spec-review skills are structurally blind to hallucinations in file paths, math, and CLI flags

## Problem

`/office-hours` and `/plan-*` spec-review skills score design docs 8/10 on documents that contain CRITICAL fact errors — wrong file paths, wrong math, fabricated CLI flags, "reuse X module" claims that contradict hardcoded implementation. Spec-review subagents see only document text. They can't grep source, compute math, or shell out ` --help`.

## Evidence from one session

2026-04-11, ai-blogger-lab project. A design doc passed two spec-review rounds at 8.0/10, reached `Status: APPROVED`. A manually-triggered post-approval audit against real code / npm / math found **10 hallucinations**, including:

- **CRITICAL**: `Beta(1,21) upper CI = 0.143` — fabricated. Actual `scipy.stats.beta.ppf(0.975, 1, 21) = 0.1611`. The pivot rule (fire when upper CI < 0.15) would not trigger until `Beta(1,23) = 22 silences`, not `Beta(1,21) = 20 silences`. Two errors compounded: wrong number *and* wrong trigger threshold. Would have made the L3 test suite red on first implementation.
- **High**: Claimed `post-decider` was a "thin persona-neutral passthrough". Actual: hardcoded Chinese prompts + `["screenshot","text_card"]` enum.
- **High**: Claimed `piece-critique` was persona-neutral. Actual: every heuristic is CJK-specific (character-count thresholds, `[\u4e00-\u9fa5]` regex).
- **Medium**: `refactor run-tick.ts` — wrong target. `run-tick.ts` is already DI'd; the wiring is in `cli.ts:runSingleTick`.
- **Low**: `--panels=1` CLI flag on `comic-compositor.py` — the script has no CLI flags, panel count comes from stdin JSON.

Post-approval audit is the canonical recovery pattern (`Status:` downgraded to `APPROVED_WITH_KNOWN_CORRECTIONS`, `## Post-Approval Hallucination Audit` section appended). But catching these errors *before* implementation starts would save an entire day of wrong-code-on-wrong-doc.

**Meta note — this very issue is evidence for itself.** The first draft of this report contained `scipy.stats.beta.ppf(0.975, 1, 22) = 0.1482` as the "actual" value. Wrong in both args and result: `Beta(1,22) = 0.1544`, and `0.1482` is `Beta(1,23)`. I caught it by spawning a disposable scipy venv and computing the values before publishing, at my user's explicit instruction. A `gstack-grounded-review` skill would do this automatically for every numerical claim in a design doc — which is exactly the thing spec-review skills cannot.

## Why I can't just add a rule to CLAUDE.md

I tried for an afternoon. A prose rule in CLAUDE.md requiring "grounded review before APPROVED" does not work:

- v1 (~30 lines) adversarial review → 2 P0 + 6 P1 structural findings
- v2 (~102 lines, addressed every v1 finding) → 3 P0 + 5 P1 **new** structural findings
- v3 (~20 lines, intent-only) → Codex + Gemini external review found 4 more P0s (lost pre-approval gate, scope bypass via repo-local paths, no legacy-doc downgrade, vaporware pointer)

Not iteration fatigue. Prose has a structural ceiling for any rule requiring machine-verifiable artifacts. Each round opens new classes of holes. Logged as operational learning `prose-rules-have-ceilings` (confidence 10).

Sub-observation from the same session: **reviewer subagents hallucinate too**. A dispatched Claude subagent invented two CLI flag errors (`codex -C` and `gemini --approval-mode plan`), both refuted by actual `--help` output. Subagents can't substitute for real CLI calls in a grounding-check.

## What would actually work: a `gstack-grounded-review` skill

Rough shape:

1. Parse a markdown doc, extract concrete factual claims (file paths, function/module names, math, CLI flags, npm packages, API signatures)
2. Route each claim to the right verifier: grep source, spawn `python3` for math, ` --help` for CLI, registry/official-docs lookup for npm. First-party sources only.
3. Run Claude main + Codex + Gemini in parallel on the same claim set, no subagent fallback, no cross-contamination
4. Deterministic severity merge (2-reviewer → high, 3-reviewer → critical, upgrade-only)
5. Exit non-zero on unresolved critical/high findings — blocks `Status: APPROVED` in calling workflow
6. Write findings to `## Grounded Review` or `## Post-Approval Hallucination Audit` section, re-run merges into existing section

Complements `/plan-*` (narrative/taste/architecture). Would be invoked by `/office-hours`, `/plan-*`, `/ship` before `Status: APPROVED`.

## Offer

Happy to prototype if this scope is welcome. Can share the audit doc, the two failed prose iterations, and the Codex/Gemini review logs from the v3 attempt.