M168: Fault injection and hostile-endpoint testing for the socket transport

[Copilot]

Implements Milestone 168: extend Phase 1's hostile-input testing tradition to the M162 out-of-process transport boundary. Queue-full, out-of-order/duplicate KV appends, DMA descriptor rejection, endpoint crash mid-decode, and malformed completions must all map to att1_status_t cleanly, without UB.

New fault-injection test suite

• tests/test_aimu_endpoint_fault_injection.c exercises five scenarios against a real att1-aimu-endpoint daemon process (plus one hostile mock daemon):
  • Queue-full over the socket transport
  • Out-of-order and duplicate KV appends
  • Malformed DMA descriptors (unaligned address, unsupported flags, zero length)
  • Daemon crash mid-decode (SIGKILL between submit and poll)
  • A minimal hostile mock daemon that speaks just enough of the wire protocol to send back deliberately malformed completions (oversized kv_key_bytes/kv_value_bytes/payload_bytes claims)
• tools/att1-aimu-endpoint.c gains --cmd-ring-depth/--comp-ring-depth CLI overrides so queue-full doesn't require filling the 64-deep default ring.

Two real defects found and fixed

Writing the crash-mid-decode and malformed-completion cases surfaced actual bugs, not just missing coverage:

• SIGPIPE kills the client on daemon crash. src/aimu_endpoint_protocol.c used a plain write(); once the peer had crashed, that raises SIGPIPE, whose default disposition terminates the client process instead of returning ATT1_ERR_IO. Fixed by switching to send(..., MSG_NOSIGNAL).
• Out-of-bounds read from a hostile/buggy daemon response. client_kv_read(), client_kv_copy_range(), and client_fabric_receive() in src/aimu_endpoint_client.c clamped daemon-reported lengths only against the caller's own buffer capacity, not against the wire payload's actual capacity (ATT1_AIMU_ENDPOINT_MAX_PAYLOAD). A malformed response claiming an oversized length, combined with a sufficiently large caller buffer, could read past the local att1_aimu_endpoint_response struct. All three now clamp against both bounds before any memcpy, and client_fabric_receive()'s reported byte count now always matches what was actually copied.

Docs

• docs/PHASE2_PLAN.md, docs/OPERATION_LOG.md, and compiler/check_docs.py updated to reflect M168 complete / M169 active.