feat: Document-type awareness via --type flag (instructions/policy/persona/reference)

[roli-lpci]

Problem

lintlang currently treats every file as a generic system prompt. This creates false positives:

• TOOLS.md (reference cheat sheet) gets flagged for missing priority ordering — irrelevant for reference docs
• SOUL.md (persona guide) gets dinged for intentional vagueness — correct for persona, wrong for instructions
• Policy files with NEVER / FORBIDDEN constraints get flagged as negative instructions — but these are security boundaries, not style issues

~30-40% of MEDIUM findings in real-world audits are type-context errors.

Proposed Solution

Add --type [instructions|policy|persona|reference] flag with auto-detection fallback:

class DocType(Enum):
    INSTRUCTIONS = "instructions"  # System prompts, SKILL.md files
    POLICY       = "policy"        # Red lines, security rules, hard constraints
    PERSONA      = "persona"       # Character, tone, soul docs
    REFERENCE    = "reference"     # Cheat sheets, tool lists, lookup tables

def detect_doc_type(text: str, filename: str) -> DocType:
    POLICY_SIGNALS    = ["red line", "never", "forbidden", "hard constraint", "non-negotiable"]
    PERSONA_SIGNALS   = ["you are", "your vibe", "who you are", "core truths", "soul"]
    REFERENCE_SIGNALS = ["## examples", "```", "cheat sheet", "quick reference", "lookup"]

Per-type scoring: skip HERM-2/3 for REFERENCE docs, reduce vagueness penalty for PERSONA, treat security negatives in POLICY as features.

Evidence

Derived from CS-001/CS-002 self-audit (AGENTS.md, SOUL.md, TOOLS.md), CS-003 (NemoClaw SKILL.md, 141 HN pts), CS-004 (Microsoft AgentAsJudge).

Estimated effort: Medium (1-2 days). Auto-detection heuristics, CLI flag, per-type weight tables.