Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@docs/plugins/index.mdx`:
- Around line 847-853: Update the docs to recommend a buffer so
REQUEST_TIMEOUT_SECONDS is strictly greater than your longest plugin timeout
(plugin.timeout) to avoid the race that causes write EPIPE; explicitly state
"REQUEST_TIMEOUT_SECONDS should be at least X seconds greater than the longest
plugin.timeout (e.g., +1–2s buffer)" and apply the same change to the other
occurrence referenced (around the lines noted).

In `@plugins/lib/worker-pool.ts`:
- Around line 107-108: The outer worker-pool uses a fixed DEFAULT_TASK_TIMEOUT
(DEFAULT_PLUGIN_TIMEOUT_MS + 5000) which ignores per-plugin request.timeout,
causing plugins with timeouts > DEFAULT_PLUGIN_TIMEOUT_MS to be cut off; update
the code that sets taskTimeout (the value passed into the Piscina safety timeout
/ Promise.race) to derive from the effective per-request timeout (use
request.timeout when present, otherwise DEFAULT_PLUGIN_TIMEOUT_MS), then add the
5000ms cleanup buffer and use that computed value instead of
DEFAULT_TASK_TIMEOUT when invoking Piscina.runTask / Promise.race so the pool
honors per-plugin timeouts.

In `@src/services/plugins/pool_executor.rs`:
- Around line 996-1000: The response_timeout calculation (using
response_timeout, timeout_secs, DEFAULT_PLUGIN_TIMEOUT_SECONDS) adds a fixed 5s
buffer but still can be shorter than the actual worker-side wait, causing
callers to get a generic timeout before the worker can return the structured
504/TIMEOUT; update the response_timeout computation to match or exceed the
worker wait by using the same buffer/delay logic the worker uses (replace the
hardcoded + Duration::from_secs(5) with the same buffer constant or value used
by the worker, or compute response_timeout =
(timeout_secs.unwrap_or(DEFAULT_PLUGIN_TIMEOUT_SECONDS) + WORKER_BUFFER_SECS +
QUEUE_BUFFER_SECS) seconds) and apply the same change to the other occurrence
referenced in the comment (the block around the second occurrence at the later
lines) so both code paths use the identical timeout math.
- Around line 1096-1118: The extra broad string check in is_dead_server_error is
preventing DeadServerIndicator::from_error_str from detecting transport-level
timeouts (e.g., PluginExecutionError("connection timed out")); remove the
(lower.contains("plugin") && lower.contains("timed out")) clause and only
special-case the explicit Node.js phrase "handler timed out" so that
connection/transport timeout messages fall through to
DeadServerIndicator::from_error_str for proper dead-server detection.
- Around line 545-549: The current code adds PLUGIN_TIMEOUT_BUFFER_SECONDS into
the timeout passed to conn.send_request_with_timeout, causing ScriptTimeout to
report the expanded wall-clock time; instead pass the original configured
timeout (timeout_secs.unwrap_or(DEFAULT_PLUGIN_TIMEOUT_SECONDS)) to
send_request_with_timeout so callers see the intended timeout, and implement the
backstop separately (e.g., run a parallel watchdog using
PLUGIN_TIMEOUT_BUFFER_SECONDS that forcibly terminates the Node process if the
plugin hangs) so the buffer is used only for enforcement and not reflected in
the reported PluginError; update the call site where timeout is computed and the
call conn.send_request_with_timeout(&request, timeout) to use the
configured_timeout and add a watchdog/select-based backstop surrounding that
call.