docs: document connection limit, async fetch, and connection pool

Update USAGE, CHEATSHEET, README, PROJECT, and TASKS for the three
performance improvements: max_connections semaphore, async HTTP
source fetching, and first-hop TCP connection pool.

docs/USAGE.md

@@ -43,6 +43,9 @@ listen: 127.0.0.1:1080
 timeout: 10
 retries: 3
 log_level: info
+max_connections: 256      # concurrent connection limit (backpressure)
+pool_size: 0              # pre-warmed TCP connections to first hop (0 = disabled)
+pool_max_idle: 30         # max idle time for pooled connections (seconds)
 
 chain:
   - socks5://127.0.0.1:9050
@@ -245,9 +248,10 @@ Settings reloaded on SIGHUP:
 | `timeout` | Per-connection timeout |
 | `retries` | Max retry attempts |
 | `log_level` | Logging verbosity |
+| `max_connections` | Concurrent connection limit |
 | `proxy_pool.*` | Sources, intervals, thresholds |
 
-Settings that require a restart: `listen`, `chain`.
+Settings that require a restart: `listen`, `chain`, `pool_size`, `pool_max_idle`.
 
 Requires `-c` / `--config` to know which file to re-read. Without a
 config file, SIGHUP is ignored with a warning.
@@ -304,6 +308,40 @@ curl --proxy socks5h://127.0.0.1:1080 https://example.com
 
 Note: use `socks5h://` (not `socks5://`) with curl to send DNS through the proxy.
 
+## Connection Limit
+
+s5p caps concurrent connections with an `asyncio.Semaphore`. When all
+slots are taken, new clients backpressure at TCP accept (the connection
+is accepted but the handler waits for a slot).
+
+```yaml
+max_connections: 256    # default
+```
+
+```bash
+s5p -m 512   # override via CLI
+```
+
+The `max_connections` value is reloaded on SIGHUP. Connections already
+in flight are not affected -- new limits take effect as active connections
+close.
+
+## First-Hop Connection Pool
+
+Pre-warms TCP connections to the first hop in the chain, avoiding the
+TCP handshake latency on each client request. Only the raw TCP
+connection is pooled -- once SOCKS/HTTP CONNECT negotiation begins,
+the connection is consumed.
+
+```yaml
+pool_size: 8          # 0 = disabled (default)
+pool_max_idle: 30     # seconds before a pooled connection is evicted
+```
+
+Connections idle longer than `pool_max_idle` are discarded and replaced.
+A background task tops up the pool at half the idle interval. Requires
+at least one hop in `chain` -- ignored if chain is empty.
+
 ## Chain Order
 
 Hops are traversed left-to-right: