif_checkpoint.owned_values(after)
+    assert "sfw" not in exif_checkpoint.owned_values(after)
+
+
+def test_analysis_keywords_are_additive_and_keep_the_safety_decision(env):
+    """The two stages share the Keywords field; the second must merge, not replace."""
+    asset_id = env.assets["prior_analysis_keywords"]
+    path = env.files["prior_analysis_keywords"]
+    SafetyService(env.sf).decide(asset_id, "sfw")
+    before = snapshot(path)
+
+    AnalysisService(
+        env.sf, provider=Provider(["harbour", "boats"]), library_roots=(env.lib,)
+    ).run([asset_id])
+    after = snapshot(path)
+
+    keywords = exif_checkpoint.owned_values(after)
+    assert {"sfw", "beach", "sunset", "harbour", "boats"} <= keywords
+    assert "nsfw" not in keywords
+    assert stage_fields(before, after) == ()
+
+
+def test_flipping_a_safety_decision_removes_only_the_opposite_keyword(env):
+    asset_id = env.assets["prior_analysis_keywords"]
+    path = env.files["prior_analysis_keywords"]
+    safety = SafetyService(env.sf)
+    safety.decide(asset_id, "nsfw")
+    before = snapshot(path)
+
+    safety.decide(asset_id, "sfw")
+    after = snapshot(path)
+
+    keywords = exif_checkpoint.owned_values(after)
+    assert "sfw" in keywords and "nsfw" not in keywords
+    assert {"beach", "sunset"} <= keywords, "analysis keywords are not safety's to remove"
+    assert stage_fields(before, after) == ()
+
+
+def test_conflicting_safety_keywords_are_resolved_to_one(env):
+    asset_id = env.assets["conflicting_safety_keywords"]
+    SafetyService(env.sf).decide(asset_id, "sfw")
+    keywords = exif_checkpoint.owned_values(snapshot(env.files["conflicting_safety_keywords"]))
+    assert keywords & {"sfw", "nsfw"} == {"sfw"}
+
+
+def test_malformed_metadata_does_not_block_the_checkpoint(env):
+    result = SafetyService(env.sf).decide(env.assets["malformed_metadata"], "nsfw")
+    assert result["exif_verified"] is True
+
+
+# ── verification and the refreshed hash ───────────────────────────────────────
+
+
+def test_a_verified_checkpoint_refreshes_the_recorded_bytes(env):
+    """exiftool rewrites the container, so the stored SHA-256 must be the new one —
+    upload compares against exactly these bytes."""
+    asset_id = env.assets["no_exif"]
+    path = env.files["no_exif"]
+    SafetyService(env.sf).decide(asset_id, "sfw")
+
+    with env.sf() as session:
+        asset = session.get(Asset, asset_id)
+        row = session.get(ExifProjection, (asset_id, "safety"))
+    import hashlib
+
+    on_disk = hashlib.sha256(path.read_bytes()).hexdigest()
+    assert asset.current_sha256 == on_disk
+    assert row.state == "verified" and row.result_file_sha256 == on_disk
+    assert json.loads(row.desired_json) == {"add": ["sfw"], "remove": ["nsfw"]}
+
+
+def test_the_projection_survives_a_restart(env):
+    asset_id = env.assets["user_exif"]
+    SafetyService(env.sf).decide(asset_id, "nsfw")
+
+    engine = create_db_engine(env.config.database_url)  # a fresh connection, as a restart is
+    try:
+        with create_session_factory(engine)() as session:
+            row = session.get(ExifProjection, (asset_id, "safety"))
+            assert row.state == "verified" and row.verified_at is not None
+    finally:
+        engine.dispose()
+
+
+# ── divergence ────────────────────────────────────────────────────────────────
+
+
+def test_a_field_the_stage_does_not_own_changing_is_divergent(env, monkeypatch):
+    """Something rewrote the artist while the safety keyword was being written.
+
+    The stage must not call that verified, must record what moved, and must not put
+    the old value back — the file is now a question for a human.
+    """
+    asset_id = env.assets["user_exif"]
+    path = env.files["user_exif"]
+    real_apply = exiftool.apply_keywords
+
+    def sabotage(target, *, add=(), remove=()):
+        ok = real_apply(target, add=add, remove=remove)
+        subprocess.run(
+            ["exiftool", "-m", "-overwrite_original", "-Artist=Someone Else", str(target)],
+            capture_output=True,
+            check=False,
+        )
+        return ok
+
+    monkeypatch.setattr(exiftool, "apply_keywords", sabotage)
+    result = SafetyService(env.sf).decide(asset_id, "nsfw")
+
+    assert result["exif_verified"] is False, "a divergent checkpoint is not verified"
+    with env.sf() as session:
+        row = session.get(ExifProjection, (asset_id, "safety"))
+        asset = session.get(Asset, asset_id)
+    assert row.state == "divergent"
+    assert "EXIF:IFD0:Artist" in json.loads(row.divergent_fields)
+    assert row.verified_at is None
+    # Not repaired, and not silently accepted as the current verified bytes.
+    assert snapshot(path)["EXIF:IFD0:Artist"] == "Someone Else"
+    assert asset.current_sha256 is None
+
+
+def test_a_divergent_asset_shows_up_in_the_review_queue(env, monkeypatch):
+    asset_id = env.assets["user_exif"]
+    monkeypatch.setattr(
+        exif_checkpoint,
+        "run",
+        lambda *args, **kwargs: exif_checkpoint.CheckpointResult(
+            exif_checkpoint.DIVERGENT, changed_fields=("EXIF:IFD0:Artist",), sha256="abc"
+        ),
+    )
+    SafetyService(env.sf).decide(asset_id, "sfw")
+
+    rows = SafetyService(env.sf).review_queue()["items"]
+    row = next(item for item in rows if item["asset_id"] == asset_id)
+    assert row["exif_state"] == "divergent"
+    assert row["exif_verified"] is False
+
+
+def test_a_write_that_does_not_take_is_a_failure_not_a_verification(env, monkeypatch):
+    monkeypatch.setattr(exiftool, "apply_keywords", lambda *a, **k: False)
+    result = SafetyService(env.sf).decide(env.assets["no_exif"], "sfw")
+    assert result["exif_verified"] is False
+    assert exif_checkpoint.state_for(env.sf, env.assets["no_exif"], "safety") == "failed"
+
+
+def test_unreadable_metadata_is_a_failure_not_an_empty_snapshot(env, monkeypatch):
+    """``None`` from exiftool means "cannot answer"; treating it as "nothing there"
+    would make every field look preserved."""
+    monkeypatch.setattr(exiftool, "read_all", lambda path: None)
+    result = exif_checkpoint.run(str(env.files["no_exif"]), add=("sfw",))
+    assert result.state == "failed" and result.reason == "metadata_unreadable"
+
+
+def test_analysis_records_its_own_projection_separately(env):
+    asset_id = env.assets["user_exif"]
+    SafetyService(env.sf).decide(asset_id, "sfw")
+    AnalysisService(env.sf, provider=Provider(["pier"]), library_roots=(env.lib,)).run([asset_id])
+
+    assert exif_checkpoint.state_for(env.sf, asset_id, "safety") == "verified"
+    assert exif_checkpoint.state_for(env.sf, asset_id, "analysis") == "verified"
+
+
+# ── the comparison rules themselves ───────────────────────────────────────────
+
+
+def test_compare_ignores_owned_and_volatile_fields_only():
+    before = {
+        "EXIF:IFD0:Artist": "Ada",
+        "IPTC:Keywords": ["holiday"],
+        "File:System:FileSize": "3.8 kB",
+        "File:CurrentIPTCDigest": "aaa",
+        "XMP:XMP-x:XMPToolkit": "old",
+    }
+    after = {
+        "EXIF:IFD0:Artist": "Ada",
+        "IPTC:Keywords": ["holiday", "sfw"],
+        "File:System:FileSize": "3.9 kB",
+        "File:CurrentIPTCDigest": "bbb",
+        "XMP:XMP-x:XMPToolkit": "new",
+    }
+    assert exif_checkpoint.compare(before, after) == ()
+
+    after["EXIF:GPS:GPSLatitude"] = "48.1"  # an addition counts as much as a loss
+    del after["EXIF:IFD0:Artist"]
+    assert exif_checkpoint.compare(before, after) == ("EXIF:GPS:GPSLatitude", "EXIF:IFD0:Artist")
diff --git a/tests/integration/test_media_hardening.py b/tests/integration/test_media_hardening.py
new file mode 100644
index 0000000..22bc474
--- /dev/null
+++ b/tests/integration/test_media_hardening.py
@@ -0,0 +1,335 @@
+"""US07-03: the golden media corpus, driven through the real decode path.
+
+Every case in ``tests/fixtures/media_corpus.py`` is exercised here. The claims:
+
+* a supported format/orientation/profile renders, with the orientation applied and
+  the decode bounded to roughly the size that was asked for;
+* a damaged, empty, lying, or gigapixel file becomes one precise item error — the
+  scan still finishes, the other assets still render, and a job that meets one keeps
+  running;
+* a failure leaves no temporary file behind, and startup removes only the temporaries
+  this service recognises;
+* cache invalidation follows the pixels, not the metadata.
+"""
+
+from __future__ import annotations
+
+import hashlib
+import shutil
+import time
+import tracemalloc
+from types import SimpleNamespace
+
+import pytest
+from PIL import Image
+
+from photo_pipeline import imaging
+from photo_pipeline.config import Config
+from photo_pipeline.db import create_db_engine, create_session_factory, run_migrations
+from photo_pipeline.models import Asset, Thumbnail
+from photo_pipeline.services.inventory import InventoryService
+from photo_pipeline.services.thumbnails import (
+    TEMP_SUFFIX,
+    ThumbnailError,
+    ThumbnailService,
+)
+from tests.fixtures.media_corpus import CASES, CASES_BY_ID, build_corpus
+
+RENDERABLE = [case for case in CASES if case.expect_error is None]
+DAMAGED = [case for case in CASES if case.expect_error is not None]
+
+needs_exiftool = pytest.mark.skipif(
+    shutil.which("exiftool") is None, reason="exiftool not installed"
+)
+
+
+@pytest.fixture(scope="module")
+def corpus(tmp_path_factory):
+    root = tmp_path_factory.mktemp("corpus")
+    return SimpleNamespace(root=root, files=build_corpus(root))
+
+
+@pytest.fixture
+def env(tmp_path, corpus):
+    """A library holding the whole corpus, scanned into a fresh database."""
+    data = tmp_path / "data"
+    data.mkdir()
+    lib = tmp_path / "lib"
+    shutil.copytree(corpus.root, lib)
+    config = Config.from_env(
+        {"PHOTO_PIPELINE_DATA_DIR": str(data), "PHOTO_PIPELINE_LIBRARY_ROOTS": str(lib)}
+    )
+    run_migrations(config.database_url)
+    engine = create_db_engine(config.database_url)
+    sf = create_session_factory(engine)
+    scan = InventoryService(sf).scan(lib)
+    by_id = {
+        case.id: next(
+            (aid for path, aid in scan.asset_ids.items() if path.endswith(case.filename)), None
+        )
+        for case in CASES
+    }
+    yield SimpleNamespace(
+        config=config, lib=lib, sf=sf, scan=scan, assets=by_id,
+        thumbs=ThumbnailService(sf, config),
+    )
+    engine.dispose()
+
+
+# ── the manifest itself ───────────────────────────────────────────────────────
+
+
+def test_the_manifest_is_internally_consistent():
+    assert len({case.id for case in CASES}) == len(CASES), "logical ids must be unique"
+    assert len({case.filename for case in CASES}) == len(CASES), "paths must be unique"
+    for case in CASES:
+        assert case.kind in {"format", "orientation", "profile", "damaged", "metadata"}
+        assert case.expect_error in (None, "unsupported_image", "image_too_large")
+
+
+def test_the_corpus_regenerates_byte_for_byte(tmp_path):
+    """A golden corpus that drifts between runs cannot be a golden corpus."""
+    first = build_corpus(tmp_path / "one")
+    second = build_corpus(tmp_path / "two")
+    digests = {
+        case_id: (
+            hashlib.sha256(first[case_id].read_bytes()).hexdigest(),
+            hashlib.sha256(second[case_id].read_bytes()).hexdigest(),
+        )
+        for case_id in first
+    }
+    drifted = [case_id for case_id, (a, b) in digests.items() if a != b]
+    assert not drifted, f"non-deterministic fixtures: {drifted}"
+
+
+# ── bounded decode ────────────────────────────────────────────────────────────
+
+
+@pytest.mark.parametrize("case", RENDERABLE, ids=lambda case: case.id)
+def test_every_supported_case_renders(case, env):
+    asset_id = env.assets[case.id]
+    assert asset_id, f"{case.id} was not discovered by the scan"
+    path = env.thumbs.generate(asset_id, 256)
+    with Image.open(path) as thumb:
+        assert thumb.format == "WEBP"
+        assert max(thumb.size) <= 256
+        assert min(thumb.size) >= 1
+
+
+@pytest.mark.parametrize("orientation", range(1, 9))
+def test_exif_orientation_is_applied_before_resizing(orientation, env):
+    """All eight tags: the 90° ones must come out portrait from a landscape source."""
+    asset_id = env.assets[f"orientation_{orientation}"]
+    with Image.open(env.thumbs.generate(asset_id, 256)) as thumb:
+        rotated = orientation in (5, 6, 7, 8)
+        assert (thumb.height > thumb.width) is rotated
+
+
+def test_transparency_and_grayscale_survive_the_pipeline(env):
+    with Image.open(env.thumbs.generate(env.assets["png_alpha"], 256)) as thumb:
+        assert "A" in thumb.getbands()
+    with Image.open(env.thumbs.generate(env.assets["grayscale"], 256)) as thumb:
+        assert thumb.size == (100, 100)  # smaller than the request: never upscaled
+
+
+def test_a_broken_colour_profile_still_produces_a_picture(env):
+    """An unusable ICC profile is a metadata problem, not a reason to lose the preview."""
+    for case_id in ("icc_tagged", "broken_icc"):
+        with Image.open(env.thumbs.generate(env.assets[case_id], 256)) as thumb:
+            assert thumb.size[0] > 0
+
+
+def test_a_large_jpeg_is_not_decoded_at_full_resolution(env):
+    """12 megapixels would be ~36 MB of pixels; the draft decode keeps it far below."""
+    tracemalloc.start()
+    try:
+        env.thumbs.generate(env.assets["large_jpeg"], 256)
+        _, peak = tracemalloc.get_traced_memory()
+    finally:
+        tracemalloc.stop()
+    assert peak < 12_000_000, f"decode peaked at {peak} bytes"
+
+
+# ── damaged inputs ────────────────────────────────────────────────────────────
+
+
+@pytest.mark.parametrize("case", DAMAGED, ids=lambda case: case.id)
+def test_every_damaged_case_is_a_precise_item_error(case, env):
+    asset_id = env.assets[case.id]
+    assert asset_id, f"{case.id} was not discovered by the scan"
+    with pytest.raises(ThumbnailError) as raised:
+        env.thumbs.generate(asset_id, 256)
+    assert raised.value.code == case.expect_error
+    # Persisted, so a broken original is not re-decoded on every request...
+    with env.sf() as session:
+        rows = [
+            row
+            for row in session.query(Thumbnail).all()
+            if row.asset_id == asset_id and row.state == "error"
+        ]
+    assert rows and rows[0].error_code == case.expect_error
+    # ...and the cached failure is the same precise error, not a generic one.
+    with pytest.raises(ThumbnailError) as again:
+        env.thumbs.generate(asset_id, 256)
+    assert again.value.code == case.expect_error
+
+
+def test_a_gigapixel_header_is_refused_quickly_and_cheaply(env):
+    """The refusal must come from the declared size, not from decoding it."""
+    tracemalloc.start()
+    started = time.monotonic()
+    try:
+        with pytest.raises(ThumbnailError):
+            env.thumbs.generate(env.assets["bomb_header"], 1280)
+        _, peak = tracemalloc.get_traced_memory()
+    finally:
+        tracemalloc.stop()
+    assert time.monotonic() - started < 5
+    assert peak < 5_000_000, f"a 3.6 gigapixel header allocated {peak} bytes"
+
+
+def test_one_broken_file_stops_neither_the_scan_nor_its_neighbours(env):
+    """The whole corpus is in the library, damaged files included."""
+    assert len(env.scan.asset_ids) >= len(CASES) - 1  # the empty file has no pixels to hash
+    for case in RENDERABLE[:5]:
+        assert env.thumbs.generate(env.assets[case.id], 256).exists()
+
+
+def test_a_job_that_meets_a_broken_file_gets_evidence_not_an_exception(env):
+    """``ensure_protected`` is what the archive lane calls; a plan must not die on
+    one unreadable original."""
+    evidence = env.thumbs.ensure_protected(env.assets["corrupt_png"])
+    assert evidence["state"] == "unsupported"
+    assert evidence["error_code"] == "unsupported_image"
+    assert env.thumbs.ensure_protected(env.assets["jpeg"])["state"] == "ready"
+
+
+def test_undecodable_files_do_not_break_hashing(env):
+    """Discovery records what it can: byte identity always, pixel identity when the
+    file has pixels."""
+    with env.sf() as session:
+        assets = {a.current_path: a for a in session.query(Asset).all()}
+    for case in DAMAGED:
+        asset = next((a for path, a in assets.items() if path.endswith(case.filename)), None)
+        if asset is None:
+            continue
+        assert asset.current_sha256, "byte identity is always available"
+        assert asset.pixel_sha256 is None, "undecodable files must not invent pixel identity"
+
+
+# ── cache lifecycle ───────────────────────────────────────────────────────────
+
+
+def test_a_failed_render_leaves_no_temporary_behind(env):
+    with pytest.raises(ThumbnailError):
+        env.thumbs.generate(env.assets["truncated_jpeg"], 256)
+    leftovers = list(env.config.thumbnail_cache_dir.rglob(f"*{TEMP_SUFFIX}"))
+    assert leftovers == []
+
+
+def test_startup_cleanup_removes_only_recognised_temporaries(env):
+    cache = env.config.thumbnail_cache_dir
+    env.thumbs.generate(env.assets["jpeg"], 256)  # creates the cache directory
+    stale = cache / f".abc123{TEMP_SUFFIX}"
+    stale.write_bytes(b"half a thumbnail")
+    innocent = cache / "keep-me.webp"
+    innocent.write_bytes(b"not a temporary")
+
+    assert env.thumbs.cleanup_temp_files() == 1
+    assert not stale.exists()
+    assert innocent.exists()
+    assert list(cache.rglob("*.webp")), "real cache entries survive"
+
+
+def test_metadata_only_change_reuses_the_thumbnail_and_a_pixel_change_does_not(env):
+    """The cache key follows the pixels: an EXIF edit must not cost a re-render."""
+    asset_id = env.assets["jpeg"]
+    first = env.thumbs.generate(asset_id, 256)
+
+    with env.sf() as session:  # what a rescan records after an EXIF-only edit
+        asset = session.get(Asset, asset_id)
+        asset.current_sha256 = "different-bytes"
+        session.commit()
+    assert env.thumbs.generate(asset_id, 256) == first
+
+    with env.sf() as session:  # a genuine pixel change
+        asset = session.get(Asset, asset_id)
+        asset.pixel_sha256 = "different-pixels"
+        session.commit()
+    assert env.thumbs.generate(asset_id, 256) != first
+
+
+def test_a_deleted_cache_file_is_regenerated(env):
+    asset_id = env.assets["png"]
+    path = env.thumbs.generate(asset_id, 256)
+    path.unlink()
+    regenerated = env.thumbs.generate(asset_id, 256)
+    assert regenerated == path and regenerated.exists()
+
+
+# ── the imaging door itself ───────────────────────────────────────────────────
+
+
+def test_open_image_translates_every_decoder_failure(corpus):
+    """Callers must be able to catch two typed errors, never bare ``Exception``."""
+    for case in DAMAGED:
+        expected = imaging.ImageTooLarge if case.expect_error == "image_too_large" else imaging.UndecodableImage
+        with pytest.raises(expected):
+            with imaging.open_image(corpus.files[case.id]) as image:
+                image.load()
+
+
+def test_open_image_refuses_more_pixels_than_the_caller_allowed(corpus):
+    with pytest.raises(imaging.ImageTooLarge):
+        with imaging.open_image(corpus.files["jpeg"], max_pixels=100):
+            pass
+    with imaging.open_image(corpus.files["jpeg"], max_pixels=100_000) as image:
+        assert image.size == (320, 240)
+
+
+@needs_exiftool
+def test_malformed_metadata_does_not_stop_the_picture(env):
+    """A broken EXIF block is a metadata fact, not a decode failure."""
+    assert env.thumbs.generate(env.assets["malformed_metadata"], 256).exists()
+    with env.sf() as session:
+        asset = session.get(Asset, env.assets["malformed_metadata"])
+    assert asset.pixel_sha256, "pixels are still identifiable"
+
+
+def test_scoring_never_relaxes_truncated_image_handling(corpus):
+    """The donor's process-global ``LOAD_TRUNCATED_IMAGES`` is gone for good.
+
+    It is global state: switching it on for the safety model would also switch it on
+    for hashing and preview rendering in the same process, and half a file would
+    silently become a valid picture (donor_ledger: nt-score-model).
+    """
+    import inspect
+
+    from PIL import ImageFile
+
+    from photo_pipeline.integrations import nsfw_model
+
+    code = [
+        line
+        for line in inspect.getsource(nsfw_model).splitlines()
+        if not line.strip().startswith("#")
+    ]
+    assert not any("LOAD_TRUNCATED_IMAGES" in line for line in code)
+
+    # And the door holds even when something else in the process turned it on —
+    # the frozen donor does exactly that when the characterization suite imports it.
+    previous = ImageFile.LOAD_TRUNCATED_IMAGES
+    ImageFile.LOAD_TRUNCATED_IMAGES = True
+    try:
+        with pytest.raises(imaging.UndecodableImage):
+            with imaging.open_image(corpus.files["truncated_jpeg"]) as image:
+                image.load()
+        assert ImageFile.LOAD_TRUNCATED_IMAGES is True, "the caller's setting is restored"
+    finally:
+        ImageFile.LOAD_TRUNCATED_IMAGES = previous
+
+
+def test_every_manifest_case_is_exercised():
+    """The corpus lint: no fixture may sit in the manifest untested."""
+    covered = {case.id for case in RENDERABLE} | {case.id for case in DAMAGED}
+    assert covered == set(CASES_BY_ID)
diff --git a/tests/story_traceability.json b/tests/story_traceability.json
index 8a1c3d3..cd1f570 100644
--- a/tests/story_traceability.json
+++ b/tests/story_traceability.json
@@ -152,6 +152,10 @@
       "tests/unit/test_security_policy.py",
       "tests/integration/test_security_boundaries.py",
       "tests/e2e/test_security.py"
+    ],
+    "US07-03": [
+      "tests/integration/test_media_hardening.py",
+      "tests/integration/test_exif_checkpoints.py"
     ]
   }
 }