QVAC-13536 Fixing benchmark workflow errors (Emdeddings + LLM)

[maxim-smotrov]

Note

Embed datasets TRECCOVID and FiQA2018 are still failing with a generic message with no meaningful server logs. This suggests that a crash may be happening somewhere in the CPP layer (segfault?).

For now, we are disabling these datasets.

🎯 What problem does this PR solve?

• Benchmark pipeline runs were failing/noisy due to context overflow errors that were not returned in a structured, retryable way.
• Workflow defaults and results-summary collection made benchmark runs less stable and less informative in CI/manual dispatch.

📝 How does it solve it?

• Adds structured context-overflow handling in embed benchmark server responses (HTTP 422, code: CONTEXT_OVERFLOW, retryable details including sequence/context metadata).
• Adds adaptive retry logic in embed benchmark client/model wrapper:
  • catches overflow responses,
  • applies targeted per-sequence truncation when details are available,
  • falls back to global truncation tightening when details are unavailable,
  • retries up to a bounded number of attempts with clearer logging.
• Normalizes embed benchmark server model config/device handling to avoid inconsistent model-key/config behavior (cpu explicit, otherwise gpu).
• Updates workflow defaults and summary publishing:
  • embed benchmark default datasets narrowed to ArguAna,NFCorpus,SciFact,SCIDOCS,
  • llm benchmark default model pair switched to Qwen3 values,
  • both workflows now guard summary extraction and append only when fresh markdown results exist.

🧪 How was it tested?

• Ran the workflows in CI from a temp branch and verified they passed.

[gianni-cor]

/review

[github-actions]

Tier-based Approval Status

**PR Tier:** TIER1

**Current Status:** ✅ APPROVED

**Requirements:**
- 1 Team Member approval ✅ (1/1)
- 1 Team Lead OR Management approval ✅ (1/1)



---
*This comment is automatically updated when reviews change.*

[gianni-cor]

/review