feat: LinkedIn profile search engine (Django + React + Elasticsearch + ELK)

Full-stack implementation of the Cyberyan full-stack trial task.

Backend (Django 5 / DRF)
- Normalised PostgreSQL model (Profile, Company, School, Skill, Experience,
  Education) as the source of truth, with GIN + trigram indexes.
- Self-healing CSV reader: the supplied export carries an injected watermark
  column, missing columns, rows split by broken quoting and a duplicated
  header. A signature-scored Needleman-Wunsch alignment reconstructs 284 clean
  records out of 336 physical rows; only 10 rows parse correctly without it.
- Pluggable search layer (Strategy): Elasticsearch by default, PostgreSQL
  full-text search as a fallback selected by a single environment variable.
- Keyword search with per-field boosts, fuzziness and phrase boosting, six
  faceted filters (OR within a filter, AND across filters), an experience
  range, four sort orders, highlighting and facet aggregations.
- JWT auth with rotation + blacklist, three roles, PII masking, scoped rate
  limits, a single error envelope, OpenAPI schema and health probes.

Frontend (React 18 / TypeScript / Vite)
- Debounced search with type-ahead, faceted filter panel, result cards,
  profile drawer, URL-synced search state, automatic token refresh.
- Highlight snippets rendered without dangerouslySetInnerHTML.

Infrastructure
- Hardened multi-stage images (non-root, read-only rootfs, healthchecks).
- docker-compose for the full stack plus dev and production overlays.
- ELK: structured JSON logs shipped asynchronously to Logstash, routed per
  event type, with a composable index template and a Kibana setup script.

Quality
- 61 backend tests (85% coverage), ruff/black/bandit/mypy, ESLint + tsc,
  GitHub Actions running lint, tests and both image builds.
- Persian documentation: README plus architecture, search, API, data,
  security and operations guides.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GnFo6fNaHCTAkJbfcdkqtG

backend/gunicorn.conf.py

@@ -0,0 +1,40 @@
+"""
+Gunicorn configuration.
+
+Sized for a container: worker count follows the CPU quota, requests are
+recycled to bound memory growth, and access logs are emitted as JSON so they
+land in Elasticsearch in the same shape as the application logs.
+"""
+
+import multiprocessing
+import os
+
+bind = os.getenv("GUNICORN_BIND", "0.0.0.0:8000")
+workers = int(os.getenv("GUNICORN_WORKERS", str(min(multiprocessing.cpu_count() * 2 + 1, 8))))
+threads = int(os.getenv("GUNICORN_THREADS", "2"))
+worker_class = "gthread"
+timeout = int(os.getenv("GUNICORN_TIMEOUT", "45"))
+graceful_timeout = 30
+keepalive = 5
+
+# Recycle workers to keep any slow leak from growing without bound.
+max_requests = 1000
+max_requests_jitter = 100
+
+preload_app = False
+forwarded_allow_ips = os.getenv("GUNICORN_FORWARDED_ALLOW_IPS", "*")
+
+# Refuse absurd request lines/headers early - cheap protection against a class
+# of denial-of-service probes.
+limit_request_line = 8190
+limit_request_fields = 100
+limit_request_field_size = 8190
+
+accesslog = "-"
+errorlog = "-"
+loglevel = os.getenv("GUNICORN_LOG_LEVEL", "info")
+access_log_format = (
+    '{"level":"INFO","logger":"gunicorn.access","message":"%(r)s",'
+    '"http_status":%(s)s,"duration_ms":%(D)s,"client_ip":"%(h)s",'
+    '"request_id":"%({x-request-id}i)s"}'
+)