JVNAUTOSCI-2514 bound recovery prompt context

[witbrock]

Summary

• Bound the recovery_decision LLM context fields with per-field and total prompt budgets.
• Drop duplicate/internal runtime telemetry from recovery context while preserving failure evidence and loop-attempt signals.
• Attach prompt-context compaction diagnostics to LLM call telemetry and step envelopes for follow-up UI surfacing in JVNAUTOSCI-2515.

Validation

• .venv/bin/python -m pytest tests/backend/test_llm_step_executor.py -q --tb=short
• .venv/bin/python -m pytest tests/backend/test_turn_execution_workflow_definitions.py -q --tb=short -k 'not test_conversation_turn_recovery_retry_progresses_across_multiple_prompt_targets'
• .venv/bin/python -m ruff check src/backend/workflows/llm_step_executor.py src/backend/workflows/recovery_prompt_compaction.py tests/backend/test_llm_step_executor.py
• .venv/bin/pyright src/backend/workflows/llm_step_executor.py src/backend/workflows/recovery_prompt_compaction.py tests/backend/test_llm_step_executor.py
• .venv/bin/python -m py_compile src/backend/workflows/llm_step_executor.py src/backend/workflows/recovery_prompt_compaction.py tests/backend/test_llm_step_executor.py
• git diff --check

Note: test_conversation_turn_recovery_retry_progresses_across_multiple_prompt_targets is excluded because it currently fails on origin/main with metadata_validation_failed:metadata_write_context_key_missing:completion_gate:completion_gate_preserved_response.