Git it done v43

This project is a general-purpose AI agent built for Track 1, designed around the actual scoring mechanic — accuracy gate first, then token-efficiency ranking — rather than a generic "route to an LLM" approach. The core idea is a tiered escalation pipeline that tries the cheapest possible path first and only spends tokens when genuinely necessary: Deterministic solvers handle math reasoning and logic puzzles with zero token cost and zero risk of accuracy loss, using purpose-built parsers and solvers rather than model calls. Local model inference (via an Ollama sidecar) handles simpler categories — sentiment, NER, summarization, straightforward factual Q&A — using a self-consistency confidence check to decide whether the answer is trustworthy enough to return without escalating. Tiered Fireworks escalation kicks in only when local confidence is low or the task requires stronger reasoning or code capability, routing across the allowed model set (MiniMax-M3, Kimi-K2p7-Code, and three Gemma variants) based on measured cost/accuracy tradeoffs rather than name-pattern guesses. A calibration harness sweeps confidence thresholds against a held-out labeled dataset to set conservative, empirically-grounded cutoffs — since the platform's accuracy threshold is unpublished, this avoids the failure mode of guessing wrong and zeroing out the submission entirely. The agent also includes full compliance hardening (no hardcoded model IDs, env-based configuration, JSON schema validation, correct platform manifest, image size limits) and deliberately routes eligible tasks through Gemma models to qualify for the bonus prize pool. The result is an agent that competes on execution quality — how well the local classifier and confidence gate generalize to unseen prompt variants, and how aggressively it can trim token spend without failing the accuracy gate — rather than on architectural novelty alone.