Reduce AIC burn and stall risk in PR Code Quality Reviewer workflow

[Copilot]

This updates the PR Code Quality Reviewer workflow to address high AIC variance and a costly driver_exit failure mode observed in recent runs. The changes focus on fail-cheap behavior for stalled executions and reducing redundant context/tool usage on longer runs.

• Guardrails for stalled runs (driver_exit)

  • Added explicit prompt-level budget constraints so the agent exits early with a partial COMMENT review when progress stalls instead of consuming full run budget.
  • Clarified stop conditions tied to turn/tool-call progress to prevent zero-output high-cost failures.

• Context reuse and cache-read reduction

  • Tightened handoff guidance so downstream steps use compact prior outputs (sub-agent JSONL + diff summary) instead of reloading broad context each turn.
  • Reinforced reuse of already-loaded data between adjudication steps.

• Tool-call budget discipline

  • Added explicit instruction not to re-read files already loaded in Step 1.
  • Reduced duplication between main-agent analysis and sub-agent output consumption.

Example of the new guardrail pattern in prompt instructions:

If tool-call count exceeds budget before reaching final adjudication,
stop and submit a partial COMMENT review with highest-confidence findings.
Do not re-read Step 1 files; use already-loaded context and sub-agent summary.

• Fixes [agentic-token-optimizer] Optimize AIC usage in PR Code Quality Reviewer (79.4 avg AIC/run, 1 driver_exit failure) #50028