perf(session): count messages without cloning session history

[dgageot]

Three call sites in the runtime and server packages needed a message count from a session — including its child sessions — and all three did it by calling GetAllMessages() and taking len of the result. GetAllMessages clones every message from every session in the tree into a new slice just to answer that question, allocating O(N) heap on every call.

AllMessageCount replaces that pattern with a read-locked recursive walk that counts nodes without materialising any slice. The three call sites in pkg/runtime/loop.go, pkg/server/server.go, and pkg/server/session_manager.go are updated to use it.

The test file covers unit correctness on flat and deeply-nested trees, equivalence against GetAllMessages on random inputs, a concurrent-writer stress test for the lock path, and a benchmark across three fixture sizes (50 messages, 300 direct + 5×40 child, 2000 + 20×100 child). Before/after benchmark results against the three fixtures:

BenchmarkAllMessageCount/small_50msg
  before: 63808 B/op  58 allocs/op
  after:      0 B/op   0 allocs/op

BenchmarkAllMessageCount/medium_300direct+5x40child
  before: 798594 B/op  551 allocs/op
  after:       0 B/op    0 allocs/op

BenchmarkAllMessageCount/large_2000+20x100child
  before: 8356224 B/op  4197 allocs/op
  after:        0 B/op     0 allocs/op

GetAllMessages is untouched; callers that need actual messages continue to use it. The count preserves its system-message filter and sub-session selection, including items containing both a message and a sub-session.

Validation passed: affected-package build, tests, and scoped golangci-lint (0 issues) with CGO_ENABLED=0, plus go test -race ./pkg/session. Full task build, task test, and task lint were attempted locally but blocked by the unaccepted Xcode licence; the full tests also encountered an unavailable Docker Model Runner. Full-suite validation remains for CI. The figures above describe allocation savings on synthetic fixtures, not whole-product speedups.