US08-01: Make the Trust Boundary Configurable and Authenticated

.gitignore

@@ -20,3 +20,12 @@ _IGNORE/
 
 # Test failure evidence (US07-04)
 .artifacts/
+
+# Any dotenv, not only the default name.
+*.env
+
+# Local virtualenv for running the app.
+.venv/
+
+# setuptools editable-install metadata.
+*.egg-info/

README.md

@@ -7,16 +7,49 @@ archive workflow. Planning lives in `INTEGRATED_PIPELINE_CONCEPT.md` and
 ## Application (`photo_pipeline`)
 
 The target application lives in `photo_pipeline/` (FastAPI + SQLAlchemy + Alembic).
-Run it with:
+Install it into a virtualenv once:
 
 ```bash
-python -m photo_pipeline migrate   # apply database migrations
-python -m photo_pipeline serve     # start the API + static review UI (127.0.0.1:8000)
+python3.12 -m venv .venv
+.venv/bin/pip install -e ".[vision]"   # drop [vision] for a review-only install
 ```
 
+Then run the two processes:
+
+```bash
+.venv/bin/python -m photo_pipeline migrate   # apply database migrations
+.venv/bin/python -m photo_pipeline serve     # API + review UI at 127.0.0.1:8000/app/
+.venv/bin/python -m photo_pipeline worker    # second terminal: runs the jobs
+```
+
+The server enqueues work and serves the UI; nothing actually scans, scores,
+analyses, uploads, or archives without a worker. `work_item/scripts/python` is the
+*helper's* launcher — it prefers Conda base and falls back to a bare system
+interpreter, so it is not how the application is run.
+
 Configuration comes from `PHOTO_PIPELINE_*` environment variables (see
 `photo_pipeline/config.py`); secrets are referenced, never logged.
 
+### Configuration file
+
+`.env` in the working directory is read at startup, or any path named by
+`PHOTO_PIPELINE_ENV_FILE`. It is parsed, never executed: `KEY=value` lines,
+`#` comments, optional quotes — no interpolation and no `export`. **Anything already
+exported wins**, so the file is the standing configuration and the shell is the
+override for one run.
+
+The archived CLI's variable names still work, so an existing `photo_analyzer.env`
+can be used as-is:
+
+| in the file | applied as |
+|---|---|
+| `LLM_API_KEY` / `GEMINI_API_KEY` | `OPENAI_API_KEY` |
+| `LLM_BASE_URL` | `OPENAI_BASE_URL` |
+| `LIBRARY` | `PHOTO_PIPELINE_LIBRARY_ROOTS` |
+
+`.env` and `*.env` are gitignored and denied by the work-item safety checks: the
+file holds a real key and must never be committed.
+
 ### API access (US07-02)
 
 The app listens on loopback, so its attacker is another page in the same browser.
@@ -36,6 +69,34 @@ not a loopback name (DNS rebinding), when `Origin` is any other origin, when
 thumbnail), or when the body exceeds `PHOTO_PIPELINE_MAX_REQUEST_BYTES`. There is no
 CORS middleware at all, so no other origin can read a response.
 
+### Reaching it through a hostname or proxy (US08-01)
+
+| variable | meaning |
+|---|---|
+| `PHOTO_PIPELINE_ALLOWED_HOSTS` | comma-separated extra names the app answers to; empty means loopback only |
+| `PHOTO_PIPELINE_ACCESS_SECRET` | traded for the session cookie at `GET /api/v1/session` via `X-Access-Secret` |
+| `PHOTO_PIPELINE_TRUSTED_PROXIES` | comma-separated peer addresses whose `X-Forwarded-Proto`/`X-Forwarded-Host` are believed |
+
+Being reachable *was* the authentication: whoever could open `127.0.0.1:8000` owned
+the library. So naming any non-loopback host — or binding to one, `0.0.0.0` included
+— makes the access secret mandatory, and `serve` refuses to start without it rather
+than publishing the library. Loopback-only deployments need no secret and behave
+exactly as before.
+
+```bash
+curl -sc /tmp/pp.jar -H "X-Access-Secret: $PHOTO_PIPELINE_ACCESS_SECRET" \
+  https://photos.example.com/api/v1/session
+```
+
+The browser asks for the secret once per tab and keeps it in `sessionStorage`.
+Wrong secrets are rate-limited (5 per minute) and logged with the caller's address
+only. `Host` and `Origin` are judged against the configured names; the *external*
+scheme and host come from the forwarded headers only when the request arrived from a
+`PHOTO_PIPELINE_TRUSTED_PROXIES` address, so a client cannot declare its own origin,
+and the session cookie is marked `Secure` when that external scheme is HTTPS. Health
+endpoints stay reachable without the secret so an orchestrator can restart the
+container; nothing else does.
+
 ## Testing
 
 One offline command runs the whole suite (unit, integration, and browser
@@ -281,6 +342,51 @@ file in the temporary library are copied to `.artifacts/<test id>/` before pytes
 deletes the directory. Point `PHOTO_PIPELINE_TEST_ARTIFACTS` elsewhere to collect
 them from CI.
 
+## Release gate (US07-07)
+
+One command runs every suite in an isolated stack and keeps the evidence:
+
+```bash
+work_item/scripts/python -m photo_pipeline release-gate --output data/release/$(date -u +%Y%m%dT%H%M%SZ)
+```
+
+It fails — and exits non-zero — when any stage fails, when a suite skips a test for
+a reason that is not a documented environment limit (`exiftool not installed`,
+`root ignores directory permissions`), or when the story matrix has a hole. The
+evidence directory holds `release-report.json` (revision, per-stage result, timings,
+summaries), `logs/<stage>.log`, and `CHECKSUMS.sha256` over both.
+
+**The story matrix** lives in `tests/story_traceability.json`: every story under
+`delivery_backlog/stories/` is either mapped to test files that exist, or listed in
+`planned` as an accepted but unimplemented story. A story that is neither, or a
+mapping to a file that has been deleted, fails the gate.
+
+**The journey** (`tests/e2e/test_release_journey.py`) takes one fresh library through
+discovery, duplicate review, safety, analysis, EXIF verification, album proposal,
+guarded rename, rescan, upload with server-side verification, archive, offline
+deduplication, and restore — over HTTP against real server and worker processes,
+with a full restart in the middle and at the end.
+
+### Real-library dry run and approval
+
+Before the application is pointed at photos that cannot be replaced:
+
+```bash
+work_item/scripts/python -m photo_pipeline dry-run --output dry-run.json
+work_item/scripts/python -m photo_pipeline approve-dry-run dry-run.json --approver "$(whoami)"
+```
+
+The dry run is strictly read-only: it opens no file for writing, writes no database
+row, and reports what it found — file counts by extension, folders, bytes, unreadable
+files, excluded directories, and a reconciliation against what the database already
+knows (already registered, new, recorded but absent). Set
+`PHOTO_PIPELINE_REQUIRE_DRY_RUN_APPROVAL=1` and **every mutating API request is
+refused with `403 dry_run_not_approved`** until a report for exactly those library
+roots has been approved. Reading stays open — you have to be able to see what was
+found in order to approve it — and so does taking a backup. Change the library roots
+and the approval no longer applies: it approves that reconciliation, not the idea of
+mutating.
+
 ## Performance budgets (US07-06)
 
 Budgets are measured, not asserted in prose. `python -m photo_pipeline benchmark`

frontend/js/api.js

@@ -7,9 +7,28 @@ export const BASE = "/api/v1";
 // what makes it proof that the caller is this app and not another page.
 let csrfToken = null;
 
+// A deployment reachable through a proxy trades an operator secret for that cookie.
+// Kept per tab: sessionStorage dies with the tab, and the secret never enters a URL.
+const SECRET_KEY = "pp_access_secret";
+
+async function bootstrap(secret) {
+  return fetch(BASE + "/session", {
+    credentials: "same-origin",
+    headers: secret ? { "X-Access-Secret": secret } : {},
+  });
+}
+
 async function session() {
   if (csrfToken === null) {
-    const response = await fetch(BASE + "/session", { credentials: "same-origin" });
+    let response = await bootstrap(sessionStorage.getItem(SECRET_KEY));
+    if (response.status === 401) {
+      sessionStorage.removeItem(SECRET_KEY);
+      const secret = prompt("Access secret");
+      if (secret) {
+        response = await bootstrap(secret);
+        if (response.ok) sessionStorage.setItem(SECRET_KEY, secret);
+      }
+    }
     const body = await response.json().catch(() => null);
     csrfToken = (body && body.csrf_token) || null;
   }

photo_pipeline/__main__.py

@@ -11,6 +11,7 @@ from __future__ import annotations
 
 import argparse
 import json
+from pathlib import Path
 from typing import Sequence
 
 from photo_pipeline.config import Config
@@ -63,6 +64,20 @@ def main(argv: Sequence[str] | None = None) -> int:
     )
     bench_cmd.add_argument("--output", help="Write the JSON report here as well as to stdout")
 
+    gate_cmd = commands.add_parser(
+        "release-gate", help="Run every suite in an isolated stack and keep the evidence"
+    )
+    gate_cmd.add_argument("--output", help="Evidence directory (default: data/release/<stamp>)")
+    dry_cmd = commands.add_parser(
+        "dry-run", help="Read-only reconciliation of the configured library (US07-07)"
+    )
+    dry_cmd.add_argument("--output", help="Write the report here as well as to stdout")
+    approve_cmd = commands.add_parser(
+        "approve-dry-run", help="Approve a dry-run report, which is what enables mutation"
+    )
+    approve_cmd.add_argument("report", help="Path to the dry-run report")
+    approve_cmd.add_argument("--approver", required=True, help="Who is accepting this")
+
     args = parser.parse_args(argv)
 
     config = Config.from_env()
@@ -115,6 +130,41 @@ def main(argv: Sequence[str] | None = None) -> int:
         # anyone reading the JSON.
         return 0 if report["ok"] else 1
 
+    if args.command == "release-gate":
+        from photo_pipeline.services import release
+
+        report = release.run_gate(config, output=args.output)
+        print(
+            json.dumps(
+                {k: v for k, v in report.items() if k not in ("stages", "matrix")}, indent=2
+            )
+        )
+        return 0 if report["ok"] else 1
+
+    if args.command == "dry-run":
+        from photo_pipeline.services import release
+
+        try:
+            report = release.dry_run(config)
+        except release.ReleaseError as error:
+            print(str(error))
+            return 1
+        if args.output:
+            Path(args.output).write_text(json.dumps(report, indent=2))
+        print(json.dumps(report, indent=2))
+        return 0
+
+    if args.command == "approve-dry-run":
+        from photo_pipeline.services import release
+
+        try:
+            record = release.approve(config, args.report, approver=args.approver)
+        except (release.ReleaseError, OSError, ValueError) as error:
+            print(str(error))
+            return 1
+        print(json.dumps(record, indent=2))
+        return 0
+
     if args.command == "diagnostics":
         from photo_pipeline.services import diagnostics
 
@@ -159,15 +209,25 @@ def main(argv: Sequence[str] | None = None) -> int:
             lock.release()
         return 0
 
+    import sys
+
     import uvicorn
 
-    from photo_pipeline.api.app import create_app
+    from photo_pipeline.api.app import ConfigurationRefused, create_app
+
+    # An exposed deployment without an access secret must not reach the port at all,
+    # and the operator needs a sentence, not a traceback (US08-01).
+    try:
+        app = create_app(config)
+    except ConfigurationRefused as error:
+        print(str(error), file=sys.stderr)
+        return 4
 
     lock = LibraryLock(config, "api")
     if (held := _acquire(lock, allow_legacy=args.allow_legacy)) is not None:
         return held
     try:
-        uvicorn.run(create_app(config), host=config.host, port=config.port)
+        uvicorn.run(app, host=config.host, port=config.port)
     finally:
         lock.release()
     return 0

photo_pipeline/api/app.py

@@ -35,7 +35,13 @@ from photo_pipeline.api.routes import (
     uploads,
     workflow,
 )
-from photo_pipeline.api.security import DEFAULT_HEADERS, SecurityMiddleware, Session
+from photo_pipeline.api.security import (
+    DEFAULT_HEADERS,
+    FailureLimiter,
+    SecurityMiddleware,
+    Session,
+    trust_refusal,
+)
 
 # Registers the safety_score / analysis job handlers on import.
 import photo_pipeline.jobs.domain_handlers  # noqa: F401
@@ -84,9 +90,16 @@ def _install_error_handlers(app: FastAPI) -> None:
         return _envelope(500, "internal_error", "internal error")
 
 
+class ConfigurationRefused(RuntimeError):
+    """The configuration would serve the library to callers it cannot authenticate."""
+
+
 def create_app(config: Config | None = None) -> FastAPI:
     config = config or Config.from_env()
     configure_logging(config.log_level, config.log_format)
+    # Before anything is built, let alone bound to a port (US08-01).
+    if (why := trust_refusal(config)) is not None:
+        raise ConfigurationRefused(why)
 
     @asynccontextmanager
     async def lifespan(app: FastAPI):
@@ -114,6 +127,10 @@ def create_app(config: Config | None = None) -> FastAPI:
     # One session per process: the browser exchanges it for a cookie + CSRF token,
     # and every other origin is refused before a route ever runs (US07-02).
     app.state.session = Session.create()
+    app.state.access_limiter = FailureLimiter()
+    # Also set in the lifespan, but the bootstrap route reads it, and a caller can
+    # arrive before anything else has touched app.state.
+    app.state.config = config
     app.add_middleware(SecurityMiddleware, session=app.state.session, config=config)
     _install_error_handlers(app)
     app.include_router(session_routes.router, prefix="/api/v1")

photo_pipeline/api/routes/session.py

@@ -3,21 +3,52 @@
 It sets the ``HttpOnly``/``SameSite=Strict`` session cookie and returns the CSRF
 token in the body. A foreign page can call this — it just cannot read the answer,
 because the app sends no CORS headers — and the cookie it received is never attached
-to a request that foreign page initiates.
+to a request that foreign page initiated.
+
+When an access secret is configured (mandatory as soon as the app is reachable from
+another machine, US08-01) this is also the authentication gate: the secret buys the
+cookie, and every route behind it keeps asking for exactly the session and CSRF token
+it asked for before. Wrong secrets are counted, and a burst of them stops being
+answered — otherwise a proxy-exposed deployment could be guessed at indefinitely.
 """
 
 from __future__ import annotations
 
+import logging
+import secrets
+
 from fastapi import APIRouter, Request
 from fastapi.responses import JSONResponse
 
-from photo_pipeline.api.security import SESSION_COOKIE
+from photo_pipeline.api.security import ACCESS_SECRET_HEADER, SESSION_COOKIE
 
 router = APIRouter(tags=["session"])
 
+log = logging.getLogger(__name__)
+
+
+def _refuse(status: int, code: str, message: str) -> JSONResponse:
+    return JSONResponse(status_code=status, content={"error": {"code": code, "message": message}})
+
 
 @router.get("/session")
 def start_session(request: Request) -> JSONResponse:
+    config = request.app.state.config
+    secret = config.access_secret
+    if secret is not None:
+        limiter = request.app.state.access_limiter
+        if limiter.blocked():
+            return _refuse(429, "too_many_attempts", "too many failed attempts; retry later")
+        offered = request.headers.get(ACCESS_SECRET_HEADER, "")
+        if not secrets.compare_digest(offered, secret.get_secret_value()):
+            limiter.record_failure()
+            # The client address is the whole record: the offered secret, the issued
+            # session, and the request body all stay out of the log.
+            log.warning(
+                "access secret rejected", extra={"client": _client(request), "path": "/session"}
+            )
+            return _refuse(401, "access_denied", "a valid access secret is required")
+
     session = request.app.state.session
     response = JSONResponse({"csrf_token": session.csrf_token})
     response.set_cookie(
@@ -25,6 +56,13 @@ def start_session(request: Request) -> JSONResponse:
         session.id,
         httponly=True,
         samesite="strict",
+        # HTTPS outside means the cookie must never travel over a plain hop, even one
+        # this process cannot see. Loopback http keeps working unchanged.
+        secure=request.scope.get("state", {}).get("external_scheme") == "https",
         path="/",
     )
     return response
+
+
+def _client(request: Request) -> str:
+    return request.client.host if request.client else "unknown"

photo_pipeline/api/security.py

@@ -19,6 +19,14 @@ The defenses stack, because each one alone has a hole:
   only in the bootstrap response body, which a foreign page cannot read (no CORS) —
   so possessing it proves the caller is same-origin.
 
+Behind a reverse proxy (US08-01) the same stack holds with two substitutions: the
+allowed host set comes from configuration instead of being the loopback names, and
+the host/scheme the policy judges is the *external* one, which is only read from
+``X-Forwarded-*`` when the request actually arrived from a configured proxy. The
+loopback check was standing in for authentication, so naming a non-loopback host
+also makes an access secret mandatory — ``trust_refusal`` refuses to start without
+one, and the secret is what the bootstrap endpoint trades for the session cookie.
+
 ``evaluate`` is a pure function over the request metadata: the whole policy is one
 table that a unit test can enumerate, and the middleware only applies its verdict.
 """
@@ -26,6 +34,7 @@ table that a unit test can enumerate, and the middleware only applies its verdic
 from __future__ import annotations
 
 import secrets
+import time
 from collections.abc import Mapping
 from dataclasses import dataclass
 from urllib.parse import urlsplit
@@ -35,6 +44,7 @@ from starlette.responses import JSONResponse
 
 SESSION_COOKIE = "pp_session"
 CSRF_HEADER = "x-csrf-token"
+ACCESS_SECRET_HEADER = "x-access-secret"
 API_PREFIX = "/api/v1"
 SAFE_METHODS = frozenset({"GET", "HEAD", "OPTIONS"})
 # Reachable without a session: liveness/readiness (an orchestrator has no cookie)
@@ -43,6 +53,9 @@ PUBLIC_PATHS = frozenset(
     {f"{API_PREFIX}/health/live", f"{API_PREFIX}/health/ready", f"{API_PREFIX}/session"}
 )
 LOOPBACK_HOSTS = frozenset({"127.0.0.1", "localhost", "::1", "[::1]"})
+# Mutating endpoints that must stay reachable while mutation itself is gated: the
+# backup a careful operator takes first, and its retention (US07-07).
+MUTATION_EXEMPT_PATHS = frozenset({f"{API_PREFIX}/backups", f"{API_PREFIX}/backups/prune"})
 
 # Applied to every response. No inline script/style is used by the frontend, so the
 # policy can stay strict; `frame-ancestors 'none'` and CORP keep other pages from
@@ -93,6 +106,28 @@ def split_host(value: str) -> tuple[str, str]:
     return host, port
 
 
+def external_view(
+    *,
+    client: str | None,
+    headers: Mapping[str, str],
+    scheme: str,
+    trusted_proxies: frozenset[str],
+) -> tuple[str, str]:
+    """The ``(scheme, host)`` the caller used, as opposed to the one this hop saw.
+
+    Forwarded headers are a client-supplied claim. Believing them from anyone lets a
+    request declare its own origin — and origin is half of this module's evidence —
+    so they count only when the connection came from a configured proxy.
+    """
+    host = headers.get("host", "")
+    if client is None or client not in trusted_proxies:
+        return scheme, host
+    # A chain appends: the first entry is what the original client asked for.
+    forwarded_proto = headers.get("x-forwarded-proto", "").split(",")[0].strip().lower()
+    forwarded_host = headers.get("x-forwarded-host", "").split(",")[0].strip()
+    return forwarded_proto or scheme, forwarded_host or host
+
+
 def evaluate(
     *,
     method: str,
@@ -100,20 +135,26 @@ def evaluate(
     headers: Mapping[str, str],
     session: Session,
     allowed_hosts: frozenset[str] = LOOPBACK_HOSTS,
+    scheme: str = "http",
     max_request_bytes: int,
 ) -> Refusal | None:
-    """Why this request must be refused, or ``None`` when it may proceed."""
+    """Why this request must be refused, or ``None`` when it may proceed.
+
+    ``headers["host"]`` and ``scheme`` are the external ones (see ``external_view``);
+    the allowed origins are the allowed hosts under that scheme and port, so there is
+    no second list that can drift away from the first.
+    """
     host_header = headers.get("host", "")
     host, port = split_host(host_header)
     if host.lower() not in allowed_hosts:
-        return Refusal(403, "host_not_allowed", "request host is not a local address")
+        return Refusal(403, "host_not_allowed", "request host is not an allowed address")
 
     origin = headers.get("origin")
     if origin is not None and origin != "":
         parts = urlsplit(origin)
         origin_host, origin_port = split_host(parts.netloc)
         if (
-            parts.scheme not in ("http", "https")
+            parts.scheme != scheme
             or origin_host.lower() not in allowed_hosts
             or origin_port != port
         ):
@@ -139,14 +180,66 @@ def evaluate(
     return None
 
 
+def exposed_hosts(config) -> list[str]:
+    """Configured names by which this application is reachable from another machine."""
+    names = {str(config.host).lower()}
+    names.update(split_host(name)[0].lower() for name in config.allowed_hosts)
+    return sorted(names - LOOPBACK_HOSTS)
+
+
+def trust_refusal(config) -> str | None:
+    """Why this configuration must not serve at all, or ``None``.
+
+    Reaching the app used to prove ownership of it. The moment a configuration makes
+    it reachable from elsewhere that stops being true, so serving without a secret
+    would publish the library — refuse at startup rather than at the first request,
+    when the operator is no longer watching (US08-01).
+    """
+    exposed = exposed_hosts(config)
+    if exposed and config.access_secret is None:
+        return (
+            f"refusing to serve: {', '.join(exposed)} is reachable from outside this "
+            "machine, so PHOTO_PIPELINE_ACCESS_SECRET must be set"
+        )
+    return None
+
+
+class FailureLimiter:
+    """Bounded failed access-secret attempts, so the secret cannot be guessed online.
+
+    ponytail: one counter for the whole process rather than per client address —
+    behind a proxy every attempt arrives from the same address anyway. Per-caller
+    buckets if the app is ever exposed without one.
+    """
+
+    def __init__(self, limit: int = 5, window: float = 60.0) -> None:
+        self.limit = limit
+        self.window = window
+        self._failures: list[float] = []
+
+    def blocked(self) -> bool:
+        now = time.monotonic()
+        self._failures = [at for at in self._failures if now - at < self.window]
+        return len(self._failures) >= self.limit
+
+    def record_failure(self) -> None:
+        self._failures.append(time.monotonic())
+
+
 class SecurityMiddleware:
     """Pure-ASGI so the SSE stream keeps streaming (BaseHTTPMiddleware buffers)."""
 
     def __init__(self, app, *, session: Session, config) -> None:
         self.app = app
         self.session = session
+        self.config = config
         self.max_request_bytes = config.max_request_bytes
-        self.allowed_hosts = frozenset(LOOPBACK_HOSTS | {str(config.host).lower()})
+        self.allowed_hosts = frozenset(
+            LOOPBACK_HOSTS
+            | {str(config.host).lower()}
+            | {split_host(name)[0].lower() for name in config.allowed_hosts}
+        )
+        self.trusted_proxies = frozenset(config.trusted_proxies)
 
     async def __call__(self, scope, receive, send) -> None:
         if scope["type"] != "http":
@@ -157,14 +250,26 @@ class SecurityMiddleware:
         # policy never has to parse a Cookie header.
         lookup = dict(headers)
         lookup["cookie-session"] = _cookie(headers.get("cookie", ""), SESSION_COOKIE)
+        client = scope.get("client")
+        scheme, lookup["host"] = external_view(
+            client=client[0] if client else None,
+            headers=headers,
+            scheme=scope.get("scheme", "http"),
+            trusted_proxies=self.trusted_proxies,
+        )
+        # What the session cookie's Secure flag is decided from, one hop later.
+        scope.setdefault("state", {})["external_scheme"] = scheme
         refusal = evaluate(
             method=scope.get("method", "GET"),
             path=scope.get("path", "/"),
             headers=lookup,
             session=self.session,
             allowed_hosts=self.allowed_hosts,
+            scheme=scheme,
             max_request_bytes=self.max_request_bytes,
         )
+        if refusal is None:
+            refusal = self._mutation_refusal(scope)
         if refusal is not None:
             response = JSONResponse(
                 status_code=refusal.status,
@@ -183,6 +288,26 @@ class SecurityMiddleware:
 
         await self.app(scope, receive, send_with_headers)
 
+    def _mutation_refusal(self, scope) -> Refusal | None:
+        """Refuse every mutating request while the library's dry run is unapproved.
+
+        One choke point for the whole API: every mutation the browser can start is a
+        non-safe method under ``/api/v1``. Reading stays open — an operator has to be
+        able to look at what the application found in order to approve it (US07-07).
+        """
+        method = scope.get("method", "GET").upper()
+        path = scope.get("path", "/")
+        if method in SAFE_METHODS or not path.startswith(API_PREFIX):
+            return None
+        if path in MUTATION_EXEMPT_PATHS:
+            return None
+        from photo_pipeline.services.release import mutation_blockers
+
+        blockers = mutation_blockers(self.config)
+        if not blockers:
+            return None
+        return Refusal(403, blockers[0]["code"], blockers[0]["message"])
+
 
 def _cookie(header: str, name: str) -> str:
     for part in header.split(";"):

photo_pipeline/config.py

@@ -7,6 +7,9 @@ real external call needs them.
 
 pydantic-settings would do this too, but a prefix-scan over the declared fields
 is a few lines and one fewer dependency.
+
+Tuple-valued settings are lists in one variable: library roots are ``os.pathsep``
+separated because they are paths, everything else is comma separated.
 """
 
 from __future__ import annotations
@@ -18,6 +21,62 @@ from typing import Mapping
 from pydantic import BaseModel, ConfigDict, SecretStr
 
 ENV_PREFIX = "PHOTO_PIPELINE_"
+ENV_FILE_VAR = f"{ENV_PREFIX}ENV_FILE"
+DEFAULT_ENV_FILE = Path(".env")
+COMMA_LIST_FIELDS = frozenset({"allowed_hosts", "trusted_proxies"})
+
+# The archived CLI's variable names, so the configuration file an operator already
+# has keeps working. The vision provider reads the OpenAI SDK's names, and the
+# library root is configuration here rather than a bare path (US07-01 donor).
+LEGACY_ALIASES = {
+    "LLM_API_KEY": "OPENAI_API_KEY",
+    "GEMINI_API_KEY": "OPENAI_API_KEY",
+    "LLM_BASE_URL": "OPENAI_BASE_URL",
+    "LIBRARY": f"{ENV_PREFIX}LIBRARY_ROOTS",
+}
+
+
+def parse_env_file(text: str) -> dict[str, str]:
+    """``KEY=value`` lines into a mapping. Comments, blanks, and quotes handled.
+
+    Deliberately not a shell: no interpolation, no ``export``, no multi-line values.
+    A configuration file that can run code is a configuration file that can be a
+    vulnerability.
+    """
+    values: dict[str, str] = {}
+    for line in text.splitlines():
+        line = line.strip()
+        if not line or line.startswith("#") or "=" not in line:
+            continue
+        key, _, raw = line.partition("=")
+        key = key.strip()
+        if not key or key.startswith("#"):
+            continue
+        value = raw.strip().strip('"').strip("'")
+        values[key] = value
+        alias = LEGACY_ALIASES.get(key)
+        if alias:
+            values.setdefault(alias, value)
+    return values
+
+
+def load_env_file(path: Path | str | None = None) -> dict[str, str]:
+    """Load ``PHOTO_PIPELINE_ENV_FILE`` (or ``./.env``) into the environment.
+
+    Anything already exported wins: a file is the standing configuration, the shell
+    is what you meant *this time*. Returns what it applied, which is what the CLI
+    prints — names only, never values.
+    """
+    candidate = path or os.environ.get(ENV_FILE_VAR) or DEFAULT_ENV_FILE
+    candidate = Path(candidate)
+    if not candidate.is_file():
+        return {}
+    applied = {}
+    for key, value in parse_env_file(candidate.read_text()).items():
+        if key not in os.environ:
+            os.environ[key] = value
+            applied[key] = value
+    return applied
 
 
 class Config(BaseModel):
@@ -30,6 +89,19 @@ class Config(BaseModel):
     log_level: str = "INFO"
     log_format: str = "json"  # "json" or "text"
 
+    # Trust boundary (US08-01). Empty means loopback only, which is what the app did
+    # before there was a setting: a request whose Host is not a loopback name is
+    # refused, and no secret is needed because nothing outside this machine can call.
+    # Naming a real hostname here is what makes the app reachable through a reverse
+    # proxy, and it is exactly then that ``access_secret`` becomes mandatory.
+    allowed_hosts: tuple[str, ...] = ()
+    # Addresses whose ``X-Forwarded-Proto``/``X-Forwarded-Host`` may be believed. A
+    # client that is not the proxy can otherwise declare its own origin.
+    trusted_proxies: tuple[str, ...] = ()
+    # Exchanged for the session cookie at the bootstrap endpoint. Once set it is
+    # required even on loopback, so a development setup cannot half-enable it.
+    access_secret: SecretStr | None = None
+
     # Largest request body the API accepts. Every endpoint takes small JSON commands;
     # anything larger is a mistake or an attempt to exhaust memory (US07-02).
     max_request_bytes: int = 1_048_576
@@ -42,6 +114,12 @@ class Config(BaseModel):
     # Free space an archive destination must keep beyond the transfer itself.
     archive_free_space_reserve_bytes: int = 1_000_000_000
 
+    # Refuse every mutating request until a read-only dry run of the configured
+    # library has been produced and explicitly approved (US07-07). Off by default so
+    # a development setup is unchanged; turn it on before pointing the application at
+    # a library whose photos cannot be replaced.
+    require_dry_run_approval: bool = False
+
     vision_api_key: SecretStr | None = None
     immich_api_key: SecretStr | None = None
     immich_server_url: str = ""
@@ -62,10 +140,18 @@ class Config(BaseModel):
     @classmethod
     def from_env(cls, environ: Mapping[str, str] | None = None) -> "Config":
         env = os.environ if environ is None else environ
+        if environ is None:
+            load_env_file()  # a file never overrides what the shell already set
+            env = os.environ
         data: dict = {}
         for name in cls.model_fields:
             raw = env.get(ENV_PREFIX + name.upper())
             if not raw:
                 continue
-            data[name] = raw.split(os.pathsep) if name == "library_roots" else raw
+            if name == "library_roots":
+                data[name] = raw.split(os.pathsep)
+            elif name in COMMA_LIST_FIELDS:
+                data[name] = [part.strip() for part in raw.split(",") if part.strip()]
+            else:
+                data[name] = raw
         return cls(**data)

photo_pipeline/services/analysis.py

@@ -264,8 +264,11 @@ class AnalysisService:
             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).
+                # the hash of what is actually on disk now (concept §3), and the
+                # recorded size has to move with it (US07-07).
                 asset.current_sha256 = checkpoint.sha256
+                if checkpoint.byte_size is not None:
+                    asset.byte_size = checkpoint.byte_size
             session.commit()
 
     def get(self, asset_id: str) -> dict | None:

photo_pipeline/services/exif_checkpoint.py

@@ -21,6 +21,7 @@ unreadable file, or a write that did not take is not evidence that metadata is f
 from __future__ import annotations
 
 import json
+import os
 import uuid
 from dataclasses import dataclass
 from datetime import datetime, timezone
@@ -58,6 +59,11 @@ class CheckpointResult:
     state: str  # verified | divergent | failed
     changed_fields: tuple[str, ...] = ()
     sha256: str | None = None
+    # exiftool rewrites the container, so the file's size moves with its hash. Both
+    # are inventory facts about the current bytes and both have to be refreshed
+    # together, or the next stage compares against a size that no longer exists
+    # (US07-07: a rename plan blocked itself forever after any EXIF write).
+    byte_size: int | None = None
     verified_at: datetime | None = None
     reason: str | None = None
 
@@ -137,9 +143,12 @@ def run(
 
     changed = compare(before, after)
     sha256 = hashing.sha256_file(path)
+    byte_size = os.path.getsize(path)
     if changed:
-        return CheckpointResult(DIVERGENT, changed_fields=changed, sha256=sha256)
-    return CheckpointResult(VERIFIED, sha256=sha256, verified_at=_now())
+        return CheckpointResult(
+            DIVERGENT, changed_fields=changed, sha256=sha256, byte_size=byte_size
+        )
+    return CheckpointResult(VERIFIED, sha256=sha256, byte_size=byte_size, verified_at=_now())
 
 
 def record(

photo_pipeline/services/release.py

@@ -0,0 +1,391 @@
+"""The release gate, the real-library dry run, and the approval that unlocks
+mutation (US07-07, concept §18 release gates).
+
+Three things live here because they are one decision:
+
+1. **The gate** — one command that provisions an isolated stack, runs every suite in
+   a fixed order, and retains versioned evidence with checksums. A release is not
+   "the tests passed on my machine last Tuesday"; it is a report that says which
+   revision, which suites, how long, and what the artefacts hash to.
+2. **The dry run** — a strictly read-only pass over the real photo library that
+   answers "what would this application do to it?" before it is allowed to do
+   anything. It opens no file for writing, creates no database rows, and touches no
+   metadata; it counts, classifies, and reconciles against whatever the database
+   already knows.
+3. **The approval** — a person reads that report and signs it off for exactly the
+   library roots it describes. Until then, with
+   ``PHOTO_PIPELINE_REQUIRE_DRY_RUN_APPROVAL`` set, every mutating request is
+   refused. Change the roots, or produce a newer report, and the approval no longer
+   matches: it approves *that* reconciliation, not the idea of mutating.
+"""
+
+from __future__ import annotations
+
+import hashlib
+import json
+import os
+import subprocess
+import sys
+import time
+from collections import Counter
+from dataclasses import dataclass
+from datetime import datetime, timezone
+from pathlib import Path
+
+from photo_pipeline import path_policy
+from photo_pipeline.config import Config
+
+SCHEMA_VERSION = 1
+APPROVAL_NAME = "dry-run-approval.json"
+CHECKSUMS_NAME = "CHECKSUMS.sha256"
+REPORT_NAME = "release-report.json"
+
+# The suites, in the order a failure is cheapest to read: units before the stacks
+# they compose. ``label`` is what the report and the operator see.
+STAGES: tuple[tuple[str, tuple[str, ...]], ...] = (
+    ("unit", ("tests/unit",)),
+    ("characterization", ("tests/characterization",)),
+    ("integration", ("tests/integration",)),
+    ("browser", ("tests/e2e",)),
+)
+
+# Skips the gate accepts, because they describe the machine rather than the code.
+ALLOWED_SKIP_REASONS = ("exiftool not installed", "root ignores directory permissions")
+
+
+class ReleaseError(RuntimeError):
+    pass
+
+
+def _now() -> datetime:
+    return datetime.now(timezone.utc)
+
+
+def sha256_file(path: Path) -> str:
+    digest = hashlib.sha256()
+    with path.open("rb") as handle:
+        for chunk in iter(lambda: handle.read(1024 * 1024), b""):
+            digest.update(chunk)
+    return digest.hexdigest()
+
+
+def sha256_bytes(payload: bytes) -> str:
+    return hashlib.sha256(payload).hexdigest()
+
+
+def revision() -> str | None:
+    """The commit this gate ran against, when the tree is a git checkout."""
+    try:
+        result = subprocess.run(
+            ["git", "rev-parse", "HEAD"],
+            capture_output=True,
+            text=True,
+            timeout=10,
+            cwd=str(Path(__file__).resolve().parents[2]),
+        )
+    except (OSError, subprocess.SubprocessError):
+        return None
+    return result.stdout.strip() or None
+
+
+# ── the story matrix ─────────────────────────────────────────────────────────
+
+
+def story_matrix(repo: Path | None = None) -> dict:
+    """Every backlog story, and how it is covered.
+
+    A story is ``delivered`` (mapped to test files that exist) or ``planned`` (an
+    accepted, not-yet-implemented story). Anything else — a story file nobody
+    mapped, or a mapping to a file that is gone — is a hole in the matrix, and the
+    gate fails on it rather than reporting a green run over missing coverage.
+    """
+    repo = repo or Path(__file__).resolve().parents[2]
+    traceability = json.loads((repo / "tests" / "story_traceability.json").read_text())
+    mapped: dict[str, list[str]] = traceability["stories"]
+    planned: list[str] = traceability.get("planned", [])
+    stories = sorted(
+        "-".join(path.stem.split("-")[:2])
+        for path in (repo / "delivery_backlog" / "stories").glob("US*.md")
+    )
+
+    missing_tests = [
+        f"{story}: {rel}"
+        for story, files in mapped.items()
+        for rel in files
+        if not (repo / rel).is_file()
+    ]
+    unmapped = [s for s in stories if s not in mapped and s not in planned]
+    unknown = [s for s in list(mapped) + planned if s not in stories]
+    overlap = sorted(set(mapped) & set(planned))
+    return {
+        "stories": len(stories),
+        "delivered": sorted(mapped),
+        "planned": sorted(planned),
+        "problems": [
+            *(f"story with no tests and not planned: {s}" for s in unmapped),
+            *(f"mapped test file is missing — {entry}" for entry in missing_tests),
+            *(f"mapped story is not in the backlog: {s}" for s in unknown),
+            *(f"story is both delivered and planned: {s}" for s in overlap),
+        ],
+    }
+
+
+# ── the gate ─────────────────────────────────────────────────────────────────
+
+
+@dataclass
+class StageResult:
+    label: str
+    command: list[str]
+    returncode: int
+    seconds: float
+    summary: str
+    skipped: list[str]
+
+    def as_dict(self) -> dict:
+        return {
+            "stage": self.label,
+            "command": self.command,
+            "returncode": self.returncode,
+            "seconds": round(self.seconds, 2),
+            "summary": self.summary,
+            "skipped": self.skipped,
+            "ok": self.returncode == 0,
+        }
+
+
+def _run_stage(label: str, paths: tuple[str, ...], *, repo: Path, log_dir: Path) -> StageResult:
+    command = [sys.executable, "-m", "pytest", *paths, "-q", "-rs"]
+    started = time.monotonic()
+    result = subprocess.run(command, cwd=str(repo), capture_output=True, text=True)
+    elapsed = time.monotonic() - started
+    output = result.stdout + result.stderr
+    (log_dir / f"{label}.log").write_text(output)
+    lines = [line for line in output.splitlines() if line.strip()]
+    summary = lines[-1] if lines else ""
+    skipped = [line for line in lines if line.startswith("SKIPPED")]
+    return StageResult(label, command, result.returncode, elapsed, summary, skipped)
+
+
+def unexpected_skips(results: list[StageResult]) -> list[str]:
+    """Skips the gate will not accept: everything but the documented environment ones."""
+    return [
+        line
+        for result in results
+        for line in result.skipped
+        if not any(reason in line for reason in ALLOWED_SKIP_REASONS)
+    ]
+
+
+def run_gate(
+    config: Config,
+    *,
+    output: Path | str | None = None,
+    stages: tuple[tuple[str, tuple[str, ...]], ...] = STAGES,
+    repo: Path | None = None,
+) -> dict:
+    """Run every suite in an isolated stack and retain checksummed evidence.
+
+    The stack is isolated by construction: each pytest run builds its own temporary
+    data directories and libraries, so the gate never reads or writes the operator's
+    photos. What it keeps afterwards is the report, the per-stage logs, and a
+    checksum file over both.
+    """
+    repo = repo or Path(__file__).resolve().parents[2]
+    directory = Path(output) if output else Path(config.data_dir) / "release" / _now().strftime(
+        "%Y%m%dT%H%M%SZ"
+    )
+    logs = directory / "logs"
+    logs.mkdir(parents=True, exist_ok=True)
+
+    matrix = story_matrix(repo)
+    results = [_run_stage(label, paths, repo=repo, log_dir=logs) for label, paths in stages]
+    skips = unexpected_skips(results)
+
+    report = {
+        "schema_version": SCHEMA_VERSION,
+        "started_at": _now().isoformat(),
+        "revision": revision(),
+        "python": sys.version.split()[0],
+        "platform": os.uname().sysname,
+        "matrix": matrix,
+        "stages": [result.as_dict() for result in results],
+        "unexpected_skips": skips,
+        "failures": [result.label for result in results if result.returncode != 0],
+    }
+    report["ok"] = not report["failures"] and not matrix["problems"] and not skips
+    report["finished_at"] = _now().isoformat()
+
+    (directory / REPORT_NAME).write_text(json.dumps(report, indent=2))
+    # The evidence is only evidence if it can be shown to be the evidence that was
+    # produced. Checksums are the honest version of "signed" without a key: a real
+    # signature belongs to whatever key management the release actually has.
+    checksums = "\n".join(
+        f"{sha256_file(path)}  {path.relative_to(directory)}"
+        for path in sorted(directory.rglob("*"))
+        if path.is_file() and path.name != CHECKSUMS_NAME
+    )
+    (directory / CHECKSUMS_NAME).write_text(checksums + "\n")
+    report["evidence"] = str(directory)
+    return report
+
+
+# ── the real-library dry run ─────────────────────────────────────────────────
+
+
+def dry_run(config: Config, *, roots: tuple[Path, ...] | None = None) -> dict:
+    """Read-only reconciliation of the configured library. Changes nothing.
+
+    Opens no file for writing, writes no database row, and reads only what
+    ``os.stat`` and the existing database already say. The point is to be able to
+    look at a real library — the one with the irreplaceable photos in it — and see
+    what the application believes about it before it is allowed to act.
+    """
+    roots = roots or tuple(Path(root) for root in config.library_roots)
+    if not roots:
+        raise ReleaseError("no library roots are configured")
+
+    by_extension: Counter = Counter()
+    folders: set[str] = set()
+    files: list[str] = []
+    unreadable: list[str] = []
+    excluded = 0
+    total_bytes = 0
+    for root in roots:
+        if not Path(root).is_dir():
+            raise ReleaseError(f"library root {root} is not a directory")
+        for path in sorted(Path(root).rglob("*")):
+            if path.is_dir():
+                # Never traverse into an excluded directory, and never report its
+                # contents: proving exclusion must not require opening it.
+                if path_policy.is_excluded(path):
+                    excluded += 1
+                continue
+            if path_policy.is_excluded(path):
+                continue
+            try:
+                stat = path.stat()
+            except OSError:
+                unreadable.append(str(path))
+                continue
+            files.append(str(path))
+            folders.add(str(path.parent))
+            by_extension[path.suffix.lower() or "(none)"] += 1
+            total_bytes += stat.st_size
+
+    known = _known_paths(config)
+    on_disk = set(files)
+    report = {
+        "schema_version": SCHEMA_VERSION,
+        "generated_at": _now().isoformat(),
+        "revision": revision(),
+        "library_roots": [str(root) for root in roots],
+        "files": len(files),
+        "folders": len(folders),
+        "bytes": total_bytes,
+        "excluded_directories": excluded,
+        "unreadable": unreadable,
+        "by_extension": dict(sorted(by_extension.items())),
+        "reconciliation": {
+            "known_to_database": len(known),
+            "already_registered": len(on_disk & known),
+            "new_to_the_application": len(on_disk - known),
+            "recorded_but_absent": sorted(known - on_disk)[:100],
+            "recorded_but_absent_total": len(known - on_disk),
+        },
+        "mutation": "none — this pass is read-only",
+    }
+    report["checksum"] = sha256_bytes(
+        json.dumps(report, sort_keys=True).encode("utf-8")
+    )
+    return report
+
+
+def _known_paths(config: Config) -> set[str]:
+    """Current asset paths the database holds, or an empty set if there is none."""
+    if not config.database_path.exists():
+        return set()
+    from sqlalchemy import select
+
+    from photo_pipeline.db import create_db_engine, create_session_factory
+    from photo_pipeline.models import Asset
+
+    engine = create_db_engine(config.database_url)
+    try:
+        with create_session_factory(engine)() as session:
+            return {
+                path
+                for path in session.scalars(select(Asset.current_path))
+                if path is not None
+            }
+    except Exception:
+        return set()
+    finally:
+        engine.dispose()
+
+
+# ── the approval ─────────────────────────────────────────────────────────────
+
+
+def approval_path(config: Config) -> Path:
+    return Path(config.data_dir) / APPROVAL_NAME
+
+
+def approve(config: Config, report: dict | Path | str, *, approver: str) -> dict:
+    """Record that a person read this reconciliation and accepts mutation for it."""
+    if isinstance(report, (str, Path)):
+        report = json.loads(Path(report).read_text())
+    if "checksum" not in report:
+        raise ReleaseError("this is not a dry-run report: it has no checksum")
+    record = {
+        "schema_version": SCHEMA_VERSION,
+        "approved_at": _now().isoformat(),
+        "approved_by": approver,
+        "report_checksum": report["checksum"],
+        "library_roots": report["library_roots"],
+        "files": report["files"],
+        "revision": report.get("revision"),
+    }
+    path = approval_path(config)
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text(json.dumps(record, indent=2))
+    return record
+
+
+def mutation_blockers(config: Config) -> list[dict]:
+    """Why mutation must stay refused, or an empty list.
+
+    Only enforced when ``require_dry_run_approval`` is configured — the loopback
+    developer setup keeps working unchanged, and an operator turns this on before
+    pointing the application at the library they cannot replace.
+    """
+    if not config.require_dry_run_approval:
+        return []
+    path = approval_path(config)
+    if not path.exists():
+        return [
+            {
+                "code": "dry_run_not_approved",
+                "message": (
+                    "run `python -m photo_pipeline dry-run` and approve its report "
+                    "before mutation is enabled"
+                ),
+            }
+        ]
+    try:
+        record = json.loads(path.read_text())
+    except ValueError:
+        return [{"code": "approval_unreadable", "message": f"{path} is not readable JSON"}]
+    approved_roots = [str(root) for root in record.get("library_roots", [])]
+    configured = [str(root) for root in config.library_roots]
+    if sorted(approved_roots) != sorted(configured):
+        return [
+            {
+                "code": "approval_scope_mismatch",
+                "message": (
+                    f"the approval covers {approved_roots}, but the configured library "
+                    f"is {configured}; run a new dry run"
+                ),
+            }
+        ]
+    return []

photo_pipeline/services/safety.py

@@ -294,6 +294,7 @@ class SafetyService:
 
         exif_verified_at = None
         result_sha256 = None
+        result_byte_size = None
         if write_exif and decision in (SFW, NSFW) and path:
             ops = exif_projection(decision)
             # The full checkpoint: write the owned keyword, read the whole file back,
@@ -314,6 +315,7 @@ class SafetyService:
             if result.verified:
                 exif_verified_at = result.verified_at
                 result_sha256 = result.sha256
+                result_byte_size = result.byte_size
 
         now = _now()
         with self._session_factory() as session:
@@ -332,6 +334,8 @@ class SafetyService:
             if result_sha256:
                 asset = session.get(Asset, asset_id)
                 asset.current_sha256 = result_sha256
+                if result_byte_size is not None:
+                    asset.byte_size = result_byte_size
             session.commit()
         return {
             "asset_id": asset_id,

pyproject.toml

@@ -8,18 +8,32 @@ dependencies = [
     "sqlalchemy>=2.0",
     "alembic>=1.13",
     "pydantic>=2.7",
-]
-
-[project.optional-dependencies]
-test = [
-    "pytest>=8",
-    "httpx>=0.27",
+    # Imaging is runtime, not test-only: thumbnails decode through Pillow, and the
+    # perceptual hash is a DCT over the decoded pixels (services/hashing.py).
     "pillow>=10",
     "numpy>=1.26",
     "scipy>=1.11",
+]
+
+[project.optional-dependencies]
+# The cloud vision provider. Optional because the analysis stage is the only thing
+# that needs it, and a local review-only install should not pull an API client.
+vision = ["openai>=1.30"]
+test = [
+    "pytest>=8",
+    "httpx>=0.27",
     "playwright>=1.40",
     "pytest-playwright>=0.4",
 ]
+
+[build-system]
+requires = ["setuptools>=68"]
+build-backend = "setuptools.build_meta"
+
+[tool.setuptools]
+# The importable application. ``migrations`` and ``work_item`` live beside it but
+# are not part of the package; without this, an editable install cannot guess.
+packages = ["photo_pipeline"]
 # Browser end-to-end tests also require: python -m playwright install chromium
 
 [tool.ruff]

tests/e2e/test_release_gate.py

@@ -0,0 +1,298 @@
+"""The release gate, the read-only dry run, and the approval that unlocks mutation
+(US07-07).
+
+The gate itself is exercised with a tiny stage set — running the whole suite from
+inside the suite would be a fork bomb with better manners. What is proven here is
+the machinery a release depends on: the story matrix is complete, a failing stage
+fails the gate, an unexpected skip fails the gate, and the evidence is written with
+checksums that match what was written.
+
+The dry run is proven to be read-only against a real temporary library, and the
+approval is proven to be what stands between a configured library and any mutation.
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import uuid
+from datetime import datetime, timezone
+from pathlib import Path
+
+import pytest
+from fastapi.testclient import TestClient
+
+from photo_pipeline.api.app import create_app
+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
+from photo_pipeline.services import release
+
+REPO = Path(__file__).resolve().parents[2]
+NOW = datetime(2026, 1, 1, tzinfo=timezone.utc)
+
+# Two throwaway stages: one that passes, one the test can point at a failure.
+PASSING = ("tests/e2e/test_traceability.py",)
+
+
+def _config(tmp_path, **extra) -> Config:
+    data = tmp_path / "data"
+    data.mkdir(parents=True, exist_ok=True)
+    lib = tmp_path / "lib"
+    lib.mkdir(exist_ok=True)
+    return Config.from_env(
+        {
+            "PHOTO_PIPELINE_DATA_DIR": str(data),
+            "PHOTO_PIPELINE_LIBRARY_ROOTS": str(lib),
+            **extra,
+        }
+    )
+
+
+# ── the story matrix ─────────────────────────────────────────────────────────
+
+
+def test_every_backlog_story_is_delivered_or_explicitly_planned():
+    matrix = release.story_matrix(REPO)
+
+    assert matrix["problems"] == [], "the story matrix has holes"
+    assert len(matrix["delivered"]) + len(matrix["planned"]) == matrix["stories"]
+    assert "US01-01" in matrix["delivered"] and "US07-07" in matrix["delivered"]
+
+
+def test_a_story_without_tests_is_a_gate_failure(tmp_path):
+    """A story file nobody covered must not pass quietly as 'no tests ran'."""
+    fake = tmp_path / "repo"
+    (fake / "delivery_backlog" / "stories").mkdir(parents=True)
+    (fake / "tests").mkdir()
+    (fake / "delivery_backlog" / "stories" / "US99-01-invented.md").write_text("# US99-01")
+    (fake / "tests" / "story_traceability.json").write_text(json.dumps({"stories": {}}))
+
+    matrix = release.story_matrix(fake)
+
+    assert matrix["problems"] == ["story with no tests and not planned: US99-01"]
+
+
+def test_a_mapping_to_a_deleted_test_file_is_a_gate_failure(tmp_path):
+    fake = tmp_path / "repo"
+    (fake / "delivery_backlog" / "stories").mkdir(parents=True)
+    (fake / "tests").mkdir()
+    (fake / "delivery_backlog" / "stories" / "US99-01-invented.md").write_text("# US99-01")
+    (fake / "tests" / "story_traceability.json").write_text(
+        json.dumps({"stories": {"US99-01": ["tests/gone.py"]}})
+    )
+
+    assert release.story_matrix(fake)["problems"] == [
+        "mapped test file is missing — US99-01: tests/gone.py"
+    ]
+
+
+# ── the gate ─────────────────────────────────────────────────────────────────
+
+
+def test_the_gate_runs_its_stages_and_keeps_checksummed_evidence(tmp_path):
+    config = _config(tmp_path)
+    evidence = tmp_path / "evidence"
+
+    report = release.run_gate(config, output=evidence, stages=(("smoke", PASSING),))
+
+    assert report["ok"] is True and report["failures"] == []
+    assert report["stages"][0]["stage"] == "smoke" and report["stages"][0]["ok"] is True
+    assert report["revision"], "the evidence must say which commit it covers"
+    assert report["matrix"]["problems"] == []
+
+    written = json.loads((evidence / release.REPORT_NAME).read_text())
+    assert written["ok"] is True
+    assert (evidence / "logs" / "smoke.log").exists()
+    checksums = (evidence / release.CHECKSUMS_NAME).read_text().splitlines()
+    assert len(checksums) >= 2
+    for line in checksums:
+        digest, name = line.split("  ", 1)
+        assert release.sha256_file(evidence / name) == digest
+
+
+def test_a_failing_stage_fails_the_gate(tmp_path):
+    config = _config(tmp_path)
+    failing = tmp_path / "failing_test.py"
+    failing.write_text("def test_no():\n    assert False\n")
+
+    report = release.run_gate(
+        config, output=tmp_path / "evidence", stages=(("broken", (str(failing),)),)
+    )
+
+    assert report["ok"] is False and report["failures"] == ["broken"]
+    assert report["stages"][0]["returncode"] != 0
+
+
+def test_an_unexpected_skip_fails_the_gate_but_an_environment_skip_does_not():
+    environment = release.StageResult(
+        "unit", [], 0, 0.1, "1 skipped", ["SKIPPED [1] x.py:1: exiftool not installed"]
+    )
+    silent = release.StageResult(
+        "unit", [], 0, 0.1, "1 skipped", ["SKIPPED [1] x.py:1: flaky, look later"]
+    )
+
+    assert release.unexpected_skips([environment]) == []
+    assert release.unexpected_skips([silent, environment]) == [
+        "SKIPPED [1] x.py:1: flaky, look later"
+    ]
+
+
+# ── the real-library dry run ─────────────────────────────────────────────────
+
+
+def _library(root: Path) -> None:
+    (root / "album").mkdir(parents=True)
+    (root / "album" / "a.jpg").write_bytes(b"a" * 128)
+    (root / "album" / "b.png").write_bytes(b"b" * 64)
+    (root / "loose.JPG").write_bytes(b"c" * 32)
+    excluded = root / "_IGNORE" / "private"
+    excluded.mkdir(parents=True)
+    (excluded / "secret.jpg").write_bytes(b"never read")
+
+
+def test_the_dry_run_describes_the_library_without_touching_it(tmp_path):
+    config = _config(tmp_path)
+    root = Path(config.library_roots[0])
+    _library(root)
+    before = {
+        str(p): (p.stat().st_mtime_ns, p.read_bytes()) for p in root.rglob("*") if p.is_file()
+    }
+
+    report = release.dry_run(config)
+
+    assert report["files"] == 3, "the excluded sentinel is not counted"
+    assert report["by_extension"] == {".jpg": 2, ".png": 1}
+    assert report["excluded_directories"] >= 1
+    assert report["mutation"] == "none — this pass is read-only"
+    assert report["checksum"]
+    assert not any("secret" in json.dumps(report) for _ in [0]), "excluded content never appears"
+    after = {
+        str(p): (p.stat().st_mtime_ns, p.read_bytes()) for p in root.rglob("*") if p.is_file()
+    }
+    assert after == before, "a read-only pass changed the library"
+
+
+def test_the_dry_run_reconciles_against_what_the_database_already_knows(tmp_path):
+    config = _config(tmp_path)
+    root = Path(config.library_roots[0])
+    _library(root)
+    run_migrations(config.database_url)
+    engine = create_db_engine(config.database_url)
+    with create_session_factory(engine)() as session:
+        session.add(
+            Asset(
+                id=str(uuid.uuid4()),
+                original_path=str(root / "album" / "a.jpg"),
+                current_path=str(root / "album" / "a.jpg"),
+                discovered_at=NOW,
+                hash_version=1,
+                byte_size=128,
+            )
+        )
+        session.add(
+            Asset(
+                id=str(uuid.uuid4()),
+                original_path=str(root / "album" / "gone.jpg"),
+                current_path=str(root / "album" / "gone.jpg"),
+                discovered_at=NOW,
+                hash_version=1,
+                byte_size=1,
+            )
+        )
+        session.commit()
+    engine.dispose()
+
+    reconciliation = release.dry_run(config)["reconciliation"]
+
+    assert reconciliation["known_to_database"] == 2
+    assert reconciliation["already_registered"] == 1
+    assert reconciliation["new_to_the_application"] == 2
+    assert reconciliation["recorded_but_absent_total"] == 1
+    assert reconciliation["recorded_but_absent"][0].endswith("gone.jpg")
+
+
+def test_a_library_root_that_is_not_there_is_refused(tmp_path):
+    config = _config(tmp_path, PHOTO_PIPELINE_LIBRARY_ROOTS=str(tmp_path / "nowhere"))
+    with pytest.raises(release.ReleaseError, match="not a directory"):
+        release.dry_run(config)
+
+
+# ── the approval ─────────────────────────────────────────────────────────────
+
+
+def test_mutation_is_refused_until_the_dry_run_is_approved(tmp_path):
+    config = _config(tmp_path, PHOTO_PIPELINE_REQUIRE_DRY_RUN_APPROVAL="1")
+    _library(Path(config.library_roots[0]))
+
+    blockers = release.mutation_blockers(config)
+    assert [blocker["code"] for blocker in blockers] == ["dry_run_not_approved"]
+
+    record = release.approve(config, release.dry_run(config), approver="domverse")
+    assert record["approved_by"] == "domverse" and record["report_checksum"]
+    assert release.mutation_blockers(config) == []
+
+
+def test_an_approval_covers_the_library_it_was_written_for(tmp_path):
+    config = _config(tmp_path, PHOTO_PIPELINE_REQUIRE_DRY_RUN_APPROVAL="1")
+    _library(Path(config.library_roots[0]))
+    release.approve(config, release.dry_run(config), approver="domverse")
+
+    other = tmp_path / "other-library"
+    other.mkdir()
+    moved = config.model_copy(update={"library_roots": (other,)})
+
+    assert [b["code"] for b in release.mutation_blockers(moved)] == ["approval_scope_mismatch"]
+
+
+def test_without_the_requirement_nothing_changes(tmp_path):
+    config = _config(tmp_path)  # the loopback development default
+    assert release.mutation_blockers(config) == []
+
+
+def test_the_api_refuses_every_mutation_until_the_report_is_approved(tmp_path):
+    config = _config(tmp_path, PHOTO_PIPELINE_REQUIRE_DRY_RUN_APPROVAL="1")
+    _library(Path(config.library_roots[0]))
+    with TestClient(create_app(config)) as client:
+        # Reading stays open: an operator has to see what was found to approve it.
+        assert client.get("/api/v1/workflow").status_code == 200
+        refused = client.post("/api/v1/inventory/scan", json={})
+        assert refused.status_code == 403
+        assert refused.json()["error"]["code"] == "dry_run_not_approved"
+        # A backup is the one mutation a careful operator takes first.
+        assert client.post("/api/v1/backups", json={}).status_code == 201
+
+        release.approve(config, release.dry_run(config), approver="domverse")
+        assert client.post("/api/v1/inventory/scan", json={}).status_code in (200, 201, 202)
+
+
+def test_approving_something_that_is_not_a_report_is_refused(tmp_path):
+    config = _config(tmp_path)
+    with pytest.raises(release.ReleaseError, match="not a dry-run report"):
+        release.approve(config, {"files": 3}, approver="domverse")
+
+
+def test_the_cli_runs_the_dry_run_and_the_approval(tmp_path):
+    config = _config(tmp_path, PHOTO_PIPELINE_REQUIRE_DRY_RUN_APPROVAL="1")
+    _library(Path(config.library_roots[0]))
+    from photo_pipeline.__main__ import main
+
+    environment = {
+        "PHOTO_PIPELINE_DATA_DIR": str(config.data_dir),
+        "PHOTO_PIPELINE_LIBRARY_ROOTS": str(config.library_roots[0]),
+        "PHOTO_PIPELINE_REQUIRE_DRY_RUN_APPROVAL": "1",
+    }
+    previous = {key: os.environ.get(key) for key in environment}
+    os.environ.update(environment)
+    try:
+        report_path = tmp_path / "dry-run.json"
+        assert main(["dry-run", "--output", str(report_path)]) == 0
+        assert json.loads(report_path.read_text())["files"] == 3
+        assert main(["approve-dry-run", str(report_path), "--approver", "domverse"]) == 0
+    finally:
+        for key, value in previous.items():
+            if value is None:
+                os.environ.pop(key, None)
+            else:
+                os.environ[key] = value
+    assert release.mutation_blockers(config) == []

tests/e2e/test_release_journey.py

@@ -0,0 +1,348 @@
+"""The release journey (US07-07): one library, one fresh environment, every stage.
+
+This is the acceptance the whole backlog builds up to — discovery, duplicate review,
+safety, analysis, EXIF verification, album proposal, guarded rename, rescan and
+reconciliation, upload, archive, offline deduplication, restore — driven over HTTP
+against real ``photo_pipeline serve`` and worker child processes, with full process
+restarts in the middle and at the end.
+
+Nothing is reached into. External services are the deterministic fakes the earlier
+phases already use, invoked through the real integration layer: a vision fake that
+records every path it was given, a real fake ``immich-go`` executable, and an
+archive medium that is an ordinary directory whose marker file is its identity.
+
+The invariants asserted along the way are the ones the concept calls non-negotiable:
+
+- an asset's identity survives a rename, an upload, an archive, and a restore;
+- an ``_IGNORE`` sentinel is never discovered, counted, analysed, or uploaded;
+- an NSFW asset never reaches the vision provider but still reaches Immich;
+- no photo's bytes are lost at any point — every hash is still reachable somewhere;
+- every stage's durable state survives a restart of both processes.
+"""
+
+from __future__ import annotations
+
+import hashlib
+import shutil
+from pathlib import Path
+
+import httpx
+import pytest
+
+from tests.e2e._pipeline_harness import (
+    SENTINEL_KEY,
+    FakeImmich,
+    Server,
+    fake_uploader,
+    image,
+    seed_library,
+    start_worker,
+    wait_until,
+)
+
+TIMEOUT = 30
+ALBUM = "rome"
+UPLOADER = 'echo "INFO  uploaded $6"\necho "Uploaded 2, duplicates 0"\nexit 0\n'
+
+
+def _sha256(path: Path) -> str:
+    return hashlib.sha256(path.read_bytes()).hexdigest()
+
+
+def _hashes(*roots: Path) -> set[str]:
+    return {
+        _sha256(path)
+        for root in roots
+        for path in root.rglob("*.jpg")
+        if path.is_file() and not path.name.startswith(".")
+    }
+
+
+def _post(base: str, path: str, **kwargs) -> httpx.Response:
+    response = httpx.post(f"{base}/api/v1{path}", timeout=TIMEOUT, **kwargs)
+    response.raise_for_status()
+    return response
+
+
+def _get(base: str, path: str, **kwargs) -> dict:
+    response = httpx.get(f"{base}/api/v1{path}", timeout=TIMEOUT, **kwargs)
+    response.raise_for_status()
+    return response.json()
+
+
+def _await_job(base: str, job_id: str, *, states=("succeeded",)) -> dict:
+    return wait_until(
+        lambda: (
+            snapshot
+            if (snapshot := _get(base, f"/jobs/{job_id}"))["state"] in states
+            else None
+        ),
+        timeout=90,
+    )
+
+
+@pytest.fixture
+def library(tmp_path):
+    """A fresh library: an album, an exact duplicate, and an excluded sentinel."""
+    seeded = seed_library(tmp_path, {}, {})
+    album = seeded.lib / ALBUM
+    image(album / "a.jpg", 11)
+    image(album / "b.jpg", 12)
+    shutil.copyfile(album / "a.jpg", album / "a-copy.jpg")  # exact duplicate
+    ignored = seeded.lib / "_IGNORE" / "private"
+    ignored.mkdir(parents=True)
+    image(ignored / "sentinel-9f3a2b.jpg", 99)
+    return seeded
+
+
+@pytest.mark.skipif(shutil.which("exiftool") is None, reason="exiftool not installed")
+def test_the_full_release_journey_survives_every_stage_and_two_restarts(library, tmp_path):
+    immich = FakeImmich()
+    uploader = fake_uploader(tmp_path, UPLOADER)
+    vision_log = tmp_path / "vision.log"
+    archive_root = tmp_path / "medium"
+    archive_root.mkdir()
+    environment = {
+        "PHOTO_PIPELINE_IMMICH_SERVER_URL": immich.url,
+        "PHOTO_PIPELINE_IMMICH_API_KEY": SENTINEL_KEY,
+        "PHOTO_PIPELINE_IMMICH_GO_BINARY": str(uploader),
+        "PHOTO_PIPELINE_ARCHIVE_FREE_SPACE_RESERVE_BYTES": "0",
+        "PHOTO_PIPELINE_FAKE_VISION_LOG": str(vision_log),
+    }
+    server = Server(library, extra_env=environment).start()
+    worker = start_worker(library, extra_env=environment)
+    base = server.base
+
+    try:
+        # ── 0. discovery ─────────────────────────────────────────────────────
+        _post(base, "/inventory/scan")
+        assets = _get(base, "/inventory/assets", params={"limit": 200})["items"]
+        assert len(assets) == 3, "the sentinel under _IGNORE is not an asset"
+        paths = {asset["current_path"] for asset in assets}
+        assert not any("_IGNORE" in path or "sentinel" in path for path in paths)
+        identity = {asset["id"]: Path(asset["current_path"]).name for asset in assets}
+
+        # ── 1. duplicate review ──────────────────────────────────────────────
+        _post(base, "/duplicates/detect")
+        clusters = _get(base, "/duplicates/clusters")["items"]
+        assert len(clusters) == 1 and clusters[0]["member_total"] == 2
+        cluster = _get(base, f"/duplicates/clusters/{clusters[0]['id']}")
+        canonical = sorted(member["asset_id"] for member in cluster["members"])[0]
+        _post(
+            base,
+            f"/duplicates/clusters/{cluster['id']}/decision",
+            json={
+                "decision": "canonical",
+                "canonical_asset_id": canonical,
+                "expected_version": cluster["version"],
+            },
+        )
+
+        # ── 2. safety, with its EXIF checkpoint ──────────────────────────────
+        queue = _get(base, "/safety/queue", params={"limit": 100})["items"]
+        assert len(queue) == 2, "a non-canonical variant is not reviewed twice"
+        decisions = {}
+        for index, item in enumerate(sorted(queue, key=lambda row: row["current_path"])):
+            decision = "nsfw" if index == 0 else "sfw"
+            decisions[item["asset_id"]] = decision
+            result = _post(
+                base, "/safety/decisions", json={"asset_id": item["asset_id"], "decision": decision}
+            ).json()
+            assert result["exif_verified"] is True, "the safety checkpoint must verify"
+
+        # ── restart: everything so far has to be durable ─────────────────────
+        server.stop()
+        server.start()
+        base = server.base
+        assert _get(base, "/safety/counts")["nsfw"] == 1
+        assert {a["id"] for a in _get(base, "/inventory/assets", params={"limit": 200})["items"]} == set(
+            identity
+        )
+
+        # ── 3. analysis, gated to confirmed-SFW assets ───────────────────────
+        job = _post(base, "/analysis/jobs").json()
+        _await_job(base, job["id"])
+        analysed = [
+            name
+            for name, decision in (
+                (identity[asset_id], decision) for asset_id, decision in decisions.items()
+            )
+            if decision == "sfw"
+        ]
+        seen = vision_log.read_text().splitlines()
+        assert len(seen) == len(analysed) == 1
+        assert not any("sentinel" in line or "_IGNORE" in line for line in seen)
+        nsfw_id = next(aid for aid, decision in decisions.items() if decision == "nsfw")
+        assert all(identity[nsfw_id] not in line for line in seen), "NSFW reached the provider"
+
+        # From here on no stage may change a photo's bytes: the metadata stages are
+        # done, and moving, uploading, archiving, and restoring only relocate them.
+        stable_hashes = _hashes(library.lib)
+
+        # ── 4. album proposal and guarded rename ─────────────────────────────
+        _post(base, "/albums/proposals", json={})
+        proposal = _get(base, f"/albums/proposals/{ALBUM}")
+        _post(
+            base,
+            f"/albums/proposals/{ALBUM}/edit",
+            json={"name": "2019 Rome", "expected_version": proposal["version"]},
+        )
+        proposal = _get(base, f"/albums/proposals/{ALBUM}")
+        _post(
+            base,
+            f"/albums/proposals/{ALBUM}/approve",
+            json={"expected_version": proposal["version"]},
+        )
+        plan = _post(base, "/rename-plans").json()
+        assert plan["blockers"] == [], [
+            (issue["code"], issue["message"])
+            for op in plan["operations"]
+            for issue in op["issues"]
+        ]
+        response = httpx.post(
+            f"{base}/api/v1/rename-plans/{plan['id']}/apply",
+            json={"expected_version": plan["version"], "expected_checksum": plan["checksum"]},
+            timeout=TIMEOUT,
+        )
+        assert response.status_code == 200, response.text
+        applied = response.json()
+        assert applied["failed"] == 0 and applied["applied"] == 1
+        assert (library.lib / "2019 Rome").is_dir() and not (library.lib / ALBUM).exists()
+
+        # ── 5. rescan and reconciliation: identity survives the move ─────────
+        _post(base, "/inventory/scan")
+        after_rename = _get(base, "/inventory/assets", params={"limit": 200})["items"]
+        assert {asset["id"] for asset in after_rename} == set(identity)
+        assert all("2019 Rome" in asset["current_path"] for asset in after_rename)
+        assert _hashes(library.lib) == stable_hashes, "a rename changed a photo's bytes"
+
+        # ── 6. upload ────────────────────────────────────────────────────────
+        report = _post(base, "/upload-preflight", json={"albums": ["2019 Rome"]}).json()
+        assert report["state"] == "ready", report["blockers"]
+        batch = _post(
+            base,
+            "/upload-batches",
+            json={"albums": ["2019 Rome"], "token": report["token"]},
+        ).json()["batches"][0]
+        started = _post(base, f"/upload-batches/{batch['id']}/start").json()
+        _await_job(base, started["job"]["id"])
+        uploaded = _get(base, f"/upload-batches/{batch['id']}")
+        assert uploaded["state"] == "succeeded"
+        # The uploader said nothing per file, so the outcome is uncertain until the
+        # server itself is asked whether it holds those exact bytes (US05-04).
+        assert uploaded["outcome_state"] == "requires_verification"
+        verified = _post(base, f"/upload-batches/{batch['id']}/verify").json()
+        assert verified["outcome_state"] == "verified", verified
+        uploaded = _get(base, f"/upload-batches/{batch['id']}")
+        # Reviewed NSFW is uploaded; it simply never reached the analyser.
+        assert {item["asset_id"] for item in uploaded["items"]} >= {nsfw_id}
+
+        # ── 7. archive ───────────────────────────────────────────────────────
+        location = _post(
+            base, "/archive-locations", json={"name": "external", "root": str(archive_root)}
+        ).json()
+        preflight = _post(
+            base, "/archive-preflight", json={"location_id": location["id"]}
+        ).json()
+        assert preflight["state"] == "ready", [
+            (asset["asset_id"], asset["blockers"])
+            for album in preflight["albums"]
+            for asset in album["assets"]
+            if asset["blockers"]
+        ] or preflight
+        archive_plan = _post(
+            base,
+            "/archive-plans",
+            json={"location_id": location["id"], "token": preflight["token"]},
+        ).json()
+        uploaded_ids = {item["asset_id"] for item in uploaded["items"]}
+        _post(base, f"/archive-plans/{archive_plan['id']}/apply")
+        wait_until(
+            lambda: all(
+                asset["availability_state"].startswith("archived")
+                for asset in _get(base, "/inventory/assets", params={"limit": 200})["items"]
+                if asset["id"] in uploaded_ids
+            ),
+            timeout=90,
+        )
+        assert _hashes(library.lib, archive_root) == stable_hashes, "archiving lost bytes"
+
+        # ── 8. offline deduplication ─────────────────────────────────────────
+        (archive_root / ".photo-pipeline-archive.json").rename(
+            archive_root / ".photo-pipeline-archive.json.away"
+        )
+        # A copy of an archived photo turns up in the library under its own name —
+        # the real shape of "I re-imported an old card" — so nothing occupies the
+        # path the archived original would be restored to.
+        returned = library.lib / "2019 Rome" / "rediscovered.jpg"
+        returned.parent.mkdir(parents=True, exist_ok=True)
+        archived_copy = next(archive_root.rglob("*.jpg"))
+        shutil.copyfile(archived_copy, returned)
+        _post(base, "/inventory/scan")
+        _post(base, "/duplicates/detect")
+        offline = _get(base, "/inventory/assets", params={"limit": 200})["items"]
+        archived = [a for a in offline if a["availability_state"].startswith("archived")]
+        assert archived, "an unmounted medium must not make assets missing"
+        assert all(a["availability_state"] != "missing_unexpected" for a in offline)
+        assert any(
+            cluster["member_total"] >= 2 for cluster in _get(base, "/duplicates/clusters")["items"]
+        ), "the rediscovered copy did not meet its archived original"
+
+        # ── 9. restore ───────────────────────────────────────────────────────
+        (archive_root / ".photo-pipeline-archive.json.away").rename(
+            archive_root / ".photo-pipeline-archive.json"
+        )
+        restore_report = _post(
+            base, "/restore-preflight", json={"location_id": location["id"]}
+        ).json()
+        restore_plan = _post(
+            base,
+            "/restore-plans",
+            json={"location_id": location["id"], "token": restore_report["token"]},
+        ).json()
+        _post(base, f"/restore-plans/{restore_plan['id']}/apply")
+        wait_until(
+            lambda: all(
+                asset["availability_state"] == "active"
+                for asset in _get(base, "/inventory/assets", params={"limit": 200})["items"]
+                if asset["id"] in identity
+            ),
+            timeout=90,
+        )
+
+        # ── 10. the final restart proves every stage was durable ─────────────
+        worker.kill()
+        worker.wait(timeout=20)
+        server.stop()
+        server.start()
+        base = server.base
+        final = {
+            asset["id"]: asset
+            for asset in _get(base, "/inventory/assets", params={"limit": 200})["items"]
+        }
+        assert set(identity) <= set(final), "an asset id did not survive the journey"
+        assert _get(base, "/safety/counts")["nsfw"] == 1
+        assert _get(base, "/upload-batches")["batches"][0]["state"] == "succeeded"
+        reachable = {
+            _sha256(path): str(path)
+            for root in (library.lib, archive_root)
+            for path in root.rglob("*.jpg")
+            if path.is_file() and not path.name.startswith(".")
+        }
+        assert stable_hashes <= set(reachable), (
+            "a photo was lost",
+            sorted(stable_hashes - set(reachable)),
+            sorted(reachable.values()),
+        )
+        workflow = _get(base, "/workflow")
+        assert {stage["key"] for stage in workflow["stages"]} >= {
+            "inventory",
+            "duplicates",
+            "safety",
+            "analysis",
+        }
+    finally:
+        worker.kill()
+        worker.wait(timeout=20)
+        server.stop()
+        immich.stop()

tests/integration/test_trusted_hosts.py

@@ -0,0 +1,253 @@
+"""US08-01: the configurable trust boundary and its authentication gate.
+
+Until now, reaching the app proved ownership of it: it answered only to loopback
+names. A container behind a reverse proxy answers to a real hostname, so these tests
+pin the two halves that replace that proof — the app refuses to start exposed without
+an access secret, and the secret is the only way to obtain the session every other
+route already required (US07-02, unchanged and re-asserted here).
+
+The suite's ``conftest`` bootstraps a session for any ``TestClient`` automatically,
+which is precisely what an unauthenticated caller does not get; ``raw_client``
+pre-seeds a placeholder CSRF header to opt out of that convenience.
+"""
+
+from __future__ import annotations
+
+import pytest
+from starlette.testclient import TestClient
+
+from photo_pipeline.api.app import ConfigurationRefused, create_app
+from photo_pipeline.api.security import ACCESS_SECRET_HEADER, CSRF_HEADER, SESSION_COOKIE
+from photo_pipeline.config import Config
+
+SECRET = "operator-secret-value"
+HOSTNAME = "photos.example.com"
+# What Starlette reports as the peer address of an in-process request.
+TESTCLIENT_ADDRESS = "testclient"
+
+# One of each route class: a read, a mutation, and a media endpoint.
+PROTECTED = [
+    ("GET", "/api/v1/workflow", None),
+    ("POST", "/api/v1/albums/proposals", {}),
+    ("GET", "/api/v1/assets/unknown-asset/thumbnail?size=256", None),
+]
+
+
+def config(tmp_path, **overrides) -> Config:
+    return Config(data_dir=tmp_path / "data", **overrides)
+
+
+def raw_client(app, base_url="http://127.0.0.1") -> TestClient:
+    client = TestClient(app, base_url=base_url)
+    client.headers[CSRF_HEADER] = "placeholder"
+    return client
+
+
+def exchange(client, secret=SECRET, headers=None):
+    return client.get("/api/v1/session", headers={ACCESS_SECRET_HEADER: secret, **(headers or {})})
+
+
+# ── startup: exposure without a secret is refused, loopback is unchanged ──────
+
+
+@pytest.mark.parametrize(
+    "exposure,exposed",
+    [({"allowed_hosts": (HOSTNAME,)}, HOSTNAME), ({"host": "0.0.0.0"}, "0.0.0.0")],
+)
+def test_an_exposed_configuration_refuses_to_serve_without_a_secret(tmp_path, exposure, exposed):
+    with pytest.raises(ConfigurationRefused) as refused:
+        create_app(config(tmp_path, **exposure))
+    assert "PHOTO_PIPELINE_ACCESS_SECRET" in str(refused.value)
+    # The message names what is exposed, so the operator knows which setting did it.
+    assert exposed in str(refused.value)
+
+
+def test_the_serve_command_reports_the_refusal_instead_of_binding(tmp_path, monkeypatch, capsys):
+    """Exit before the port, the lock, and the database, with a sentence not a trace."""
+    from photo_pipeline.__main__ import main
+
+    monkeypatch.setenv("PHOTO_PIPELINE_DATA_DIR", str(tmp_path / "data"))
+    monkeypatch.setenv("PHOTO_PIPELINE_ALLOWED_HOSTS", HOSTNAME)
+    monkeypatch.delenv("PHOTO_PIPELINE_ACCESS_SECRET", raising=False)
+
+    assert main(["serve"]) == 4
+    assert "PHOTO_PIPELINE_ACCESS_SECRET" in capsys.readouterr().err
+
+
+def test_an_exposed_configuration_with_a_secret_starts(tmp_path):
+    app = create_app(config(tmp_path, allowed_hosts=(HOSTNAME,), access_secret=SECRET))
+    with raw_client(app, base_url=f"http://{HOSTNAME}") as client:
+        assert exchange(client).status_code == 200
+
+
+def test_a_loopback_configuration_still_needs_no_secret(tmp_path):
+    """An unset trust boundary must behave exactly as it did before this story."""
+    with raw_client(create_app(config(tmp_path))) as client:
+        response = client.get("/api/v1/session")
+        assert response.status_code == 200
+        assert response.json()["csrf_token"]
+        assert "secure" not in response.headers["set-cookie"].lower()
+
+
+# ── the exchange: secret in, session out ─────────────────────────────────────
+
+
+@pytest.fixture
+def gated(tmp_path):
+    app = create_app(
+        config(
+            tmp_path,
+            allowed_hosts=(HOSTNAME,),
+            access_secret=SECRET,
+            trusted_proxies=(TESTCLIENT_ADDRESS,),
+        )
+    )
+    with raw_client(app, base_url=f"http://{HOSTNAME}") as client:
+        yield client
+
+
+def test_the_secret_buys_the_session_and_the_session_buys_the_routes(gated):
+    response = exchange(gated)
+    assert response.status_code == 200
+    cookie = response.headers["set-cookie"].lower()
+    assert "httponly" in cookie and "samesite=strict" in cookie
+    gated.headers[CSRF_HEADER] = response.json()["csrf_token"]
+
+    # The session and CSRF requirements behind the gate are the ones US07-02 set.
+    assert gated.get("/api/v1/workflow").status_code == 200
+    assert gated.post("/api/v1/albums/proposals", json={}).status_code == 200
+    refused = gated.post("/api/v1/albums/proposals", json={}, headers={CSRF_HEADER: "guessed"})
+    assert refused.json()["error"]["code"] == "csrf_failed"
+
+
+@pytest.mark.parametrize("offered", ["", "wrong-secret", SECRET + "x", SECRET.upper()])
+def test_a_wrong_secret_buys_nothing(gated, offered):
+    response = exchange(gated, secret=offered)
+    assert response.status_code == 401
+    assert response.json()["error"]["code"] == "access_denied"
+    assert "set-cookie" not in response.headers
+
+
+def test_a_refusal_never_echoes_the_secret_or_the_session(gated, caplog):
+    with caplog.at_level("WARNING"):
+        response = exchange(gated, secret="wrong-secret")
+    assert SECRET not in response.text and "wrong-secret" not in response.text
+    assert SECRET not in caplog.text
+    # Logged as an event with its caller, without the session it did not get.
+    assert "access secret rejected" in caplog.text
+
+
+def test_guessing_is_rate_limited(gated):
+    codes = [exchange(gated, secret=f"guess-{n}").status_code for n in range(6)]
+    assert codes.count(401) == 5 and codes[-1] == 429
+    assert gated.get("/api/v1/session").status_code == 429
+    # The right secret is refused too while the limiter holds: that is the point.
+    blocked = exchange(gated)
+    assert blocked.status_code == 429
+    assert SECRET not in blocked.text
+
+
+def test_every_route_class_is_unreachable_without_the_secret(gated):
+    for method, path, body in PROTECTED:
+        response = gated.request(method, path, json=body)
+        assert response.status_code == 401, path
+        assert response.json()["error"]["code"] == "unauthenticated", path
+    # Health stays open: an orchestrator restarting the container holds no secret.
+    assert gated.get("/api/v1/health/live").status_code == 200
+    assert gated.get("/api/v1/health/ready").status_code == 200
+
+
+def test_a_session_from_another_process_is_not_replayable(tmp_path):
+    """Sessions live in the process, so a cookie captured from a previous one — a
+    restarted container, or a second deployment — must not open this one."""
+    settings = dict(allowed_hosts=(HOSTNAME,), access_secret=SECRET)
+    first, second = (create_app(config(tmp_path / str(n), **settings)) for n in (1, 2))
+    with raw_client(first, base_url=f"http://{HOSTNAME}") as client:
+        exchange(client)
+        stolen = client.cookies[SESSION_COOKIE]
+
+    with raw_client(second, base_url=f"http://{HOSTNAME}") as client:
+        client.cookies.set(SESSION_COOKIE, stolen, domain=HOSTNAME)
+        response = client.get("/api/v1/workflow")
+    assert response.status_code == 401
+    assert response.json()["error"]["code"] == "unauthenticated"
+
+
+def test_a_cross_site_request_is_still_refused_behind_the_gate(gated):
+    gated.headers[CSRF_HEADER] = exchange(gated).json()["csrf_token"]
+    refused = gated.post(
+        "/api/v1/albums/proposals", json={}, headers={"Origin": "https://evil.example"}
+    )
+    assert refused.json()["error"]["code"] == "origin_not_allowed"
+    embedded = gated.get(
+        "/api/v1/assets/unknown-asset/thumbnail?size=256", headers={"Sec-Fetch-Site": "cross-site"}
+    )
+    assert embedded.json()["error"]["code"] == "cross_site_blocked"
+
+
+def test_an_unconfigured_host_is_refused_even_with_a_valid_session(gated):
+    gated.headers[CSRF_HEADER] = exchange(gated).json()["csrf_token"]
+    for host in ("other.example.com", "192.168.1.10"):
+        response = gated.get("/api/v1/workflow", headers={"Host": host})
+        assert response.status_code == 403, host
+        assert response.json()["error"]["code"] == "host_not_allowed", host
+
+
+# ── forwarded headers: believed from the proxy, ignored from anyone else ──────
+
+
+def test_a_trusted_proxys_https_makes_the_cookie_secure(gated):
+    """The proxy speaks HTTPS outward and HTTP to this app, so only the header knows."""
+    assert "secure" in exchange(gated, headers={"X-Forwarded-Proto": "https"}).headers[
+        "set-cookie"
+    ].lower()
+    assert "secure" not in exchange(gated).headers["set-cookie"].lower()
+
+
+def test_the_external_scheme_is_part_of_the_accepted_origin(gated):
+    gated.headers[CSRF_HEADER] = exchange(gated).json()["csrf_token"]
+    allowed = gated.post(
+        "/api/v1/albums/proposals",
+        json={},
+        headers={"X-Forwarded-Proto": "https", "Origin": f"https://{HOSTNAME}"},
+    )
+    assert allowed.status_code == 200
+    # The scheme is part of the origin: the same name over plain HTTP is not it.
+    refused = gated.post(
+        "/api/v1/albums/proposals",
+        json={},
+        headers={"X-Forwarded-Proto": "https", "Origin": f"http://{HOSTNAME}"},
+    )
+    assert refused.json()["error"]["code"] == "origin_not_allowed"
+
+
+def test_a_trusted_proxys_forwarded_host_is_the_host_that_is_judged(tmp_path):
+    """The proxy terminates the operator's hostname and dials this app by address."""
+    app = create_app(
+        config(
+            tmp_path,
+            allowed_hosts=(HOSTNAME,),
+            access_secret=SECRET,
+            trusted_proxies=(TESTCLIENT_ADDRESS,),
+        )
+    )
+    with raw_client(app, base_url="http://10.0.0.5") as client:
+        forwarded = {"X-Forwarded-Host": HOSTNAME}
+        assert exchange(client, headers=forwarded).status_code == 200
+        # Without the header the address it was dialled by is not an allowed name.
+        assert exchange(client).json()["error"]["code"] == "host_not_allowed"
+
+
+def test_forwarded_headers_from_an_untrusted_client_are_ignored(tmp_path):
+    """Otherwise any caller could declare the hostname and scheme of its choosing."""
+    app = create_app(config(tmp_path, allowed_hosts=(HOSTNAME,), access_secret=SECRET))
+    with raw_client(app, base_url="http://evil.example") as client:
+        forged = exchange(client, headers={"X-Forwarded-Host": HOSTNAME})
+        assert forged.json()["error"]["code"] == "host_not_allowed"
+
+    with raw_client(app, base_url=f"http://{HOSTNAME}") as client:
+        # A forged scheme would flip the cookie's Secure flag on a plain connection,
+        # which is how a cookie gets set and then never sent again.
+        response = exchange(client, headers={"X-Forwarded-Proto": "https"})
+        assert response.status_code == 200
+        assert "secure" not in response.headers["set-cookie"].lower()

tests/story_traceability.json

@@ -15,9 +15,10 @@
       "tests/characterization/test_webapp_query.py"
     ],
     "US01-02": [
-      "tests/unit/test_config.py",
+      "tests/integration/test_app_lifecycle.py",
       "tests/integration/test_migrations.py",
-      "tests/integration/test_app_lifecycle.py"
+      "tests/unit/test_config.py",
+      "tests/unit/test_env_file.py"
     ],
     "US01-03": [
       "tests/unit/test_path_policy.py",
@@ -169,6 +170,21 @@
     ],
     "US07-06": [
       "tests/integration/test_performance_budgets.py"
+    ],
+    "US07-07": [
+      "tests/e2e/test_release_gate.py",
+      "tests/e2e/test_release_journey.py"
+    ],
+    "US08-01": [
+      "tests/unit/test_security_policy.py",
+      "tests/integration/test_trusted_hosts.py"
     ]
-  }
+  },
+  "planned": [
+    "US08-02",
+    "US08-03",
+    "US08-04",
+    "US08-05"
+  ],
+  "_planned_comment": "Accepted backlog stories that are not implemented yet. The release gate (US07-07) requires every story file to be either mapped to tests or listed here, so an unimplemented story is a visible decision rather than a hole in the matrix."
 }

tests/unit/test_env_file.py

@@ -0,0 +1,69 @@
+"""Configuration from a dotenv file, including the archived CLI's variable names.
+
+An operator who already has a ``photo_analyzer.env`` should not have to rewrite it
+to run the application it was replaced by. The file is standing configuration; the
+shell is what you meant this time, so the shell always wins.
+"""
+
+from __future__ import annotations
+
+import os
+
+from photo_pipeline.config import Config, load_env_file, parse_env_file
+
+SAMPLE = """
+# The archived CLI's shape, comments and all.
+LLM_API_KEY=not-real
+LLM_BASE_URL="https://example.invalid/v1beta/openai/"
+LLM_MODEL='gemini-2.5-flash'
+LIBRARY=/tmp/pictures
+
+MAX_WORKERS=4
+# commented=ignored
+malformed line without an equals sign
+"""
+
+
+def test_the_file_is_parsed_and_never_executed():
+    values = parse_env_file(SAMPLE)
+
+    assert values["LLM_BASE_URL"] == "https://example.invalid/v1beta/openai/"  # quotes stripped
+    assert values["LLM_MODEL"] == "gemini-2.5-flash"
+    assert values["MAX_WORKERS"] == "4"
+    assert "commented" not in values and "malformed line without an equals sign" not in values
+
+
+def test_the_archived_cli_names_still_configure_the_application():
+    values = parse_env_file(SAMPLE)
+
+    assert values["OPENAI_API_KEY"] == "not-real"
+    assert values["OPENAI_BASE_URL"] == "https://example.invalid/v1beta/openai/"
+    assert values["PHOTO_PIPELINE_LIBRARY_ROOTS"] == "/tmp/pictures"
+
+
+def test_an_explicit_shell_variable_beats_the_file(tmp_path, monkeypatch):
+    path = tmp_path / "photo_analyzer.env"
+    path.write_text(SAMPLE)
+    monkeypatch.setenv("OPENAI_API_KEY", "from-the-shell")
+    monkeypatch.delenv("PHOTO_PIPELINE_LIBRARY_ROOTS", raising=False)
+
+    applied = load_env_file(path)
+
+    assert "OPENAI_API_KEY" not in applied, "the file overrode an exported value"
+    assert os.environ["OPENAI_API_KEY"] == "from-the-shell"
+    assert os.environ["PHOTO_PIPELINE_LIBRARY_ROOTS"] == "/tmp/pictures"
+    assert Config.from_env().library_roots[0].name == "pictures"
+
+
+def test_the_file_is_found_through_its_variable(tmp_path, monkeypatch):
+    path = tmp_path / "custom.env"
+    path.write_text("PHOTO_PIPELINE_PORT=9123\n")
+    monkeypatch.delenv("PHOTO_PIPELINE_PORT", raising=False)
+    monkeypatch.setenv("PHOTO_PIPELINE_ENV_FILE", str(path))
+    monkeypatch.chdir(tmp_path)  # no ./.env here, so only the variable can find it
+
+    assert Config.from_env().port == 9123
+
+
+def test_a_missing_file_is_not_an_error(tmp_path):
+    assert load_env_file(tmp_path / "nothing-here.env") == {}

tests/unit/test_security_policy.py

@@ -13,18 +13,31 @@ import pytest
 
 from photo_pipeline.api.security import (
     CSRF_HEADER,
+    LOOPBACK_HOSTS,
     PUBLIC_PATHS,
+    FailureLimiter,
     Session,
     evaluate,
+    exposed_hosts,
+    external_view,
     split_host,
+    trust_refusal,
 )
+from photo_pipeline.config import Config
 
 SESSION = Session(id="session-id", csrf_token="csrf-token")
 HOST = "127.0.0.1:8000"
 LIMIT = 1024
+HOSTNAME = "photos.example.com"
 
 
-def check(method="GET", path="/api/v1/workflow", **headers):
+def check(
+    method="GET",
+    path="/api/v1/workflow",
+    allowed_hosts=LOOPBACK_HOSTS,
+    scheme="http",
+    **headers,
+):
     """Evaluate a request that is authenticated and same-origin unless overridden."""
     sent = {
         "host": HOST,
@@ -38,6 +51,8 @@ def check(method="GET", path="/api/v1/workflow", **headers):
         path=path,
         headers=sent,
         session=SESSION,
+        allowed_hosts=allowed_hosts,
+        scheme=scheme,
         max_request_bytes=LIMIT,
     )
 
@@ -155,6 +170,115 @@ def test_refusals_name_no_path_secret_or_internal():
         assert "/" not in refusal.message
 
 
+# ── US08-01: the same table with a configured trust boundary ─────────────────
+
+CONFIGURED = frozenset(LOOPBACK_HOSTS | {HOSTNAME})
+
+
+def test_a_configured_host_is_accepted_and_its_neighbours_are_not():
+    assert check(host=HOSTNAME, allowed_hosts=CONFIGURED) is None
+    for host in ("other.example.com", f"evil-{HOSTNAME}", "192.168.1.10"):
+        refusal = check(host=host, allowed_hosts=CONFIGURED)
+        assert (refusal.status, refusal.code) == (403, "host_not_allowed"), host
+
+
+def test_the_loopback_default_refuses_a_host_nobody_configured():
+    """The default set is what the app enforced before there was a setting."""
+    refusal = check(host=HOSTNAME)
+    assert (refusal.status, refusal.code) == (403, "host_not_allowed")
+
+
+def test_the_origin_must_match_the_external_scheme():
+    for scheme in ("http", "https"):
+        assert (
+            check(
+                method="POST",
+                host=HOSTNAME,
+                origin=f"{scheme}://{HOSTNAME}",
+                allowed_hosts=CONFIGURED,
+                scheme=scheme,
+            )
+            is None
+        )
+    # An HTTPS deployment whose caller claims plain HTTP is a different origin.
+    refusal = check(
+        method="POST",
+        host=HOSTNAME,
+        origin=f"http://{HOSTNAME}",
+        allowed_hosts=CONFIGURED,
+        scheme="https",
+    )
+    assert (refusal.status, refusal.code) == (403, "origin_not_allowed")
+
+
+def view(client, *, trusted=(), **headers):
+    sent = {name.replace("_", "-"): value for name, value in headers.items()}
+    return external_view(
+        client=client,
+        headers={"host": HOST, **sent},
+        scheme="http",
+        trusted_proxies=frozenset(trusted),
+    )
+
+
+def test_forwarded_headers_are_ignored_without_a_trusted_proxy():
+    forged = {"x_forwarded_proto": "https", "x_forwarded_host": HOSTNAME}
+    assert view("10.0.0.9", **forged) == ("http", HOST)
+    assert view(None, **forged) == ("http", HOST)
+    # Configuring *a* proxy does not trust a caller that is not it.
+    assert view("10.0.0.9", trusted=("10.0.0.1",), **forged) == ("http", HOST)
+
+
+def test_a_trusted_proxy_defines_the_external_scheme_and_host():
+    assert view(
+        "10.0.0.1", trusted=("10.0.0.1",), x_forwarded_proto="https", x_forwarded_host=HOSTNAME
+    ) == ("https", HOSTNAME)
+    # A chain: the first entry is what the original client asked for.
+    assert view(
+        "10.0.0.1",
+        trusted=("10.0.0.1",),
+        x_forwarded_proto="https, http",
+        x_forwarded_host=f"{HOSTNAME}, inner.internal",
+    ) == ("https", HOSTNAME)
+    # Trusted but silent: this hop's own view stands.
+    assert view("10.0.0.1", trusted=("10.0.0.1",)) == ("http", HOST)
+
+
+@pytest.mark.parametrize(
+    "settings,exposed",
+    [
+        ({}, []),
+        ({"host": "127.0.0.1"}, []),
+        ({"allowed_hosts": ("localhost", "127.0.0.1")}, []),
+        ({"allowed_hosts": (f"{HOSTNAME}:8443",)}, [HOSTNAME]),
+        ({"host": "0.0.0.0", "allowed_hosts": (HOSTNAME,)}, ["0.0.0.0", HOSTNAME]),
+    ],
+)
+def test_exposed_hosts_names_only_what_another_machine_can_reach(settings, exposed):
+    assert exposed_hosts(Config(**settings)) == exposed
+
+
+def test_an_exposed_configuration_without_a_secret_must_not_serve():
+    refusal = trust_refusal(Config(allowed_hosts=(HOSTNAME,)))
+    assert HOSTNAME in refusal and "PHOTO_PIPELINE_ACCESS_SECRET" in refusal
+    assert trust_refusal(Config(allowed_hosts=(HOSTNAME,), access_secret="s")) is None
+    # Loopback-only, with and without a secret, is unchanged.
+    assert trust_refusal(Config()) is None
+    assert trust_refusal(Config(access_secret="s")) is None
+
+
+def test_failed_attempts_are_bounded_per_window():
+    limiter = FailureLimiter(limit=2, window=60.0)
+    assert not limiter.blocked()
+    limiter.record_failure()
+    assert not limiter.blocked()
+    limiter.record_failure()
+    assert limiter.blocked()
+    # Attempts age out, so a locked-out operator is not locked out forever.
+    limiter._failures = [-120.0, -120.0]
+    assert not limiter.blocked()
+
+
 @pytest.mark.parametrize(
     "value,expected",
     [