feat: implement compute workload budget policy, adversarial checks, and telemetry mapping

[maniradhakrishnan-dev]

Overview

This PR implements the budget and capability routing policies for Session 15 of EAGV3. We chose a custom domain workload for GPU Compute & AI Model Sizing consisting of 15 tasks of varying complexity, configured a custom 3-tier capability ladder, implemented hard budget controls, and verified them under adversarial runaway loop attacks.

---

What is implemented:

Part 1: Floor Reproduction

• Ran the agent event suites and captured event logs, SQLite database records, and OpenTelemetry trace IDs for 4 distinct runs.
• Documented key limitations regarding provider availability vulnerabilities and telemetry-ledger disconnects.

Part 2: Custom Policy & Workload Evaluation

• Configured a custom capability ladder in config/tiers.yaml (Economy: GPT-OSS-120B, Standard: Gemini-3.1-Flash-Lite, Frontier: Gemini-3.1-Pro).
• Built a budget controller that propagates tenant/project/user principal tags to the gateway and performs automated tier downgrades.
• Evaluated performance using a Gemini judge in proofs/run_compute_eval.py:
  • Cost Savings: Reduced cost per resolved task by 48% (from $0.000480 down to $0.000249).
  • Break-Even Rate: Evaluated the ladder's break-even resolution rate at 46.2%.
  • Mistake Analysis: Analyzed task t09_flops_estimation where downgrading to economy saved $0.000566 but lost correctness due to model math limitations.

Part 3: Adversarial Budget Attack

• Created proofs/attack_budget.py to drive an infinite runaway planning loop under budget limits.
• Proven that the controller hard-refuses provider calls once budget limits (e.g. $0.002 ceiling) are neared, capping spend at $0.001927 instead of $0.009241.
• Refusals are fully logged as task_failed inside the telemetry event trace.

---

File Changes

• config/tiers.yaml: Configured custom rungs and pricing.
• s15code/gateway.py: Added principal propagation and overrides passthrough filters.
• s15code/economics/controller.py: Enabled proper default fallback and principal mapping.
• proofs/tasks/compute_analysis.jsonl: The 15 custom model sizing tasks.
• proofs/run_compute_eval.py: Task runner and judge evaluation script.
• proofs/attack_budget.py: Runaway loop adversarial attack driver.
• proofs/capture_floor.py: Floor reproduction capture script.
• tests/test_cross_model_ladder.py: Adjusted model name assertion.
• README.md: Documented all Parts 1, 2, and 3 findings under the "Submission Evidence" section.