fix(frontmatter): knowledgePageFromMarkdown strips an unbalanced trailing quote and breaks 39.7% of recorded checks

[drewstone]

What happens

knowledgePageFromMarkdown strips a trailing quote from a front-matter value it reads as a wrapper, even when the value's leading character is not that quote. A shell command that ends in a quoted path therefore comes back unbalanced.

import { knowledgePageFromMarkdown } from '@tangle-network/agent-knowledge'

const md = [
  '---',
  'id: x',
  'rung: 4',
  'check: bash "$DISCOVERY_RUN_WORKSPACE/checks/clean.sh"',
  'expect: OK',
  '---',
  '',
  'body',
  '',
].join('\n')

knowledgePageFromMarkdown('/tmp/x.md', md).frontmatter.check
// 'bash "$DISCOVERY_RUN_WORKSPACE/checks/clean.sh'      <- trailing quote gone

Why it matters

verifyGradeableEvidence runs bash -n -c <check> before executing. The mangled command does not parse, so it raises UncheckableClaimError with INVALID_SHELL_SYNTAX_NOTE, and the claim is recorded unrunnable for a reason that has nothing to do with the claim.

Measured 2026-09-01 over every rung >= 4 claim page in one lab (discovery-lab/pursuits/**/kb/pages/*.md): the parser disagrees with the raw ^check: (.*)$ line on 1,492 of 3,762 checks (39.7%). Every one of those grades unrunnable when it is executed through verifyGradeableEvidence.

The dominant shape is a quoted workspace path:

check: python3 "$DISCOVERY_RUN_WORKSPACE/checks/verify_q36_binary_N27.py"
check: python3 -c "import json;r=json.load(open('$DISCOVERY_RUN_WORKSPACE/work/census/results.json'))..."

Expected

A value is unquoted only when the quote is balanced — the first non-space character and the last character are the same quote. bash "..." starts with b, so nothing is stripped.

Versions

@tangle-network/agent-knowledge@11.0.0, node 24.18.0.

Workaround in place

The consumer reads check, expect and evidencePath from the raw ^key: (.*)$ line and keeps the parser for everything else. It carries this issue number in a comment and is removed when this lands.