fix(intel): newline forges prompt rows in three sibling prompt-context modules (out of #5857's scope) #5881

Found while sweeping #5857. That issue scoped the fix to chat-analyst-context.ts + brief-story-context.ts; the same weakness is present in three sibling prompt-context modules that were outside its scope.

Problem

Same root cause as #5850/#5857: sanitizeForPrompt (server/_shared/llm-sanitize.js:79) strips injection phrases but deliberately preserves a lone newline — it splits on \n, filters role-prefixed lines, rejoins, then collapses only runs of 2+ whitespace. Any block composed as `- ${x}` (or `Label: ${x}`) and joined with \n therefore lets one feed item forge an extra row the model reads as a separate, real datum.

sanitizeForPromptLine() now exists in server/_shared/llm-sanitize.js (added by the #5857 fix) — it is sanitizeForPrompt plus a full whitespace-run collapse. These sites should adopt it.

Affected sites (by inspection — not separately reproduced)

1. server/worldmonitor/intelligence/v1/deduct-situation.ts:69-77 — `- "${title}" — Yes ${pct}% (${vol} volume)` joined by \n under a ## Prediction Market Odds header. title goes through sanitizeForPrompt only. Exact same shape as buildPredictionMarkets, which fix(intel): newline in feed text forges extra bullets in analyst prompt blocks (sibling formatters missed by #5850) #5857 fixed.

2. server/worldmonitor/intelligence/v1/brief-why-matters-prompt.ts:270-278 — Headline: / Description: / Source: / Severity: / Category: / Country: lines joined by \n. Fields come from sanitizeStoryFields(story); a newline in any of them forges an extra Key: value line the model reads as a real field.

3. server/worldmonitor/intelligence/v1/_country-brief-context.ts:210-213 — headlineLines are raw item.title values, not sanitized at all, joined by \n under a Headlines: marker. This one is worse than the fix(intel): newline in feed text forges extra bullets in analyst prompt blocks (sibling formatters missed by #5850) #5857 sites: it is missing both the delimiter guard and the LLM prompt injection: news headlines use structural-only sanitizer, bypassing semantic injection filters #3724 content sanitization.

Also worth checking scripts/lib/brief-llm.mjs:284-288 and :522-526 for the same shape.

Fix

sanitizeForPromptLine(...) at each line-composing call site; keep plain sanitizeForPrompt on prose bodies whose internal newlines are legitimate. For (3), that also closes the missing-sanitization gap.

Acceptance criteria

• A failing test first, per affected module: one feed item containing \n must yield exactly the number of rows its payload declares.
• Prove the guard has teeth — neutering sanitizeForPromptLine back to sanitizeForPrompt must turn each new test red (this is how the fix(intel): newline in feed text forges extra bullets in analyst prompt blocks (sibling formatters missed by #5850) #5857 guards were verified: 17 tests went red under that mutant).

Severity

P2, same threat model as #5857: exploiting it needs attacker-influenced text in a seeded feed, which is upstream- rather than user-controlled — but that is exactly what #3724's hard-sanitization policy at these prompt boundaries exists for, and the fix is a few characters per call site.