TopicForge 0.5.5

Tool outputs were reworked after testing them with LLM agents on the author's
16 test scenarios (live DDS buses with planted faults).

Breaking

• detect_qos_mismatches returns a MismatchScan envelope instead of a bare
  list. Migrate by reading ["reports"] where you used the list.

Added

• list_endpoints, the twelfth tool: every announced DDS writer and reader with
  owning participant, structured QoS and a per-topic roll-up that flags orphans
  (no_reader, no_writer). Cyclone and mock only.
• Continuous discovery tracking on Cyclone, so a node restarted three times
  shows as 3 lost and 4 discovered events dated by DDS, not by poll time.
• announced_ns, lost_ns, time_source and related timestamp fields on
  participant_events and list_participants.
• health_check reports now_ns, tracker status and an observed-domain note.
• QosProfile gains Liveliness, Ownership, Partition, LatencyBudget,
  DestinationOrder and DataRepresentation.
• peek_dds_samples on builtin discovery topics returns structured endpoint
  and participant fields instead of a raw repr string.
• Topic filters accept rt/x and x interchangeably; a filter that matches
  nothing returns the known topics.
• detect_qos_mismatches also reports matched and not_matched pairs, hints
  (near-miss topic names, path suffixes, type id differences), and the policies
  it did and did not check.

Changed

• detect_qos_mismatches checks Partition first (with * and ? wildcards),
  then type names, then the RxO rules; Liveliness, LatencyBudget, Ownership,
  DestinationOrder and DataRepresentation join Reliability, Durability and
  Deadline.
• topic_metrics and peek_dds_samples on a user topic say they have no data
  instead of returning zeros or a placeholder.
• Tool descriptions no longer carry internal history.

Fixed

• Infinite durations are None instead of 9223372036854775807.
• Races between the tracker and tool calls could mark a live participant lost
  or keep endpoints of a departed one.
• Concurrent cyclonedds calls could corrupt the heap on Windows; every binding
  call now takes one lock.
• A failing tracker no longer adds 3 s to every tool call.
• An unreadable QoS duration is treated as unknown, not infinite.
• Typo hints stay bounded on a bus with a thousand topics.

Known limits

• A hung writer (alive, no data) cannot be observed without a data probe.
• A crash and a clean leave look the same; lost_ns is an upper bound.
• A participant that cycles faster than the history depth between two tracker
  passes can be missed.
• The Fast DDS backend has never run on a bus; list_endpoints is not
  supported there.
• DDS Security is not supported, and user-topic payload decoding is disabled.