Migrate MCP SDK: mark3labs/mcp-go -> official go-sdk via new abstraction layer

[LackOfMorals]

Summary

• The MCP spec and its SDKs move fast, including breaking changes, and this codebase had zero abstraction around mark3labs/mcp-go — SDK types were spelled out directly in the server, every tool spec/handler, and every test tier.
• Introduces internal/mcpsdk as the single package allowed to import an MCP SDK (plus internal/mcpsdk/mcpsdktest, a matching client wrapper used only by this repo's own test suites), and migrates the server, all five tools, and every test tier onto it.
• mark3labs/mcp-go is fully removed from go.mod; the server and tools are now backed by the official github.com/modelcontextprotocol/go-sdk, which was already an indirect dependency of this repo.
• Fixes 4 pre-existing golangci-lint findings (ElementId -> ElementID naming in internal/database/json_tagged_values.go) found while linting the branch.
• Adds .github/workflows/test.yml to actually run the unit/integration/e2e suites in CI — previously no workflow ran tests at all.
• Adds a Changie fragment (.changes/unreleased/mcp-sdk-abstraction.yaml) documenting the change for the next release's changelog.

Two real behavioral gaps were found and fixed as part of the migration (not just mechanical renames):

• The official SDK's stdio transport returns an error (rather than nil) when stdin is closed/exhausted, which broke stdio-mode tests that relied on the old SDK swallowing that case — mcpsdk.Server.ServeStdio now treats a closed/EOF stdin as a clean shutdown, matching prior behavior.
• CallToolRequest.BindArguments could panic on a nil Params — hardened with a nil check.

Test plan

• go build ./..., go vet ./..., go test ./... all pass
• golangci-lint run reports 0 issues
• go build -tags=e2e ./test/e2e/... and go build -tags=integration ./test/integration/... compile cleanly
• task test:int / task test:e2e against a real Neo4j (needs Docker for testcontainers) — not run in this environment, recommended before merge
• task build produces a working binary (--version reports correctly)

🤖 Generated with Claude Code

[LackOfMorals]

Root-caused the CI failure: TestServerLifecycle in test/integration/server_test.go polled a fixed 4s window for s.Start() to fail rather than waiting on the goroutine itself. The invalid-host/invalid-database cases only resolve once the driver's connection/DNS-resolution attempt actually times out, which can take longer over the CI runner's network than locally — so the test saw a false nil at the 4s mark. This file has no MCP SDK dependency; it's a pre-existing flake that had never been exercised in CI until this PR added a workflow that actually runs the test suites. Fixed in 82e2fe5 by waiting on Start()'s real result via a channel, with a 30s safety net for the network-timeout cases. Note the failing integration step also meant the e2e step never got a chance to run in that CI attempt — worth re-checking once this re-runs.

[LackOfMorals]

CI is now green (run https://github.com/neo4j-labs/neo4j-mcp-canary/actions/runs/34537845656) — unit, integration, and e2e all pass. Two more pre-existing issues surfaced and fixed along the way, both unrelated to the MCP SDK migration itself (neither file has an MCP SDK dependency), just never exercised in CI until this PR added a workflow that runs the suites:

1. test/integration/server_test.go's invalid-host/invalid-database case raced the neo4j driver's own default 30s MaxTransactionRetryTime (an unreachable host is classified as a transient/retryable error, so the driver quietly retries for its full default budget) — fixed by giving that test's driver a short 3s retry budget instead of padding the external wait further (verified locally: failure now surfaces in ~6s).
2. test/e2e/server_initialization_test.go's read-only-mode tool-count assertions (3/4) predated the give-feedback tool (added 2026-08-19) and were never updated — corrected to 4/5 to match what internal/server/tool_register_test.go already asserts.