feat: add verified-code-answers plugin to fix unverified assertions (fixes #29753)

[HarshalJain-cs]

Summary

Fixes #29753 — Claude Code presents unverified assertions with the same confidence as verified ones.

Problem

When Claude Code answers a question about code behavior, there is no discernible difference in tone, confidence, or presentation between:

1. An answer that was verified by reading the actual source code with tools
2. An answer that was pattern-matched from training data without checking

The existing system prompt instruction "Do not say things you haven't verified" is insufficient on its own, as Claude still generates confident unverified answers.

Solution

This PR adds a new plugin verified-code-answers that injects a SessionStart hook with three explicit behavioral rules:

Rule 1: Read Before You Assert

Claude must use file reading tools (Read, Glob, Grep) to inspect actual source code before answering any factual question about code behavior.

Rule 2: Caveat Unverified Answers

If Claude cannot read the relevant source file, it must explicitly caveat its answer (e.g., "I haven't read the source file, but based on common patterns...").

Rule 3: Never Present Unverified Information as Verified

Code-behavior questions must never be answered with the same confidence as verified answers unless the code was actually read.

Files Added

plugins/verified-code-answers/
  .claude-plugin/plugin.json       # Plugin metadata
  hooks/hooks.json                  # SessionStart hook configuration
  hooks-handlers/session-start.sh  # Core instruction injection script
  README.md                         # Documentation

Testing

This plugin works identically to the existing explanatory-output-style plugin in structure and hook mechanism. Install by copying to your Claude Code plugins directory.