fix: tool reliability improvements for sql-classify, edit, and webfetch

[anandgupta42]

Problem

Three tool reliability issues affecting agent effectiveness:

1. sql-classify.ts: computeSqlFingerprint references undefined core variable after safe-import refactor, silently breaking fingerprinting. Fallback classifier defaults to "read" for unknown statements (should be "write" for safety). Fallback doesn't handle multi-statement SQL.

2. edit.ts: When oldString not found, error message provides no context for self-correction. Substring matching in buildNotFoundMessage can latch onto short irrelevant lines.

3. webfetch.ts: No caching of failed URLs — agent retries 404s repeatedly. Error messages don't tell the model whether to retry. URLs with auth tokens in query strings leak into error messages.

Expected Behavior

• sql-classify should gracefully degrade with correct security guarantees in fallback mode
• Edit tool should provide nearest-match snippets to help the model self-correct
• Webfetch should cache permanent failures and provide actionable error messages