[s2x]

Closes #275.

Summary

• TimeoutConfig::batchDeadlineMs now defaults to DEFAULT_BATCH_DEADLINE_MS (60000). An explicit 0 keeps the old unbounded behaviour (opt-out).
• When > 0, the constructor rejects a deadline below any of the seven per-call timeouts that run inside a batch fan-out (readTimeoutMs is excluded: only point reads use it).
• BatchAsyncExecutor::executeParallelCapped(): all windows now share one deadline budget, as its docblock promised. The deadline exception reports the configured deadline, the total elapsed time and every sub-batch without a result.
• RawKvScanner: the budget is shared across the windows of one region batch, and a deadline no longer falls back to the unbounded sequential scan.
• TxnKV: a deadline in the secondary-commit fan-out (after the commit point) is logged and commit() does not throw. A deadline in a rollback fan-out propagates and the transaction stays rollback-able.
• The [RC-5] A single request can block a worker indefinitely #184 regression tests are kept; the new tests are in RawKvBatchDeadlineEnforcementTest.
• Docs: configuration, operations, troubleshooting, error-handling, architecture, faq, decisions. The worst case is batchDeadlineMs plus one region's full retry loop.

BREAKING (CHANGELOG ### Changed, with a migration note): batches are now bounded by default. Set batchDeadlineMs to 0 to keep the old behaviour.

Not covered: the experimental BatchCommands path (RawKvBatch::tryMultiplex, off by default) has only a per-call limit.

Test plan

• composer lint
• composer test:unit (1973 tests)
• composer test:grpc (321 tests; the 1 warning also occurs on master)
• E2E in CI (local Docker is not available)