diff --git a/SPEC.md b/SPEC.md
index ba21956..1d5f886 100644
--- a/SPEC.md
+++ b/SPEC.md
@@ -184,6 +184,5 @@ shipped since this list was written.) Add when someone asks.
    states, autoplay default, `run.sh`.
 7. Auth — login screen, then multi-user CLI-created accounts with
    per-user history + test suite.
-8. Later (see git log, `docs/CHANGES-2026-09.md`): provider layer with
-   hianime primary, SQLite storage, stats/events, 2026-09 security
-   audit batch.
+8. Later (see git log): provider layer with hianime primary, SQLite
+   storage, stats/events, 2026-09 security audit batch.
diff --git a/docs/CHANGES-2026-09.md b/docs/CHANGES-2026-09.md
deleted file mode 100644
index fc48ca7..0000000
--- a/docs/CHANGES-2026-09.md
+++ /dev/null
@@ -1,52 +0,0 @@
-# Code-audit fix batch (2026-09)
-
-Findings from the 2026-09-15 code audit, in shipping order. The systemd
-cutover was explicitly deferred — the process stays on run.sh's nohup
-model; only the restart mechanism's blast radius is fixed here.
-
-## 1. login handler blocks the event loop (HIGH)
-
-`login` is `async def`, so the anti-brute-force `time.sleep(0.5)` inside
-the lockout check runs ON the event loop: every wrong password stalls
-every in-flight request (all streams included) on the single worker.
-Fix: make the handler a plain `def` (uvicorn runs it on the threadpool);
-move the lockout logic into a side-effect-free `server/auth.py` so it is
-unit-testable without booting the app. Regression test measures
-`/api/cdn` latency while a wrong-password login is mid-sleep.
-
-## 2. run.sh global pkill (HIGH)
-
-`pkill -f "uvicorn .*app:app"` also matches the test server
-(`uvicorn server.app:app --host 127.0.0.1 --port 8399`): an ani-web
-restart kills running test sessions (and vice versa). Fix: PID file
-(`server.pid`) becomes the primary kill path; any pattern fallback is
-port-scoped to the real instance.
-
-## 3. lockout is per-username only (MEDIUM)
-
-Spreading failures across usernames bypasses the per-username lockout
-entirely (and anyone can lock a known account for 15 min). Add a
-per-IP failure budget (`ANIWEB_IP_FAILS` / window) on top of the
-per-username lockout. Client IP = `request.client.host`; the reverse
-proxy's `X-Forwarded-For` is consulted only when `ANIWEB_TRUST_XFF=1`
-(set by run.sh), so a spoofed header can't widen or narrow the gate.
-
-## 4. login is exempt from the CSRF header check (MEDIUM)
-
-The custom-header check that makes cross-site form POSTs fail excludes
-`/login`. The exemption is load-bearing only by accident (JSON body
-content-type); login becomes covered by the same check as every other
-mutation, and login.html sends the header.
-
-## 5. dependencies unpinned (MEDIUM)
-
-`curl_cffi` version churn = impersonation churn = silent upstream 403
-waves. requirements.txt becomes a full freeze; bump deliberately.
-
-## 6. live-upstream e2e gated off by default
-
-The Playwright tests scrape and play real provider content — they go red
-whenever a provider rotates or a title is taken down (observed repeatedly).
-A suite that fails for non-regressions can't gate anything, so the default
-run is now hermetic only; run the live e2e deliberately with
-`ANIWEB_LIVE_TESTS=1`.