Act on the wj architectural review: races, limits, pool and contract gaps

Findings F2, F4, F5, F6, F7, F8, F9, F10, F11 and F13 from
wj-architectural-review.md. F1/F1a (the non-sargable LOWER pattern) and F3
(the warranty sync holding a pooled connection) are being handled separately
and are untouched here.

- A collector check-in that raced another for the same PC lost everything:
  the unique index on the application fingerprint raised, nothing caught it,
  and the whole check-in rolled back, discarding state, events and hardware,
  answering the agent with form copy meant for a person. The insert now runs
  in a savepoint, reconciles what the other transaction already wrote, one row
  per savepoint so a second collision skips a row rather than the batch, and
  the check-in commits with a warning. This is the shape the delete side of
  the same function already had.
- Enrollment is deliberately unauthenticated, so it now carries a per-address
  rate limit and a cap on pending requests, and the cap is a setting an admin
  can see and change. A flood can no longer bury the approval queue or grow
  the table without bound.
- The fleet alert poll no longer formats debug strings on every request.
- BusinessUnit is on the contract surface (0.34.0), so the endpoints plugin
  stops reaching into a core table with raw SQL.
- A test walks every asset type descriptor and proves the endpoint it names
  really refuses a caller without the declared permission.
- The plugin settings routes now need a permission; they were telling anyone
  which modules a site runs and where its components live.
- The connection pool is sized explicitly with a short timeout, and the
  arithmetic against the server's connection ceiling is written next to the
  worker count instead of being left to defaults nobody chose.
- Fleet alerts carry both the correct lowercase keys and the old camelCase
  pair. Trays older than 1.7.0 match keys case-sensitively and would have gone
  quiet on every PC not yet upgraded; the old pair goes when they are.
- The dev server proxy points at localhost again, with environment variables
  for the remote-box workflow, and one developer's address is out of the
  server certificate page.

shopdb/config.py

@@ -10,6 +10,18 @@ class ConfigError(Exception):
     """Raised when required configuration is missing or unsafe."""
 
 
+# Connection pool per worker process. Env-driven so a site can size it against
+# its own MySQL max_connections without editing code; see the arithmetic in
+# docker-compose.yml next to the worker count.
+#
+# Deliberately NOT named SQLALCHEMY_POOL_SIZE and friends: those were
+# Flask-SQLAlchemy 2.x config keys, removed in 3.x, and reusing the names would
+# suggest the extension still reads them. It reads SQLALCHEMY_ENGINE_OPTIONS.
+DB_POOL_SIZE = int(os.environ.get('DB_POOL_SIZE', 10))
+DB_MAX_OVERFLOW = int(os.environ.get('DB_MAX_OVERFLOW', 10))
+DB_POOL_TIMEOUT = int(os.environ.get('DB_POOL_TIMEOUT', 5))
+
+
 def _required_env(varname):
     """Read an env var; raise ConfigError if missing or empty."""
     value = os.environ.get(varname)
@@ -32,9 +44,32 @@ class Config:
         'mysql+pymysql://root:password@localhost:3306/shopdb_flask',
     )
     SQLALCHEMY_TRACK_MODIFICATIONS = False
+
+    # Connection pool, sized on purpose rather than left to SQLAlchemy.
+    #
+    # The defaults (pool_size 5, max_overflow 10, pool_timeout 30) were never
+    # chosen: they simply applied, and nothing in the repo related them to the
+    # gunicorn worker count or to the MySQL max_connections ceiling. The
+    # arithmetic that ties the three together is written next to the worker
+    # count in docker-compose.yml.
+    #
+    # Under gevent a pooled connection is held for the duration of whatever the
+    # greenlet is doing, so a request that sits inside a slow external HTTP call
+    # WITH THE SESSION OPEN pins a connection for that whole call. Hundreds of
+    # greenlets can enter such a path at once and drain the pool, at which point
+    # every other request blocks on checkout. Hence the fail-fast timeout: 5
+    # seconds, not 30. A pool that is empty for five seconds is not going to be
+    # fine in thirty, and a fast 500 is diagnosable where a 30-second hang looks
+    # like a dead database.
+    #
+    # See PLUGIN-DEVELOPMENT / the plugin docs: a plugin must not hold a DB
+    # session across external I/O.
     SQLALCHEMY_ENGINE_OPTIONS = {
         'pool_pre_ping': True,
         'pool_recycle': 300,
+        'pool_size': DB_POOL_SIZE,
+        'max_overflow': DB_MAX_OVERFLOW,
+        'pool_timeout': DB_POOL_TIMEOUT,
         # Pin the client connection to utf8mb4 so reads/writes match the
         # utf8mb4 schema regardless of the server default charset. TestingConfig
         # (SQLite) replaces this whole dict, so the charset arg never reaches it.
@@ -192,6 +227,13 @@ class TestingConfig(Config):
     # Off by default so login-heavy fixtures do not trip the limiter; tests
     # that exercise rate limiting flip it on via app.config override.
     AUTH_RATELIMIT_ENABLED = False
+    # Same reason: the enrollment suites post to /enroll dozens of times, and
+    # the limiter counts in a process-wide cache that outlives a test, so with
+    # it on by default a budget spent in one module surfaces as an unrelated
+    # failure in another - and which modules share a worker depends on the
+    # xdist distribution, so the failure moves around. The tests that are about
+    # the limiter turn it on themselves.
+    ENDPOINTS_ENROLL_RATELIMIT_ENABLED = False
     SQLALCHEMY_DATABASE_URI = 'sqlite:///:memory:'
     SQLALCHEMY_ENGINE_OPTIONS = {
         'connect_args': {'check_same_thread': False},