fix(detector): detect first-call state replay for red 382

[prasannakotyal]

KernelGuard-Red-Submission: 382

Adds a narrow pass to detect_last_call_replay for first-call output replay through an input-independent state slot.

The pass requires:

• an entrypoint if guard that references no live input parameter
• the same guarded state root to be assigned from a live input-derived expression
• the same state root to be returned after initialization, or returned in the fast path before later initialization

This targets replay caches that initialize from the first tensor and then return that stored tensor for later calls without checking the new input.

Local validation:

• UV_CACHE_DIR=/tmp/uvcache uv run python -m py_compile competition_state/kernelguard_mirror/kernelguard.py
• UV_CACHE_DIR=/tmp/uvcache uv run python kernelguard_bypasses/eval_blue_patch.py competition_state/kernelguard_mirror/kernelguard.py
• external accepted reds filtered locally: 257, 260, 261, 264, 266, 268, 297, 323, 360, 367, 382, 390, 412, 422

Target red artifact: SarahLacard red 382, first-call state replay variant.

[github-actions]

KernelGuard Blue Evaluation

• Evaluation job: 162
• Blue submission: 134
• Result: passed
• Target red submission caught: True
• Validation suite: TP 20/20, FP 20/20
• Surgicalness score: 1.0
• Workflow run: https://github.com/SinatrasC/kernelguard/actions/runs/25257219193

[SinatrasC]

Thanks for the KernelGuard Flywheel Campaign contribution. This PR is now superseded by the consolidated rule-family implementation in #273, which folds this detector coverage together with the related passing-eval variants.