feat(humanizer): update agent system prompts [skip-docs]

Wave 5 Step 16 — final wave step. Threads humanizer-aware rendering
rules through the three agent prompts that produce user-facing output,
and adds a shape test that locks the structure.

- agents/analyzer-agent.md: documents the humanizer envelope shape
  (userImpactCategory, userActionLanguage, relevanceContext) in the
  Input section; new "Humanizer-aware rendering rules" subsection
  instructs the agent to: render humanized title/description/
  recommendation verbatim, group findings by userImpactCategory, lead
  each line with userActionLanguage, surface relevanceContext when
  not affects-everyone, and skip jargon-translation subroutines.
  --raw fallback documented (v5.0.0 verbatim severity prefiks).
- agents/planner-agent.md: documents the same vocabulary; instructs
  the planner to consume humanized fields from the analysis report,
  preserve titles verbatim, and order actions by both dependencies
  AND userActionLanguage urgency. Translation duties explicitly
  removed from the plan.
- agents/feature-gap-agent.md: replaces the inline t1/t2/t3/t4
  tier-to-prose section ladder with userActionLanguage-driven
  groupings ("Fix soon" → High Impact, "Fix when convenient" →
  Worth Considering, "Optional cleanup"/"FYI" → Explore When Ready);
  instructs skipping findings whose relevanceContext is
  test-fixture-no-impact; --raw fallback documented.

tests/agents/agent-prompt-shape.test.mjs (new, +6 tests, 786 → 792):
  - structural: humanized field reference + frontmatter preserved
  - per-agent anchors: analyzer groups by userImpactCategory; planner
    orders by userActionLanguage; feature-gap references
    test-fixture-no-impact
  - global: no "explain what {jargon} means" / "translate jargon" /
    "jargon-translation duty" prose anywhere

Self-audit: Grade A unchanged (config 97/100, plugin 100/100).

plugins/config-audit/agents/analyzer-agent.md

@@ -27,12 +27,23 @@ Analyze all discovered configuration files to:
 You will receive:
 1. Session ID with findings in `~/.claude/config-audit/sessions/{session-id}/findings/`
 2. Scope configuration from `~/.claude/config-audit/sessions/{session-id}/scope.yaml`
-3. Scanner JSON envelope (if available) from scan-orchestrator.mjs
-4. Knowledge base at `{CLAUDE_PLUGIN_ROOT}/knowledge/` for best practices and anti-patterns
+3. Scanner JSON envelope (if available) from scan-orchestrator.mjs — in default mode each finding carries humanizer fields: `userImpactCategory` (e.g., "Configuration mistake", "Conflict", "Wasted tokens", "Missed opportunity", "Dead config"), `userActionLanguage` (e.g., "Fix this now", "Fix soon", "Fix when convenient", "Optional cleanup", "FYI"), and `relevanceContext` ("affects-everyone", "affects-this-machine-only", "test-fixture-no-impact"). The humanizer also replaced `title`/`description`/`recommendation` strings with plain-language equivalents.
+4. Mode flag — when `$RAW_FLAG` is `--raw`, the envelope is v5.0.0 verbatim and humanizer fields are absent; fall back to grouping by raw severity.
+5. Knowledge base at `{CLAUDE_PLUGIN_ROOT}/knowledge/` for best practices and anti-patterns.
+
+## Humanizer-aware rendering rules
+
+- **Render the humanizer's `title`/`description`/`recommendation` verbatim.** Do not paraphrase. The humanizer owns the plain-language vocabulary; if you re-derive prose, the toolchain ends up with two competing voices.
+- **Group findings by `userImpactCategory`.** This replaces severity-bucket grouping in the report. The categories are pre-translated — do not invent your own bucket names.
+- **Lead each finding line with `userActionLanguage`.** This replaces raw severity prefiks ("critical", "high", "medium") in the report. Order findings within each category by urgency: "Fix this now" → "Fix soon" → "Fix when convenient" → "Optional cleanup" → "FYI".
+- **Surface `relevanceContext` when it isn't `affects-everyone`.** The user wants to know whether a fix touches shared config or just their own machine; mention "affects only this machine" or "test-fixture, no real impact" inline.
+- **Do not include "explain what X means" subroutines.** Jargon translation is owned by the humanizer; if a term still feels obscure, that's a humanizer-data gap to file as a follow-up, not a paraphrase to invent here.
+
+In `--raw` mode, fall back to v5.0.0 severity prefiks and verbatim scanner titles — but flag in the report header that the output is unhumanized.
 
 ## Task
 
-1. **Load all findings**: Read all `*.yaml` files from findings directory
+1. **Load all findings**: Use the Read tool on all `*.yaml` files from findings directory
 1.5. **Load scanner results**: If a scanner JSON envelope exists in the session directory, extract all findings. Cross-reference against `knowledge/anti-patterns.md` to add remediation context. Note any CA-{prefix}-NNN finding IDs in the report.
 2. **Build hierarchy map**: Order files by level (managed -> global -> project), visualize inheritance
 3. **Detect conflicts**: Compare settings across hierarchy levels, note which level wins
@@ -40,7 +51,7 @@ You will receive:
 5. **Identify optimizations**: Rules to globalize, missing configs, orphaned files
 6. **Security scan**: Aggregate secret warnings, check for insecure patterns
 7. **CLAUDE.md quality assessment**: Score each file against rubric, assign letter grades
-8. **Generate report**: Write comprehensive markdown report
+8. **Generate report**: Write comprehensive markdown report — group findings by `userImpactCategory`, lead with `userActionLanguage`
 
 ## Output