mcp: bound tool_name metric label to registered tools

[a-palamarchuk]

What

The MCP server's metrics middleware labels mcp_tool_invocations_total,
mcp_tool_execution_duration_ns and mcp_tool_concurrent_executions with
tool_name, taken verbatim from the tools/call request. Tool names are chosen
by the client, so every call naming a tool that doesn't exist creates new series
in all three metrics. That can happen through a typo, a tool name hallucinated by
an LLM, or a deliberate scan, and the series count grows without bound
(CONTRIBUTING §1.2.1: metrics should "avoid excessive cardinality").

This change has the resources wrapper record the tool names it registers, and the
metrics middleware reports calls to any other name under tool_name="unknown" (the
value it already used as a fallback).

Reproduce

redpanda-connect mcp-server --address … --observability-address … with one
processor tool (upper) and prometheus metrics. Then make 50 tools/call
requests with random nonexistent names and one call to upper.

• Before: /metrics has 51 distinct tool_name series per metric (one per random
  name, plus upper).
• After: exactly two, tool_name="unknown" (counter = 50) and tool_name="upper".

Clients still receive the same -32602 unknown tool JSON-RPC error; only the
metric labelling changes.

Design notes

• The go-sdk server has no public "is this tool registered" lookup, and inferring
  "unknown tool" from the handler's error would be fragile and would happen too
  late for the concurrency gauge, which is incremented before the call. So
  ResourcesWrapper, the only place tools are registered, now routes every
  registration through a small addTool helper and exposes HasTool.
• HasTool is guarded by an RWMutex. Today all tools are registered before the
  server starts serving, but the lock keeps this race-free if tools are ever added
  while serving.
• A real tool literally named unknown would share its series with unregistered
  calls. That label value was already this code's fallback, so I kept it rather
  than changing what existing dashboards see.

Tests

• New TestToolMetricsCollapseUnknownToolNames (internal/mcp/metrics, which had
  no tests before) builds resources with the prometheus exporter on an
  in-memory mux, sends tools/call requests through the middleware for one
  registered and two unregistered names, then scrapes /metrics. It asserts the
  registered tool has its own series, the two unregistered calls are counted
  under tool_name="unknown", and no requested made-up name appears anywhere.
  With the new check disabled it fails and prints the scrape showing the
  per-name series.
• New TestResourcesWrapperHasTool (internal/mcp/tools) covers the derived cache
  tool names (get-/set-), an MCP-enabled processor, a processor with MCP
  disabled, and an unregistered name.
• go test -race -shuffle=on ./internal/mcp/..., the in-process MCP integration
  tests (-run '^TestIntegrationMCP') and ./internal/cli/ pass;
  golangci-lint run and golangci-lint fmt --diff are clean.