Measure the PTY receive queue under slow, absent and burst consumers #184

What

Measure the terminal receive queue when the consumer reads slowly, stops reading, or receives a
burst: queue length, and retained and peak memory over time, for PtySession and
PersistentPtySession. Use a macro harness (a long-running scenario whose socket produces output
faster than it is read), not only BenchmarkDotNet's MemoryDiagnoser. Record the numbers with their
environment.

Why in 1.0

The receive queue is unbounded by contract, and ADR-0023 says a memory limit and an overflow policy
need workload evidence before that contract changes. The 1.0 contract keeps the queue unbounded only
if the measurement shows that this is acceptable; a default bound after 1.0 would be a behaviour
change. This measurement is that evidence. An opt-in bound, if one is wanted, follows after 1.0.

Evidence

• src/OpenCode.Sdk/Internal/TerminalReceivePump.cs:12: Channel.CreateUnbounded; :152: one
  16 KiB receive buffer per connection (src/OpenCode.Sdk/Internal/TerminalSocketBounds.cs:19).
• docs/adr/0023-terminal-consumer-cancellation.md:7-11.
• docs/ROADMAP.md:92-93 ("a bound for the terminal receive queue, which is unbounded today").
• tests/OpenCode.Sdk.Performance.Tests/Benchmarks/PtySessionReadBenchmarks.cs and
  PersistentPtySessionReadBenchmarks.cs measure receive and decode only.

Done when

• Numbers for the three consumer patterns on Windows (and on Linux if cheap), with the harness
  committed.
• A written conclusion: the unbounded default stands for 1.0, or a bound is needed, with the
  evidence; the contract sentence in client-runtime.md matches it.