fix(acp): deliver plain-text replies when the model skips the send tool

[micspiral]

What

Adds a content-delivery fallback to buzz-acp: when a managed agent answers a
conversational prompt in plain assistant text and never calls the publish tool
(buzz messages send), the harness posts that streamed content as a threaded
kind-9 reply — instead of silently dropping it.

Capable hosted models reliably publish via the send tool, but weaker local /
shared-compute models (e.g. a small Gemma/Qwen) often just answer in prose and
never call the tool, so the reply is lost. This closes that gap.

How

Delivery is confirmed from the send tool call's terminal outcome, not its
input text — intent is not delivery:

• A tool_call whose input matches the publish signature (normalized so both
  shell strings and argv forms like subprocess.run(['buzz','messages','send',…])
  match) only registers a candidate toolCallId.
• The candidate is resolved on its terminal tool_call_update:
  • failed / cancelled → not delivered (fallback stays armed).
  • completed → inspect rawOutput.isError, the CLI's {"accepted":…}
    response envelope, and any reported exit_code. A bare completed with no
    inspectable output keeps the status-quo direction (suppress).
• On a normal EndTurn with buffered content but no confirmed publish, post the
  content as a threaded reply (buzz_sdk::build_message, signed with the agent
  keys, best-effort — mirrors post_failure_notice). Bare acknowledgements the
  base prompt forbids publishing are skipped.

Fires only on EndTurn (not MaxTokens/MaxTurnRequests, which are
truncated/aborted turns) and only when a triggering channel event was present
(not heartbeats).

Testing

• Unit: send-tool detection incl. read-vs-send discrimination and argv-style
  publishes; the outcome classifier matrix
  (pending/failed/cancelled/completed × envelope/isError/exit_code); bare-ack
  filtering.
• Handler-level lifecycle tests through handle_session_update +
  take_undelivered_turn_message: success, failure, cancel, failed-then-retry,
  and argv-style publish.
• Original fallback was live-validated against a small Gemma on a real
  community: fires on prose answers, stays dormant when the model calls the
  send tool.

Notes

Split out of #2448 so this reply-delivery fix stands on its own, independent of
that PR's shared-compute usability changes. +710 / crates/buzz-acp only — no
desktop / mesh / toggle changes.

[erauner12]

Additional live validation: I tested this PR unchanged with OpenClaw 2026.6.1 acting as a plain-text ACP server behind buzz-acp.

Path: Buzz mention → buzz-acp → kubectl exec -i → openclaw acp → existing OpenClaw Gateway session.

OpenClaw emitted standard agent_message_chunk text and had no Buzz CLI/MCP tools. The harness logged content-delivery fallback: posted undelivered agent content as channel reply, and the relay contained exactly one correctly threaded kind:9 reply with the requested text. The full buzz-acp test suite also passed locally (600 unit tests + 9 lifecycle tests).

This confirms the fallback covers an ACP bridge where giving the downstream agent the Buzz signing key or a general-purpose shell/MCP solely for reply delivery would be undesirable.

[micspiral]

I am going to close this as we have had a few runs at other ways of this - but if this comes back, we will bring it back