Codemode script preamble — follow-ups

Deferred from #2431 (preamble results array + setPreamble). Field notes from preview-5 testing, in Misha's words with initial takes.

• Eval for the preamble prompt: make sure agents actually reach for results[0].data / .load(itx) instead of re-fetching or re-pasting. The field test caught one regression (fenced readFile recipe outcompeted the loader footnote) that an eval would have caught pre-merge.
• "Result" tab should show what the agent actually sees. Today it has its own truncation ("first 64 KB as YAML, no highlighting") that is a different mechanism from the history render (scriptResultHistoryLimit, spill, inferred-type + preview). One representation, the agent's.
• Why did the agent defensively writeFile("sopranos-tvmaze-full.json", text) and return { status, bytes, body: JSON.parse(text) }? It didn't need the file OR the full body — the preamble keeps the result available. Suspects: old docs/examples still teach the save-then-read pattern (itx examples, docs.search hits); the prompt says results are retained but doesn't say "so don't save copies". Audit examples + prompt line.
• Do we still need the spill-to-workspace process at all? Loaders read the settlement back through the host — more reliable (spill is best-effort, agent-door-only) and likely faster. The spill file's one remaining use is itx.workspace paging of huge text. Consider retiring the spill and pointing paging at getScriptResult too — would also delete the render's dual-pointer complexity.