fix(btw): compensate a work submission that is never handed over

[YUZHEthefool]

Summary

A work submission interrupted between its acknowledgement and the background
hand-off left a queued task no run owned. Expected the session to end
cancelled/failed and be cleaned up once; actual /work status kept reporting
a pending task, with zero background tasks registered and the event already
cleaned up. Pending sessions are exempt from retention, so the record survived
until a reload or restart.

Related issue

Fixes #155

Root cause

submit() creates the queued session, yields the acknowledgement, and registers
the background run only once the generator resumes. When a later stage stops the
event, the scheduler breaks out of the generator and never resumes it — an
async for that breaks does not close its generator — so no code path cancelled
the session. close() only walks registered _tasks, and the retention sweep
skips PENDING, so nothing reaped it either.

Reproduction

Build a real PipelineScheduler and WorkLoop, add a stage after the
acknowledgement that calls event.stop_event(), then run scheduler.execute and
read the WorkSession. The production equivalents are a plugin that stops the
event while decorating the acknowledgement, or a cancelled request.

Implementation notes

Reclaim the session from both ends, with the transition itself deciding who did
it:

• submit() cancels the session in a finally unless the background run
  accepted it. This covers a generator closed by aclose()/GeneratorExit.
• PipelineScheduler.execute's existing finally is the one place that runs even
  when the generator is never resumed, so it asks runtime_registry to cancel a
  work session whose background run never took over. This reuses the same hook
  point the scheduler already has for BTW (btw_detached_work,
  finalize_detached_event), so the coupling is not new.
• WorkSessionManager.cancel_if_pending only moves PENDING to CANCELLED, so
  the two paths cannot cancel twice or clean up twice. It cannot touch a
  legitimate background run that is merely queued behind the semaphore, because
  the hand-off marks the event before the task starts.
• config_id_of moves to runtime_registry (the module the pipeline already uses
  to reach work-loop state); submission imports it from there.

Validation

.venv/Scripts/python.exe -m pytest tests/unit -q
  5220 passed, 6 skipped

.venv/Scripts/python.exe -m ruff format --check <touched files>   # clean
.venv/Scripts/python.exe -m ruff check <touched files>            # clean
.venv/Scripts/python.exe -m pyright --project pyrightconfig.quality.json <touched files>
  0 errors

The new test was run against the unmodified tree first, where it fails with
WorkSessionStatus.PENDING and zero background tasks — the exact reported
symptom.

make check, make quality, and the coverage-gated --test-profile all run were
not run locally on this Windows checkout; this PR relies on CI for those, as #168
did.

After #168 was rebased onto the updated master head, this branch was rebased onto
the new head as well and the whole stack was re-validated at its tip:
pytest tests/unit -q 5233 passed / 6 skipped and
pytest --test-profile blocking -q 5354 passed / 6 skipped / 1 deselected.
The count above was taken on this branch before that rebase.

Compatibility and risk

• PipelineScheduler.execute's finally now makes one extra call per event
  (a registry lookup plus a mark check). It is a no-op for every event that does
  not carry a BTW work session.
• config_id_of is still importable from astrbot.core.agent.btw.submission.
• No route, schema, configuration, or dependency change. BTW stays off by default.

Checklist

• The change is focused and does not include unrelated refactoring.
• I added or updated a regression test, or explained why a test is not practical.
• I ran the relevant formatting, lint, build, and test commands.
• User-visible behavior updates both docs/zh/ and docs/en/ when needed. (No user-visible text changed.)
• OpenAPI, generated client, docs/public/openapi.json, and tests change together when routes or schemas change. (No route or schema change.)
• No secrets committed. Runtime Python deps update pyproject.toml, requirements.txt, and uv.lock together. (No dependency change.)
• I did not restore legacy shims, Python <3.14 fallbacks, or upstream publish/docs URLs as fork artifacts.
• Breaking API or behavior changes use ! and a BREAKING CHANGE: footer. (None.)
• I will not merge this PR myself. Merge needs a human maintainer review plus a separate AI-assisted review (AI_POLICY.md).
• AI use follows AI_POLICY.md. Keep exactly one author note below. Do not fabricate the other.

Agent note

Goal: resolve #155 by compensating between the acknowledgement
and the background hand-off, keeping BTW off by default and preserving
authorization and request identity.

Paths touched: astrbot/core/agent/btw/work_loop.py,
astrbot/core/agent/btw/work_sessions.py,
astrbot/core/agent/btw/runtime_registry.py,
astrbot/core/agent/btw/submission.py, astrbot/core/pipeline/scheduler.py, and
tests/unit/test_btw_delivery.py, tests/unit/test_work_tools.py.

Checks run: the commands under Validation, plus the new test against the
unmodified tree.

Residual risk: the compensation depends on the scheduler's finally running. An
isolated WorkLoop used with submit() and no scheduler has neither the
generator resumption nor that backstop, so a submission abandoned there could
still linger; the finally in submit() covers the close paths that do occur,
and no such caller exists in the tree.

Tools used: Claude Code (Opus 5) with the repository's AGENTS.md and AI_POLICY.md.

This PR is stacked on #170; merge that first.