[dgageot]

Why

terminalManager split cleanup into separate stop/wait calls and cached the result in done/closeErr fields. That split ownership made the lifecycle harder to reason about and race-prone: callers had to coordinate stop-then-wait themselves, and the cached fields could be read/written from multiple goroutines without a clear single owner.

What

Collapses the split stop/wait into a single cleanup(ctx) call, and makes Session.close the sole owner of the cleanup result:

• terminalManager drops the done/closeErr fields in favor of a single cleanup(ctx) method.
• Session.close buffers the cleanup result locally, retires/cancels immediately, and blocks on the manager cleanup — while still running terminal/runtime/MCP shutdown in parallel.
• Callers and tests migrated accordingly, including fixture ownership in terminal_test.go and tracing_test.go.
• Added terminal_cleanup_test.go with race-focused coverage: synchronous retirement, simultaneous first closers, exactly-one final release, cached failure results, late terminal creation, and trace context propagation.

Validation

• task build, task test, task lint all pass.
• go test ./pkg/acp/... -race -count=1 run 3x, all green (no flakes).
• Cross-compilation (go vet) verified for linux/amd64 and windows/amd64.