Harden /review_prs Stage-2 sketch against `args` arriving as a JSON string

## Problem

The `/review_prs` command (`.abca/commands/review_prs.md`) documents Stage 2 with an *illustrative* `Workflow` sketch (lines ~142–164) that the operator regenerates per invocation. That sketch iterates the PR descriptors with `args.map(...)` directly.

In practice the `Workflow` harness may deliver `args` to the script as a **JSON-encoded string** rather than a parsed array — even when the caller passes a genuine JSON array in the `Workflow` tool call. When that happens the script throws immediately:

```
Error: args.map is not a function ('args.map' is undefined)
```

before any review agent spawns. Because the skill is (correctly) a sketch that gets regenerated each run, every regeneration risks re-hitting this footgun. Observed live during the carve S1–S8 stacked-PR review train (2026-07-28): the first Workflow launch died in ~7ms with `agent_count: 0`; a one-line defensive unpack fixed it and the resumed run proceeded normally.

## Why not freeze the whole script into the skill

The sketch is intentionally adaptable — N==1 skips the Workflow, arg modes differ (explicit list vs. `review-requested:@me` vs. NL filter), and diff baselining differs for a *stack* (each PR vs its parent branch) vs. a flat batch (each vs `main`). Freezing a full script would bake away that variance and pull orchestration mechanics back into a skill that deliberately stays a thin orchestrator. So the fix is a **doc hardening**, not a canned script.

## Proposed fix (small, docs-only)

In the Stage-2 sketch, unpack defensively before iterating and add a half-sentence caveat:

```js
// Workflow may deliver `args` as a JSON string; tolerate either form.
const prs = typeof args === 'string' ? JSON.parse(args) : args
const results = await parallel(prs.map((pr) => () => agent(...)))
```

Add one line to the Notes / Stage-2 prose noting that `Workflow` args can arrive JSON-encoded, so generated scripts should normalize before `.map`.

## Scope

- Docs only: `.abca/commands/review_prs.md`.
- No code, no CDK, no agent-runtime change.
- Self-sufficiency bar for the skill is preserved: a competent agent regenerating the script from the prose will no longer trip on the string/array ambiguity.

## Acceptance criteria

- [ ] Stage-2 sketch normalizes `args` (string-or-array) before `.map`.
- [ ] A one-line caveat in the prose explains why.
- [ ] No other behavior change; the sketch remains a sketch (not a frozen script).

Öffne .abca/commands/review_prs.md und lies die Stage-2-Skizze im Bereich der Zeilen 142–164 sowie den Prosa-Text unter Notes. Aktualisiere die Skizze und die Erläuterung so, dass Workflow args normalisiert werden, unabhängig davon, ob sie vor der Verwendung von .map als JSON string oder als Array eintreffen. Erledigt ist die Aufgabe, wenn die beiden Akzeptanzkriterien in dieser Datei umgesetzt sind, ohne das anpassbare Verhalten der Skizze zu verändern.

Vom Indexierungsmodell aus dem Issue-Text verfasst.