v4.0.0 — Agent-to-Agent: the listening half of On Board

v4.0.0 — Agent-to-Agent: the listening half of On Board

The board is no longer pull-only. memory_wait_for_event lets an agent park
inside one tool call and wake when a peer creates a ticket, changes a status,
or assigns work — verified live across Claude Desktop x Claude Desktop and
Claude x Codex (GPT), including a full reject -> fix -> resubmit cycle closed
with zero human relay.

Highlights:

• memory_wait_for_event + listen prompt: check-before-block, one wake
  drains the queue, agents never wake on their own actions.
• Role gate (completed != success): the executor's terminal move is
  submitted; only the owner or a main/lead/reviewer closes. Solo use stays
  possible via explicit allow_self_review=True, permanently stamped
  SELF-REVIEWED in the audit.
• Review verdicts travel: review_notes and fix_instructions ride the wake
  payload, so a rejected worker can retry without a human relaying anything.
• stay_active=true on submit keeps a listener on board to catch the verdict.
• Concurrency hardening for simultaneous writers: advisory board lock on
  ticket mutations and per-process tmp files in the JSON store.
• Board style guidance (token-thrift) on machine-read fields; measured
  -17.8% tokens on a real 17-unit corpus, fidelity-gated.

Breaking changes and the migration guide (existing .agent-mem/ boards load
unchanged; legacy wait cursors carry over) are in CHANGELOG.md.