Add bounded privacy-aware Prometheus diagnostics

[YangYuS8]

What changed

• adds a lazy standard-library Prometheus /api/v1 client without adding third-party dependencies
• adds prometheus_server_info for bounded build and runtime diagnostics
• adds prometheus_target_list for bounded active-target health summaries
• adds prometheus_metric_snapshot for constrained instant metric snapshots
• exposes all three tools through the shared Registry, CLI, Ark Agent, and MCP server
• keeps OpsPilot on Go 1.26.5

Query boundary

prometheus_metric_snapshot does not accept arbitrary PromQL. OpsPilot generates the expression from:

• one validated ASCII metric name
• up to eight exact-match labels from a fixed diagnostic allowlist
• optional sum, avg, min, max, or count aggregation
• up to five allowlisted grouping labels
• a hard series limit

Regular expressions, range vectors, subqueries, offsets, functions, arbitrary labels, arbitrary paths, and user-supplied PromQL are not interfaces. Metric queries use Prometheus's URL-encoded POST form so matcher values do not appear in request URLs.

Configuration and transport safety

• initialization is lazy, so missing Prometheus configuration does not block MCP startup or unrelated tools
• HTTPS is required by default; HTTP requires OPSPILOT_PROMETHEUS_ALLOW_HTTP=true
• URL user information, query strings, fragments, redirects, and ambient proxies are rejected or disabled
• HTTPS requires TLS 1.2 or newer
• optional bearer authentication uses a bounded absolute token file that is read for every request to support rotation
• only fixed build-info, runtime-info, active-target, and instant-query endpoints are implemented
• configuration, flags, rules, alerts, label/series enumeration, administration, and arbitrary API paths are not exposed

Privacy boundary

Raw Prometheus API data is never returned. The projection omits:

• scrape and global URLs
• discovered labels
• arbitrary target and metric labels
• target error text
• runtime hostname and working directory
• API warning and info text
• raw API and HTTP error bodies
• bearer-token values

Target errors are represented by error_present. API warnings and infos are represented by counts only. Metric labels are restricted to job, instance, cluster, namespace, pod, container, node, service, and endpoint.

Bounds and response validation

• request timeout, Prometheus query timeout, response bytes, target count, and series count are bounded
• target and series limits are enforced locally even if the server returns more data
• only instant vector results are accepted
• malformed values, unexpected result types, oversized responses, non-finite timestamps, and timestamps outside the RFC3339 year range are rejected
• explicit zero and out-of-range tool limits are rejected rather than silently treated as defaults

Validation

• clean go mod tidy with no module-file changes
• clean gofmt
• go vet ./...
• go test -race -coverprofile=coverage.out ./...
• go build ./cmd/opspilot
• HTTP tests cover fixed endpoints, GET/POST method selection, POST form encoding, token headers, redirects, timeouts, cancellation, response limits, status errors, API errors, and safe URL prefixes
• privacy tests inject scrape URLs, credentials, discovered labels, arbitrary labels, target errors, runtime host paths, warnings, infos, and raw error details and verify they never enter results
• query tests cover deterministic selector generation, escaping, allowlists, aggregations, grouping, limits, label filtering, vector sorting, and invalid shapes
• timestamp tests cover fractional seconds, NaN, infinities, and out-of-range values
• Registry tests verify Prometheus tools remain lazily discoverable without configured credentials