fix(tracing): preserve inherited baggage tags when adding local tags

[lumman0]

What does this PR do?

Fixes langfuse/langfuse#18379.

Draft / dependency: #1919 must land first (or its agreed tag decoder must be available). This branch cherry-picks Ambuj Upadhyay's #1919 commit 22292d5fe13e530a1deb89214e9b3b95ada18d0a with original authorship and a cherry picked from trailer. The second commit is this PR's tag-merge change. I will update the branch against main once that dependency lands; this is not intended to replace #1919.

After W3C baggage crosses a process boundary, adding local tags with propagate_attributes(tags=[...]) currently drops the inherited tags. The merge reads only the SDK's local context key, which is not restored by header extraction. In-process nesting already appends tags.

This change seeds the merge from the existing _get_propagated_attributes_from_context(context) reader. It reuses #1919's tag decoding and preserves the reader's precedence of local context over baggage. The existing append/dedup logic and wire format are unchanged.

The regression tests cover:

• Real W3C header serialization and extraction into a fresh context.
• Updating the active observation and propagating the merged tags to its child.
• Overlapping tags, order, and a tag containing a comma.
• Restoring the inherited tags for observations created after leaving the local scope.
• as_baggage=False: merge locally, keep outgoing upstream baggage unchanged.
• as_baggage=True: send the merged tags to the next service.
• Keeping local context authoritative when it differs from baggage.

The propagate_attributes docstring now states the additive behavior. The JS counterpart is langfuse/langfuse-js#959 / langfuse/langfuse-js#966.

Type of change

• Bug fix
• New feature
• Breaking change
• Refactor
• Documentation update
• Tooling, CI, or repo maintenance

Verification

Python 3.12.14, Windows, OTel API/SDK 1.45.0. Commands used the existing locked development environment's Python interpreter, with PYTHONPATH pointing at this isolated worktree:

# Baseline with #1919 only: 136 passed
python -m pytest tests/unit/test_propagate_attributes.py -q -o log_cli=false --tb=short --timeout=20

# Before the fix: 2 failed (False/True baggage modes), 1 passed (precedence control)
python -m pytest tests/unit/test_propagate_attributes.py -q -k 'tags_append_' -o log_cli=false --tb=short --timeout=20

# After the fix: 204 passed, 2 skipped
python -m pytest tests/unit/test_propagate_attributes.py tests/unit/test_otel.py -q -o log_cli=false --tb=short --timeout=20

ruff check .
ruff format --check langfuse/_client/propagation.py tests/unit/test_propagate_attributes.py
python -m mypy langfuse --no-error-summary
git diff --check

Lint, formatting of both changed files, and type checking passed. The test run reports one existing set_trace_io deprecation warning. An initial sandboxed baseline timed out while Windows asyncio created its socket pair; the baseline and final tests above passed outside that sandbox.

I also reran an external three-process A -> B -> C localhost HTTP probe against the final source, using the real SDK with in-memory exporters. All hops kept the same trace ID:

B operation	B observation tags	C observation tags
No enrichment	memory, retry	memory, retry
Add worker, as_baggage=False	memory, retry, worker	memory, retry
Add worker, as_baggage=True	memory, retry, worker	memory, retry, worker

No backend ingestion/UI e2e or live-provider tests were run; the defect and fix are observable at the SDK export boundary. The full unit suite was not run; verification focused on propagation and OTel tests.

Checklist

• I self-reviewed the diff using code_review.md.
• I added or updated tests for behavior changes.
• I updated docs, examples, or .env.template if needed. (Updated the public propagation docstring.)
• I did not hand-edit generated files; if generated files changed, I used the upstream regeneration path.
• I did not commit secrets or credentials.