feat(code-review): add thematic triage grouping

docs/skills/ce-code-review.md

@@ -14,7 +14,7 @@ The compound-engineering ideation chain is `/ce-ideate → /ce-brainstorm → /c
 |----------|--------|
 | What does it do? | Selects reviewer personas based on diff content, dispatches them in parallel, merges findings into one report with confidence gating and auto-fix routing |
 | When to use it | Before opening a PR for sensitive/large work; explicit deep review requested; harness has no built-in `/review` |
-| What it produces | A structured findings report — interactive review, applied fixes, residual work routed via the gate |
+| What it produces | A structured findings report with thematic triage groups, interactive review, applied fixes, and residual work routed via the gate |
 | Modes | Interactive (default), Autofix, Report-only, Headless |
 
 ---
@@ -38,6 +38,7 @@ Generalist code review prompts collapse in predictable ways:
 - **Diff-aware persona selection** — 4 always-on reviewers + 2 CE always-on agents, plus cross-cutting and stack-specific personas chosen for what the diff actually touches
 - **Parallel persona dispatch** — each reviewer focuses on its lens; results return in parallel
 - **Confidence-gated synthesis** — findings merge, dedupe, promote on cross-persona agreement, and route by autofix class
+- **Thematic triage groups** — related findings are summarized by shared context and preferred resolution so large reviews are easier to work through
 - **Severity scale (P0-P3) + autofix class** — separates urgency from action ownership
 - **Four modes** — Interactive, Autofix, Report-only, Headless — for different invocation contexts
 - **Residual Work Gate** — when autofix doesn't resolve everything, structured options for accept / file tickets / continue / stop
@@ -96,7 +97,9 @@ After all dispatched personas return, synthesis:
 - Auto-promotes safe-auto candidates that meet the bar
 - Routes by tier — applied fixes, gated/manual, FYI
 
-The output is one report with calibrated severity, evidence quotes, and explicit ownership — not a flat list of every reviewer's raw output.
+The output is one report with calibrated severity, thematic triage groups, evidence quotes, and explicit ownership — not a flat list of every reviewer's raw output.
+
+Grouping is separate from dedupe. Dedupe collapses duplicate reports of the same finding; triage grouping keeps distinct findings intact while showing which ones share a root cause, subsystem, or resolution path. Large reviews can then be worked by theme without losing stable finding numbers.
 
 ### 6. Plan discovery for requirements verification
 
@@ -180,6 +183,9 @@ Concurrent use note: `mode:report-only` is the only mode safe to run alongside b
 | `<branch name>` | Checks out and reviews against detected base |
 | `base:<sha-or-ref>` | Skips scope detection; reviews current checkout against that ref |
 | `plan:<path>` | Loads the plan for requirements verification |
+| `grouping:auto` | Default; creates thematic triage groups when they reduce review load |
+| `grouping:off` | Disables thematic triage groups and renders severity-only findings |
+| `grouping:always` | Always renders triage groups, even for small reviews |
 | `mode:autofix` | No prompts; apply `safe_auto` only; emit Residual Actionable Work summary |
 | `mode:report-only` | Strictly read-only; safe with concurrent browser tests |
 | `mode:headless` | Skill-to-skill; structured text output |