Fix inference Task4 hint, stale docstring, Dockerfile port comment

- inference.py: replace broken CREATE-INSERT-DROP-RENAME hint for Task 4
with correct ALTER TABLE RENAME COLUMN instructions; DROP TABLE on
inventory/categories is an instant -1.0 safety violation
- server/environment.py: update reset() docstring from seed%3/tasks 1-3
to seed%5/tasks 1-5
- Dockerfile: fix comment example port from 8000 to 7860

Dockerfile

@@ -4,7 +4,7 @@
 #   docker build -t db-er .
 #
 # Run (server mode):
-#   docker run -p 8000:8000 db-er
+#   docker run -p 7860:7860 db-er
 #
 # Run (inference mode):
 #   docker run --rm \

inference.py

@@ -87,7 +87,7 @@ You must respond with a SINGLE JSON object -- no other text, no markdown fences.
 - Task 1 (Phantom Duplicates): Look for duplicate emails in users. Keep the OLDEST record per email (lowest id).
 - Task 2 (Cascading Failure): admin_audit_logs contains record_json with deleted product data. Parse it and re-INSERT.
 - Task 3 (Payroll Black Hole): The "timeout" is MISLEADING. Check employees_old for duplicate IDs (no PRIMARY KEY). Delete corrupted duplicates (salary=0, active=0), then migrate remaining employees.
-- Task 4 (Schema Drift): Columns were renamed. Use schema_changelog to find original names. Recreate tables with original column names using CREATE-INSERT-DROP-RENAME pattern.
+- Task 4 (Schema Drift): Columns were renamed. Use schema_changelog (old_column to new_column mapping). Reverse each rename with: ALTER TABLE <table> RENAME COLUMN <current_name> TO <original_name>. Do NOT use DROP TABLE - inventory and categories are protected core tables and will trigger instant -1.0 failure.
 - Task 5 (Referential Maze): Multi-table FK crisis. Check decommission_log before restoring anything. Use sync_audit for missing data. Projects 5 and 9 are decommissioned and must NOT be restored.
 
 Always respond with valid JSON. Never include explanations outside the JSON."""

server/environment.py

@@ -106,9 +106,9 @@ class DBEREnvironment(Environment[DBERAction, DBERObservation, DBERState]):
 
         Args:
             seed:       Controls task selection when *task_id* is not given.
-                        seed % 3 -> 0=task 1, 1=task 2, 2=task 3.
+                        seed % 5 -> 0=task 1, 1=task 2, 2=task 3, 3=task 4, 4=task 5.
             episode_id: Opaque identifier stored in state.
-            task_id:    Override task selection (1, 2 or 3).
+            task_id:    Override task selection (1-5).
         """
         #  1. Pick task 
         if task_id is not None: