diff --git a/CLAUDE.md b/CLAUDE.md
new file mode 100644
index 0000000..ee1accc
--- /dev/null
+++ b/CLAUDE.md
@@ -0,0 +1,106 @@
+# api-template
+
+FastAPI + async PostgreSQL starter: cookie-based JWT auth (FastAPI-Users) with
+refresh-token rotation, SQLAlchemy 2.0 + Alembic, slowapi rate limiting,
+structlog, OpenTelemetry, and security headers. The `notes` router is the example
+CRUD resource — copy it as the pattern for new resources. See `README.md` for the
+full tour.
+
+## How to run things
+
+```bash
+uv sync                          # install deps
+docker compose up -d db          # local Postgres
+uv run alembic upgrade head      # apply migrations
+uv run uvicorn app.main:app --reload
+
+# Checks (exactly what CI runs)
+uv run pytest                    # test suite (async; in-memory SQLite)
+uv run ruff check .              # lint
+uv run ruff format --check .     # format (drop --check to apply)
+
+# Migrations
+uv run alembic revision -m "describe change"   # stub (autogenerate can miss changes — review it)
+uv run alembic upgrade head
+```
+
+CI (`.github/workflows/ci.yml`) runs ruff + pytest on every PR to `main`. It does
+**not** deploy — see the hardening checklist.
+
+## Conventions
+
+- **Caching is opt-in (default `no-store`).** `cache_control_middleware` in
+  `app/main.py` defaults every response to `no-store`; an endpoint that serves
+  genuinely public, cacheable data opts in by setting its own `Cache-Control`.
+  Never blanket-cache authenticated or mutable responses.
+- **Rate limiting keys on the client IP via forwarded headers.**
+  `get_remote_address` reads `request.client.host`, which is the real client IP
+  only because `start.sh` runs uvicorn with `--proxy-headers`. Behind a proxy/CDN
+  that doesn't overwrite `X-Forwarded-For`, key on the CDN's real-client header
+  (e.g. `CF-Connecting-IP`) instead, or the whole audience shares one bucket.
+- **Tests build the schema with `Base.metadata.create_all` on SQLite — they never
+  run Alembic** (`tests/conftest.py`). A migration is only ever exercised against
+  real Postgres at deploy time. See hardening.
+- **Structured logging via structlog** (`app/logging.py`); JSON in prod. Log
+  exception *types*, not just messages (see hardening).
+
+## Production hardening — gotchas learned under real live-event load
+
+> Distilled from running a sibling service through a high-traffic live event.
+> These are the general, provider-agnostic ones — worth a read before shipping
+> anything derived from this template to production.
+
+- [ ] **Validate migrations against real Postgres before deploy.** The suite uses
+  in-memory SQLite via `metadata.create_all` and never runs Alembic, so a bad
+  DDL/backfill (a Postgres-only type, constraint timing, a wrong `down_revision`)
+  passes CI green and first fails at the deploy's `alembic upgrade head`. Spin up
+  a throwaway `postgres:16`, seed realistic pre-migration rows incl. edge cases,
+  `upgrade head`, assert data + constraints, and test the `downgrade`.
+- [ ] **Schema changes under a rolling deploy: expand → transition → contract,
+  never one destructive step.** A rolling deploy serves the old and new revision
+  simultaneously during the rollover, so never drop or rename a column the
+  still-running revision reads. Three phases: *expand* (add the column,
+  dual-write) → *transition* (move reads to it, keep the old column) → *contract*
+  (drop the old column once nothing reads it). Aim for zero 5xx per rollover.
+- [ ] **Startup readiness must not block on a remote dependency you might need to
+  redeploy *through*.** If a process `await`s a DB/upstream call before it binds
+  its port and reports healthy, a degraded dependency means the container never
+  goes ready — and the deploy that *fixes* that dependency can't roll out. Bind
+  the port first; do dependency loading in the background with retry.
+- [ ] **Log `type(e).__name__`, not just `str(e)`.** Many transport errors (httpx
+  timeouts/connect errors) stringify to `""`, producing un-triageable,
+  un-groupable log lines. Always capture the exception type.
+- [ ] **Keep full-table reads of static reference data off hot paths.** Reading a
+  rarely-changing reference table on every request puts a DB connection
+  *checkout* on the hot path — and under a small pool the checkout, not the
+  query, is what saturates you under load. Cache it in-process (TTL or
+  refresh-on-write).
+- [ ] **Don't let a graceful low-level fallback swallow the signal the caller
+  needs.** If a helper returns a "usable default" on both a transient failure and
+  a permanent one, a retry/backoff loop above it can't tell "retry soon" from
+  "this won't get better" and will churn. Surface the failure (raise, or return a
+  status); let the caller own the fallback *and* the cadence.
+- [ ] **Don't let a correctness-critical value depend on a single soft-failing
+  upstream with no floor.** If a denormalized/derived column is computed from one
+  upstream field that can come back empty, treat "empty" as a failure mode: encode
+  a static floor for the stable mappings, merge upstream over it, and emit a loud
+  signal when upstream returns nothing.
+- [ ] **Aggregate fan-out failures.** A loop over N entities that logs once per
+  failed item turns one upstream blip into an N-line log/alert storm. Aggregate
+  per cycle into a single summary (counts by status/type).
+- [ ] **Don't double-instrument one call path.** Enabling both ORM and driver
+  tracing (e.g. SQLAlchemy *and* asyncpg instrumentation) emits two spans per
+  query and multiplies trace volume against your backend's quota for no extra
+  signal. Pick one layer.
+- [ ] **`merge ≠ deploy` unless CI deploys.** This template's CI lints and tests
+  but does not ship — wire your own deploy step and confirm a merge actually
+  reaches production, or a merged "fix" can sit undeployed mid-incident.
+- [ ] **Don't gate a recovery step behind the step it recovers.** If a
+  cleanup/relief step (pruning, scaling, cache-clear) is chained after a step that
+  can fail under the very pressure the relief would ease, the failure disables its
+  own mitigation. Make recovery steps independent.
+- [ ] **Stacked migration PRs: don't rebase a code-only phase onto `main` if a
+  *later* stacked phase adds a migration.** Rebasing for a clean diff drops the
+  earlier phase's migration from the later branch's history, so its
+  `down_revision` resolves to the wrong head. Base the migration-bearing phase on
+  the earlier migration branch and cherry-pick the code-only phase onto it.
diff --git a/app/main.py b/app/main.py
index bc07cc8..7e54d7a 100644
--- a/app/main.py
+++ b/app/main.py
@@ -148,12 +148,24 @@ async def request_id_middleware(request: Request, call_next) -> Response:
 
 @app.middleware("http")
 async def cache_control_middleware(request: Request, call_next) -> Response:
-    """Set Cache-Control headers: no-store for auth paths, public caching for GETs."""
+    """Default every response to ``no-store``; caching is strictly opt-in.
+
+    A blanket ``public, max-age=...`` on all GETs is a foot-gun: it lets shared
+    caches and browsers store authenticated, per-user, or mutable data — a user
+    would not see their own edits until the TTL expired, and a shared/CDN cache
+    could serve one user's response to another. So the safe default is
+    ``no-store``. An endpoint that serves genuinely public, cacheable data opts
+    in by setting its own ``Cache-Control`` in the handler, e.g.
+    ``response.headers["Cache-Control"] = "public, max-age=60"`` — this
+    middleware leaves an already-set header intact.
+
+    CDN note: an edge cache needs its own cache rule to cache JSON at all, and a
+    cacheable response must avoid ``Vary: Cookie`` (it fragments the cache per
+    cookie value and silently defeats CDN coalescing).
+    """
     response = await call_next(request)
-    if request.url.path.startswith(f"{API_V1_PREFIX}/auth/"):
+    if "cache-control" not in response.headers:
         response.headers["Cache-Control"] = "no-store"
-    elif request.method == "GET" and response.status_code == 200:
-        response.headers["Cache-Control"] = "public, max-age=3600"
     return response