feat: add pre-task-calibration — mechanical pre-action verification

[YuhaoLin2005]

What

A PreToolUse hook that enforces three mandatory checks before any Edit/Write operation: concept review, input consistency, and verification planning. Any check fails → exit 2, tool call blocked.

Why

The most expensive mistakes happen before the first keystroke — editing in the wrong direction. Pre-task calibration catches these at the only moment prevention is cheaper than correction.

How It Works

PreToolUse (Edit/Write/Bash) → Q1: Approach validated? → Q2: Inputs consistent? → Q3: Verification planned? → All three pass → allow.

Integration

Module A of a composable quality pipeline. Works standalone. Complements:

• persona-review (feat: add persona-review — multi-persona adversarial review skill #1420): adversarial depth for complex decisions
• delivery-verification: session-end completeness (separate PR)
• neural-gate: post-session behavioral verification (separate PR)

Each module is independent — adopt any, all, or none. Together they form defense in depth across the full session lifecycle.

Evidence

50+ session production use. Causal experiment (n=30, p=0.0092): config rules measurably change agent behavior. Mechanical enforcement via exit codes is more reliable than aspirational prompts.