Proposal: opt-in server request admission and response budgets #319

After verifying #266's stdio serialization and #298's early-cancellation changes,
there is a separate resource-control need for hosts that expose deterministic tools
with bounded input and output contracts. A bounded input frame alone does not bound
SDK-created request tasks, queued calls, or an aggregate JSON-RPC batch response.

Would an opt-in configuration surface along these lines fit the SDK?

Server.Configuration(
    strict: true,
    maxConcurrentRequests: 4,
    maxQueuedRequests: 8,
    maxBatchItems: 8,
    maxResponseBytes: 65_535,
    maxPendingControlResponses: 64)

// Separate proposed transport configuration, naming still to be agreed:
StdioTransport.Configuration(
    maxInboundFrameBytes: 16_384,
    maxOutboundFrameBytes: 65_536,
    maxBufferedMessages: 1,
    maxPendingWrites: 4)

The numbers are host policy examples. Existing callers would retain unlimited
execution/batch/response defaults. This is a distinct proposal from #266/#298 and
would benefit from agreement on the contracts before an upstream patch:

• Reserve request ownership and admission before any suspending context lookup.
  Queue request values without allocating handler tasks. Return explicit overload
  errors when capacity is full. A separately bounded control-reply queue keeps
  reception available for cancellation while those errors await output. Count
  cancelled active work until it actually exits; retain correlation IDs through
  delayed batch delivery, separately from execution slots.
• Allow cancellation notifications through while admitted work executes. Remove
  cancelled queued calls before their handlers start. Scope correlation/admission
  by the appropriate Server/session boundary; preserve HTTP transport routing
  rather than assuming a bare wire ID is globally unique.
• Bound both per-item responses and the encoded aggregate batch envelope. Preserve
  IDs in explicit overflow errors; close the connection when even an error cannot
  fit. Never truncate a success payload. Count bytes before transport framing.
• Retain completed batch results separately under a bounded accumulator, so a batch
  larger than the execution limit can finish. The final item holds a delivery slot
  while sending the aggregate. Define how configuration changes affect work already
  admitted; immutable per-connection limits would make the contract simpler.
• Define host-driven shutdown completion and EOF cleanup. Waiting is cooperative;
  process deadlines remain a host responsibility. Avoid a handler awaiting a
  shutdown operation that itself waits for that handler.

The transport configuration should reject an oversized frame before accumulating
beyond its bound, count the delimiter consistently, bound pending/buffered messages,
and preserve FIFO writes through partial I/O. Backpressure must avoid silently
dropping a cancellation/control frame. A partial write failure needs an explicit
connection-failure policy before another frame may be written. These transport
APIs are proposed separately from Server admission; stock stdio presently remains
an unbounded input producer. Neither layer bounds arbitrary handler allocations
or notification-handler work. No networking/domain policy
belongs in these APIs.

A local development prototype based on release 0.12.1 and tests establish feasibility: the full native
macOS SDK suite passes 573 tests, and 28 focused native ARM64 Linux tests cover
4 active / 8 queued admission, overload, queued/running cancellation, retained
capacity during response delivery, batch8 through slots4, cancellation during a
batch, aggregate/individual budgets, too-small error budgets, negotiated batch
eligibility, EOF, and cooperative cleanup. This prototype is unpublished and is
not an upstream API commitment. No new dependency or broad protocol rewrite is
required for the stdio pilot. Full Linux test-target qualification remains blocked
by unrelated baseline URLProtocol helper diagnostics, separately documented.