fix(stream): emit a tool_call for zero-argument tools so their card is never lost

[Luhaozhu]

Problem

A tool declared with no parameters (its arguments are always {}) never showed up in the chat's tool list, even though it ran and its output reached the answer. With tools running in parallel, a sibling tool's card also briefly rendered the missing tool's output.

Root cause

_tool_args_ready (orchestration/tool_payloads.py) exists to hold a tool card back while the model is still streaming the call's arguments — the first chunks arrive with incomplete args. It cannot distinguish "not written yet" from "there is nothing to write":

if not tool_args:
    if tool_name in _FAST_EMIT_TOOLS:
        return True
    return False

A zero-argument tool keeps empty args forever, so it never clears the gate and astream_chat_workflow emits no tool_call event for it at all.

Its tool_result then arrives orphaned, and two things go wrong downstream:

1. Frontend (hooks/chatStream.ts) — findToolCallIndex fell back to "bind this result to the last still-running card". Under parallel tool calls that is an unrelated sibling: the sibling's card gets overwritten with the wrong output (until its own result lands and overwrites it back), and the real call never appears.
2. Persistence — attach_tool_result appends a bare log entry with no display name, no args and no content_offset. buildHistorySegments requires every entry to carry an offset, so one such entry drops the entire message off the offset-ordered history replay.

Fix

• Backend — at the tool_result stage, synthesize the tool_call that was never emitted when the tool_id never produced a card, in both streaming paths (astream_chat_workflow and _astream_subagent_direct). The card keeps the tool's normal display name; update_plan, which deliberately renders no card, is still skipped beforehand. Because the synthesized call goes through the regular log path, the persisted entry regains its display name, args and content_offset.
• Frontend — drop the blind "last running card" fallback for results that carry a tool_id which matched nothing. That combination means the result belongs to a card that was never created, and claiming an unrelated one only hides the real call. Results with no tool_id at all keep the old behaviour.

Tests

New tests/orchestration/test_missing_tool_call_synthesis.py covers the gate precondition, synthesis for an unseen tool_id, idempotency once the id is claimed, the skill-load display name, the raw-name fallback, and that the persisted log entry is complete while a sibling card does not absorb the orphan result.