fix(instrument): Brand leak in agentforce trust layer YAML + missing STREAMING/REPLAY capability declarations

[mmercuri]

Two related fixes from the depth audit (A:/tmp/adapter-depth-audit.md).

1. Brand leak in agentforce trust layer YAML

Audit finding: agentforce/trust_layer.py:144 was emitting

# Stratix Policy - Imported from Einstein Trust Layer
# Generated by: stratix.sdk.python.adapters.agentforce.trust_layer

into customer-visible YAML files. These files end up in customers' source trees, version control, and audit packages, so the legacy STRATIX brand and the obsolete stratix.sdk.python.* import path were leaking into customer-owned artifacts.

Fix: header now emits the LayerLens brand and the actual current import path:

# LayerLens Policy - Imported from Einstein Trust Layer
# Generated by: layerlens.instrument.adapters.frameworks.agentforce.trust_layer

A regression test test_trust_layer_yaml_has_no_stratix_brand_leak covers both the canonical method and the deprecated to_stratix_policy alias, asserting:

• Positive: # LayerLens Policy and the new generator path are present.
• Negative: STRATIX, Stratix, stratix.sdk, stratix.sdk.python, ateam are all absent.

Sweep (audit task #4)

grep -rn "Stratix\|stratix" src/layerlens/instrument/adapters/frameworks/ | grep -v "deprecated\|backward\|alias" surfaced additional leaks. All non-deprecation contexts fixed:

• embedding/embedding_adapter.py — module docstring, author="STRATIX Team" (surfaced via AdapterInfo)
• embedding/vector_store_adapter.py — module docstring + author
• embedding/__init__.py, benchmark_import/__init__.py, benchmark_import/adapter.py, semantic_kernel/__init__.py, semantic_kernel/lifecycle.py, semantic_kernel/filters.py — module docstrings
• agentforce/auth.py — error messages now reference layerlens agentforce connect (current CLI binary; both layerlens and stratix are registered scripts)
• agentforce/events.py — Platform Events thread name (stratix-sf-events → layerlens-sf-events)
• agentforce/mapper.py — internal docstrings
• google_adk, llama_index, openai_agents, semantic_kernel — method docstrings
• llama_index/lifecycle.py — closure-only StratixEventHandler renamed to LayerLensEventHandler (the name surfaces in LlamaIndex dispatcher logs / UI)

Intentionally left alone (would require deprecation aliases — out of scope for a fix PR):

• class StratixMemoryStore (semantic_kernel) — public class with backward-compat note added to the docstring
• class Stratix re-exported from layerlens — actual API client class name
• to_stratix_policy method — already exists as a deprecated alias

2. Missing STREAMING / REPLAY capability declarations

Audit finding: 6 adapters wrap streaming methods but do not declare AdapterCapability.STREAMING. Only langfuse declares REPLAY despite every adapter implementing serialize_for_replay().

STREAMING (added to all 6)

Adapter	Streaming entry-point wrapped
agno	Agent.arun (async stream)
ms_agent_framework	ChatCompletionAgent.invoke_stream
openai_agents	TraceProcessor receives GenerationSpanData per chunk
google_adk	BeforeModelCallback / AfterModelCallback fire per chunk
llama_index	Instrumentation Module emits per-chunk events
bedrock_agents	invoke_agent returns an EventStream completion

REPLAY (added to every adapter that implements serialize_for_replay())

agno, bedrock_agents, google_adk, llama_index, pydantic_ai, strands, openai_agents, ms_agent_framework, embedding (both EmbeddingAdapter and VectorStoreAdapter), semantic_kernel, smolagents, agentforce.

Without these declarations, the atlas-app catalog UI surfaces incorrect feature support — telling customers they cannot replay traces from an adapter that supports it, or cannot stream from one that wraps every streaming entry-point.

3. Tests

• Per-adapter test extension: every adapter test file gained a declares_streaming_and_replay_capabilities (6) or declares_replay_capability (7) test that asserts the capability appears in get_adapter_info().capabilities.
• New lint guard tests/instrument/adapters/frameworks/test_capability_consistency.py enforces both rules consistently across every adapter discovered in the branch — the in-tree counterpart of the upstream manifest_consistency lint guard from the manifest emitter PR:
  • test_replay_capability_matches_serialize_for_replay — REPLAY declared iff serialize_for_replay implemented (with non-stub body).
  • test_streaming_capability_declared_for_streaming_adapters — every adapter on the canonical streaming list declares STREAMING.

4. Drive-by fixes

• tests/instrument/adapters/frameworks/test_bulk_ported_smoke.py was hard-coded to import every adapter package, failing collection on any branch missing one of them. Replaced the per-adapter import block with a try/except importlib loop so the smoke suite tests whatever adapters are present.
• test_agentforce.test_package_does_not_eagerly_import_requests was deleting agentforce.* from sys.modules without restoring them, breaking class identity (is) checks in subsequent tests within the same session. Saved/restored the original module objects so the cleanup is hermetic.

5. Sample regenerated trust-layer YAML

# LayerLens Policy - Imported from Einstein Trust Layer
# Generated by: layerlens.instrument.adapters.frameworks.agentforce.trust_layer
# Source: Salesforce Einstein Trust Layer

policy:
  name: customer_policy
  version: "1.0.0"
  description: "Policy imported from Salesforce Einstein Trust Layer"
  source: salesforce_agentforce

settings:
  data_masking: true
  zero_data_retention: true
  audit_trail: true

rules:
  - name: toxicity_detection
    description: "Imported from Einstein Trust Layer: toxicity"
    type: toxicity
    enabled: true
    threshold: 0.7
    action: block
    source: einstein_trust_layer
  - name: pii_detection
    description: "Imported from Einstein Trust Layer: pii"
    type: pii
    enabled: true
    threshold: 0.8
    action: block
    source: einstein_trust_layer
  ...

Verification

$ uv run python -m pytest tests/instrument/adapters/frameworks/
======================= 231 passed, 1 skipped in 2.14s ========================

$ uv run python -m pytest tests/instrument/adapters/frameworks/test_agentforce.py -k brand_leak
====================== 1 passed, 37 deselected in 0.66s =======================

$ uv run python -m mypy --strict src/layerlens/instrument/adapters/frameworks/agentforce
Success: no issues found in 11 source files

$ uv run python -m ruff check src/layerlens/instrument/adapters/frameworks/ tests/instrument/adapters/frameworks/
All checks passed!

Branch base note

This branch merges feat/instrument-frameworks-agentforce (brand leak source) with feat/instrument-frameworks-agents (the 13 agent-tier framework adapters that need capability declarations) so all enumerated audit items can land in a single PR. Conflicts resolved in pyproject.toml (combined the agentforce + agent-tier extras blocks) and frameworks/__init__.py (combined the docstring adapter list).

Test plan

• uv run python -m pytest tests/instrument/adapters/frameworks/ (231 passed, 1 skipped)
• uv run python -m pytest tests/instrument/adapters/frameworks/test_agentforce.py -k brand_leak (1 passed)
• uv run python -m mypy --strict src/layerlens/instrument/adapters/frameworks/agentforce (clean)
• uv run python -m ruff check src/layerlens/instrument/adapters/frameworks/ tests/instrument/adapters/frameworks/ (clean)
• Manual sample of regenerated trust-layer YAML — LayerLens branding, no STRATIX leaks.