fix(eval): parse code fences like lib.sh in the outputs eval

[Tong-bit-art]

What does this PR do?

test-convert-outputs.sh's split-integrity model (fence_blocks) read code
fences only at column 0, while CommonMark — and lib.sh's
fence_open_p / fence_closes_p, aligned to it in #855 and #1028 — allows up
to three spaces of indentation. An indented opening fence was therefore
invisible to the eval, its column-0 closer became the model's opener, and
the following lines were reported as a torn fenced block across
SOUL.md/AGENTS.md:

  FAIL openclaw: fence-model-fixture source fenced block at lines 4-6 is torn
       across SOUL.md/AGENTS.md — a `## ` heading inside the fence was treated
       as a section boundary
  Results: 30 passed, 1 failed
  FAILED

The source is valid: GitHub renders the block, lint-agents.sh accepts it,
and convert_openclaw() splits it correctly (the model's "block at lines 4-6"
is its own mis-pairing, not the real three-line fence). Because the Check
Tools workflow runs this eval as a hard gate on every PR, a contributor
adding such an agent gets a failure that says their fence is torn when it is
not.

The model now mirrors lib.sh exactly: up to three spaces of indentation,
same character, a closer at least as long as the opener with nothing but
whitespace after the run, and an info-string run inside a block stays content.

Evidence

Reproduced in a throwaway repo with one fixture:

---
name: Fence Model Fixture
description: Fixture agent with an indented opening code fence
color: blue
---
## Identity
  ```text
  code sample
```
## Core Mission
mission text

On current main the eval fails with the report above; with this change all 15
tools validate. The existing roster is unchanged: the full eval reports
32 passed, 0 failed (282 agents × 15 tools) with zero manifest drift.

Tests

New scripts/test-convert-fence-model.sh builds the throwaway repo around the
real eval and that fixture, runs it with --update, and requires a clean
validation with no torn-block report. It fails on the previous model with the
false positive above and passes here.

Validation on the exact commit:

• full bash suite 19/19 and Hermes plugin checks 3/3;
• test-lint-fences.sh, strict test-convert-outputs.sh (zero drift),
  bash -n, git diff --check: clean.

Scope notes

• Two existing files plus one new test; no generated output, no manifest
  change, no converter behavior change.
• The eval's model is the only thing that changes: lib.sh and
  convert_openclaw() already handled indented fences.
• One step is added to the existing Check Tools workflow to run the new
  regression; no new workflow.

AI assistance was used in preparing this patch; the reproduction and test
runs above were executed locally.