docs(mcp-builder): make evaluation command examples runnable - #1921

Reviewed at head 271b027.

Scope: skills/mcp-builder/reference/evaluation.md (+15/−15), docs only. Single commit, nothing else vs merge-base 3337550.

The failure is real at the parent — every Bash block containing evaluation.py was extracted from the merge-base doc and its tokens fed to the script's real argparse (-a/--args, -e/--env, -H/--header are nargs='+'; ArgumentParser.parse_args is spied so main() can never reach an MCP server):

• 4 of 6 examples exit 2 with evaluation.py: error: the following arguments are required: eval_file: the stdio quickstart (-t stdio -c python -a my_mcp_server.py evaluation.xml), the env variant (-e DEBUG=true evaluation.xml), and both SSE/HTTP examples (-H "Authorization: Bearer ***" evaluation.xml). The trailing path is consumed by the last multi-value flag.
• The two examples that end on a single-value flag (-o … evaluation.xml) parse fine at the parent — the "four of six" claim is exact.
• The script itself is byte-identical at both states (only the .md changed), so both sweeps were parsed by the same argparse declarations — the delta is purely the command text. No evaluation is ever executed: the parse spy exits right after parse_args.

Fix verified (head)

• ✅ All 6 examples now parse: eval_file resolves correctly in each (evaluation.xml, my_evaluation.xml), transports/args/output come through as written.
• ✅ Multi-value examples keep every value — asserted from the parsed namespace, not the raw text: -e API_KEY=abc123 DEBUG=true → env == ['API_KEY=abc123', 'DEBUG=true']; -H "Authorization: Bearer ***" "X-Custom-Header: value" → headers == ['Authorization: Bearer ***', 'X-Custom-Header: value'].
• ✅ Every flag used in the examples (-t -c -a -e -u -H -o) exists in the script's --help with the documented ARGS [ARGS ...] / ENV [ENV ...] / HEADERS [HEADERS ...] forms, and the doc's usage block matches the real help output. Both added rules are accurate for all four failing cases.
• ✅ Markdown hygiene: zero trailing-whitespace lines, final newline added (the parent had none), fence tags untouched.

Notes (non-blocking):

1. The same-class defect still lives in the script itself: the --help epilog examples at scripts/evaluation.py lines 312 and 315 (-a my_server.py eval.xml and -H "Authorization: Bearer ***" eval.xml) fail identically at this head with required: eval_file; only the third epilog example parses. A docs-only PR cannot fix these — flagging for a follow-up in the script.
2. The "one flag, many values" form also sidesteps the repeated-flag overwrite tracked in fix(mcp-builder): accumulate repeated -e/--env and -H/--header values #1896: with the path first, -e A=1 -e B=2 still yields env == ['B=2'] and repeated -H yields only the last header (verified against this head).
3. No other reference/*.md in the repo embeds runnable evaluation.py commands, so the sweep is complete.

No remaining findings at 271b027.