fix: resilience improvements for retry logic, context trimming, blob listing, and logging

[Dongbumlee]

Purpose

• Add Qdrant-backed shared memory for cross-agent context sharing across workflow steps (analysis -> design -> convert -> documentation)
• Fix critical blob listing bug causing intermittent "files not found" failures across all steps
• Fix singleton cache bug causing memory store to be closed/uninitialized on subsequent runs
• Improve Azure OpenAI retry resilience for rate limits, transient errors, and context-length overflows
• Reduce LLM context window usage by keeping only 15 recent messages (down from 30/50)
• Convert all runtime print() to structured logging
• Externalize ResultGenerator prompts from inline code to text files

Does this introduce a breaking change?

• Yes
• No

Golden Path Validation

• I have tested the primary workflows (the "golden path") to ensure they function correctly without errors.

Deployment Validation

• I have validated the deployment process successfully and all services are running as expected with this change.

What to Check

Verify that the following are valid

• Blob listing fix: Agents can consistently list and find source/converted files (no more intermittent "folder is empty" errors)
• Memory store persistence: [MEMORY] Stored memory from logs appear across all steps, NOT _flush_memory skipped — memory store not initialized
• Cross-step memory injection: [MEMORY] Injecting N memories for {Agent} in design/convert/documentation steps
• Memory cleanup: [MEMORY] Workflow complete — closing memory store (N memories) with N > 0
• All 4 steps complete with PASS sign-offs on consecutive runs (not just first run)
• No [MEMORY] _embed skipped — client=None warnings
• Retry behavior: [AOAI_RETRY] logs show exponential backoff with 5s base delay
• Step timings show single yaml key (no duplicate yaml_conversion)
• Documentation reports use inline hyperlinks, not Markdown footnotes
• SHARED_MEMORY_ENABLED=false falls back to original v2 behavior without errors

Key Changes

Critical Bug Fixes

• Fix list_blobs_in_container trailing-slash bug — Non-recursive listing returned empty when folder_path lacked trailing /. Both name_starts_with prefix and relative_path computation now use normalized path with trailing /
• Fix singleton instance cache for QdrantMemoryStore — AppContext._instances cache held the old closed store from previous runs, causing _initialized=False on all subsequent runs. Now cleared before re-registering

Retry and Resilience

• Retry config: 8 retries, 5s base, 120s max (was 5/2s/30s)
• Cooldown delays on context-trim retries
• Retry transient errors: empty messages, 5xx server errors
• Embedding retry (3 retries with exponential backoff) in QdrantMemoryStore

Context Window

• Remove tool-result truncation; only summarize blob writes
• keep_last_messages 30 -> 15; disable per-message truncation
• Protect last message from truncation

Logging and Telemetry

• Fix duplicate yaml_conversion/yaml telemetry key
• Convert all print() to logger across 8+ files; remove text2art
• Add diagnostic logging for memory storage failures (_embed, _flush_memory, invoked, flush)
• Clear client cache between processes

Other

• Prohibit Markdown footnotes in documentation reports
• Workspace context injected into agent system instructions

Other Information

Configuration:

• Enabled by default (SHARED_MEMORY_ENABLED=true)
• Requires AZURE_OPENAI_EMBEDDING_DEPLOYMENT_NAME in App Config (added to Bicep)
• Gracefully degrades: if embedding deployment is missing, workflow runs without memory