What this PR does

Msg.append_usage() now accumulates a copy of the Usage it is handed, so a
message never adopts the caller's object.

append_usage is the accumulate-over-many-model-calls helper, and it is the first
call that decides who owns self.usage. Adopting the argument made that object
shared: every later call — on this message or on any other message that had been
given the same Usage — wrote through to it. Reported per-message token costs
then included other messages' calls, and the caller's own bookkeeping object
changed underneath it.

Description

Fixes #2801 (reported, triage/confirmed, claimed via /assign).

shared = Usage(input_tokens=4, output_tokens=2)
first, second = AssistantMsg(name="a", content=[]), AssistantMsg(name="b", content=[])

first.append_usage(shared)
second.append_usage(shared)
first.append_usage(Usage(input_tokens=1, output_tokens=1))

second only ever accumulated a 4/2 call, so 4/2 is what it should report, and
the caller's shared still holds the value the caller wrote. The same
write-through makes repeated appends of one object compound (a 3/1 usage totals
12/4 over three appends instead of 9/3), because accumulator and source are the
same object.

Copying matches the convention the repo already settled on for Usage ownership:
Agent._get_reply_usage() (src/agentscope/agent/_agent.py, added in #2562) is
documented as "Get a copy of the accumulated usage for the current reply" and
returns last_msg.usage.model_copy(). The three in-repo append_usage call
sites (agent/_agent.py, agent/_realtime/_agent.py, message/_base.py) each
build a fresh Usage, so nothing in the library relied on the aliasing — the
change only removes the hazard for callers of the public method.

How is it tested

Two cases appended to the existing tests/message_test.py, each asserting whole
model_dump() structures per .github/copilot-instructions.md §3:

• two calls accumulate field by field (17/8/6/1) while both argument objects keep
  the values their owner wrote;
• two messages fed the same Usage keep independent totals (5/3 and 4/2) and do
  not share one object.

Measured locally with the repo's own environment (Python 3.12.11, Windows 11
26200), pytest tests/message_test.py -q:

• RED — these two tests against unpatched 6d32be1 src: 2 failed, 2 passed.

• GREEN — against the patched src on this branch: 4 passed.

• Adjacent suites (event_to_message_test.py, event_test.py, agent_basic_test.py,
  compress_context_test.py, agent_structured_output_test.py,
  pipeline_goal_test.py): 67 passed, 4 subtests passed on this branch and the
  same 67 passed, 4 subtests passed on the unpatched baseline — no behaviour
  moved elsewhere.

• Full suite pytest tests, same environment, run twice:

  • this branch: 3 failed, 2325 passed, 305 skipped, 105 subtests passed in 695.75s
  • unpatched 6d32be1: 3 failed, 2323 passed, 305 skipped, 105 subtests passed in 786.96s

  So the same three failures on both sides, the only extra passes being the two
  tests added here. All three are in tests/realtime_agent_test.py
  (test_barge_in_truncates_to_what_was_heard, plus test_confirmed_tool_runs
  on both runs and one of test_provider_timeout_reconnects_on_next_audio /
  test_denied_tool_reports_denial depending on the run) — wall-clock-sensitive
  realtime audio cases in files this branch does not touch.

• GitHub Actions on this PR, commit 81be210: Pre-commit and
  Python Unittest Coverage both completed, and all three matrix jobs
  (ubuntu-latest/macos-26/windows-latest, Python 3.11) report success.

Lint, with the versions and arguments pinned in .pre-commit-config.yaml
(black --line-length=79 23.3.0, flake8 --extend-ignore=E203 6.1.0, mypy
1.7.0 and pylint 3.0.2 with the CI --disable list) — clean on both changed
files, pylint 10.00/10. pre-commit run --all-files itself could not be run
here (the hook environments need to be downloaded), so it is left to the CI run
on this PR.

AgentScope Version

2.0.8 (main @ 6d32be1dfcf18ed860df15a9079999fde8a3a9b7, the branch this PR is
cut from; append_usage is unchanged on the current main tip 083cbd1).

Checklist

• An issue has been created for this PR — [Bug]: Msg.append_usage() adopts the caller's Usage object, so messages given one Usage report each other's totals #2801, triage/confirmed, claimed via /assign
• I have read the CONTRIBUTING.md
• Docstrings are in Google style
• Related documentation has been updated (e.g. links, examples, etc.) in documentation repository — left unchecked because no documentation change is needed: the message's usage field is documented as the accumulated token usage of that message, which is the behaviour after this fix, and nothing documents the caller giving up ownership of the object it passes in.
• Code is ready for review

Authorship disclosure: this PR was written and measured autonomously by an AI
agent operating the Lesereingrape account; no human read the diff before
submission. The numbers above were produced by that agent with the commands
quoted and are reproducible.