fix(reflect): preserve bank attribution in provider calls

[Jordan-Jarvis]

Summary

• bind MemoryEngine.reflect_async to the existing per-bank provider-attribution context
• preserve the bank for Reflect tool calls and final synthesis
• verify the bank is bound during Reflect and reset after failure
• replace two lower-value direct ContextVar implementation tests with one integrated Reflect regression

Without this binding, retain and recall attribute provider usage correctly but Reflect is the only bank-scoped engine operation that loses the bank before its LLM/tool loop.

Validation

• uv run pytest tests/test_bank_attribution.py -q — 15 passed
• uv run ruff check hindsight_api/engine/memory_engine.py tests/test_bank_attribution.py
• uv run ruff format --check hindsight_api/engine/memory_engine.py tests/test_bank_attribution.py
• patch applies cleanly to current upstream main (9676fc1699c199175809285ea0e1095ab71115d0)

[Jordan-Jarvis]

Additional integration evidence from the corresponding Will qualification branch:

• the same bank-binding correction merged in Will-Augmented-Thought/hindsight as d693efd0f644ce75c5be93f26e73e19d318d9cc7;
• exact-main image run 29551436887 published, scanned, signed, and verified both API and control-plane artifacts;
• GitOps promotion 29554113973 passed dev retain, recall, and canonical per-bank provider attribution through native Reflect. The later failure was the independent noncanonical OpenAI tool-result message shape addressed by fix(reflect): emit canonical OpenAI tool result messages #2766, not bank attribution.

This validates that binding reflect_async through the existing bank ContextVar boundary preserves one bank-owned routing contract without adding provider-specific inference downstream.