fix(loongsuite): isolate advice and stream failures

[sipercai]

Description

This PR adds a small loongsuite-instrumentation-common package that provides
failure-isolated advice helpers and protocol-preserving synchronous/asynchronous
stream proxies.

The application call remains outside the advice boundary. Instrumentation
callbacks are best effort, while application return values, stream chunks,
original exceptions, close results, and context-manager suppression semantics
remain unchanged. This follows the
OpenTelemetry error-handling requirement
that instrumentation errors must not escape into application code.

The PR also routes LiteLLM stream accumulation through call_advice, registers
the common package in the LoongSuite test/lint matrix, and marks it as a shared
runtime package that must not be treated as an auto-enabled instrumentation by
the bootstrap generator.

No framework-owned or Python-ABI-specific dependency is added to the shared
runtime package.

Fixes # (N/A)

Type of change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• This change requires a documentation update

How Has This Been Tested?

• tox -e precommit
• tox -e generate-workflows
• tox -c tox-loongsuite.ini -e py38-test-loongsuite-instrumentation-common,py39-test-loongsuite-instrumentation-common,py310-test-loongsuite-instrumentation-common,py311-test-loongsuite-instrumentation-common,py312-test-loongsuite-instrumentation-common,py313-test-loongsuite-instrumentation-common,lint-loongsuite-instrumentation-common
• tox -c tox-loongsuite.ini -e py312-test-loongsuite-instrumentation-litellm,lint-loongsuite-instrumentation-litellm
• python -m pytest -q scripts/loongsuite/tests/test_generate_loongsuite_bootstrap.py
• Live DashScope non-streaming, streaming, concurrent, and error-path smoke
• Weaver live-check against the LoongSuite GenAI profile

Does This PR Require a Core Repo Change?

• Yes. - Link to PR:
• No.

Checklist:

See contributing.md for styleguide, changelog guidelines, and more.

• Followed the style guidelines of this project
• Changelogs have been updated
• Unit tests have been added
• Documentation has been updated

Validation Evidence

Spec and Scope

• Linked issue/spec: no public issue; the scoped design was approved before implementation.
• Approved scope: common fail-open advice and stream lifecycle primitives, one LiteLLM integration, CI registration, and bootstrap opt-out.
• Changed surface: loongsuite-instrumentation-common, LiteLLM stream accumulation, LoongSuite tox/workflow generation, and bootstrap package discovery.

Local Checks

Check	Command	Result	Notes
Static readiness	LoongSuite pipeline readiness checker	pass	No errors or warnings.
Precommit	tox -e precommit with a fresh hook cache	pass	Working tree remained clean.
Common tests	tox -c tox-loongsuite.ini -e py38-test-loongsuite-instrumentation-common,py39-test-loongsuite-instrumentation-common,py310-test-loongsuite-instrumentation-common,py311-test-loongsuite-instrumentation-common,py312-test-loongsuite-instrumentation-common,py313-test-loongsuite-instrumentation-common,lint-loongsuite-instrumentation-common	pass	26 tests on every supported Python version; lint passed.
LiteLLM tests	tox -c tox-loongsuite.ini -e py312-test-loongsuite-instrumentation-litellm,lint-loongsuite-instrumentation-litellm	pass	48 tests; lint passed.
Workflow generation	tox -e generate-workflows	pass	Generated workflows remained current.
Bootstrap tests	python -m pytest -q scripts/loongsuite/tests/test_generate_loongsuite_bootstrap.py	pass	5 tests.
Package build	uv build --wheel instrumentation-loongsuite/loongsuite-instrumentation-common and LiteLLM equivalent	pass	Both wheels are py3-none-any; common has no runtime dependencies.
External review	Three-round structured external review	pass	Two lifecycle findings fixed and re-reviewed; remaining P3 suggestions are non-blocking.
Privacy scan	Branch diff and cassette privacy scans	pass	No credentials, private paths, tokens, or private infrastructure references found.

Fix Verification

The same LiteLLM stream reproduction was executed against the merge base,
current main, and this branch. An accumulator callback failure escaped and
prevented delivery of the application chunk on both the merge base and current
main; this branch returned the original chunk and completed normally.

Real E2E Matrix

Scenario	Status	Command or Demo	Evidence
non-streaming	pass	litellm_genai_smoke.py with DashScope	One bounded live request completed and emitted an LLM span.
streaming	pass	litellm_genai_smoke.py with DashScope	Live stream completed with multiple chunks and a finalized LLM span.
concurrency	pass	Bounded two-call async DashScope smoke	Both calls completed with distinct trace IDs and no content/parent crossover.
agent/tool/ReAct	N/A	LiteLLM client instrumentation	This integration emits LLM spans and does not own an agent or tool execution lifecycle.
tool-heavy	N/A	LiteLLM client instrumentation	The changed boundary does not execute tools; streamed tool-call delta accumulation remains covered by focused tests.
error path	pass	Live invalid-model request through the instrumented LiteLLM path	The provider error remained visible and an error-status LLM span was finalized.

Telemetry and Weaver

Check	Status	Command or Artifact	Notes
Span tree / span kinds	pass	In-memory exporter from the bounded live smoke	Five root LLM spans were emitted for five calls.
Content capture modes	pass	Focused LiteLLM tests plus live NO_CONTENT smoke	NO_CONTENT emitted no message attributes; SPAN_ONLY behavior is covered by focused telemetry tests. SPAN_AND_EVENT is N/A because this integration does not emit message events.
Concurrency isolation	pass	Two-call async live smoke	Every call used a distinct trace ID.
Weaver live-check	pass	weaver registry live-check -r <registry>/model --advice-profile loongsuite-genai --input-format json ...	Five live spans, zero violations; only registry stability/informational advice remained.

CI

• GitHub checks: pending PR publication.
• Known unrelated failures: none observed locally.
• Follow-up needed: inspect the full generated LoongSuite matrix after publication; the shared workflow-generator change intentionally selects full LoongSuite CI.

[ralf0131]

@/tmp/pr245_review_body.md