qualia-framework 6.7.0 → 6.8.0

package/bin/security-scan.js

@@ -192,7 +192,7 @@ function categoryScore(findings) {
   return { weighted_sum: ws, score: Math.max(1, 5 - Math.floor(ws / 8)), counts };
 }
 
-function renderMarkdown({ findings, paths, score }) {
+function renderMarkdown({ findings, paths, score, skipped = [] }) {
   const lines = [];
   lines.push(`# Qualia security scan — ${new Date().toISOString()}`);
   lines.push("");
@@ -202,10 +202,16 @@ function renderMarkdown({ findings, paths, score }) {
   lines.push(`**Score:** ${score.score} / 5 (weighted_sum=${score.weighted_sum})`);
   lines.push("");
 
-  if (findings.length === 0) {
+  if (skipped.length > 0) {
+    lines.push(`⚠️ **${skipped.length} file(s) could not be scanned** — result is NOT clean (score degraded, non-zero exit):`);
+    for (const s of skipped) lines.push(`- \`${s.file}\` — ${s.error}`);
+    lines.push("");
+  }
+
+  if (findings.length === 0 && skipped.length === 0) {
     lines.push("✅ Clean.");
     lines.push("");
-  } else {
+  } else if (findings.length > 0) {
     findings.sort((a, b) => severityOrder(a.severity) - severityOrder(b.severity));
     for (const f of findings) {
       lines.push(`### [${f.severity}] ${f.name}`);
@@ -346,6 +352,7 @@ function main() {
   const paths = args.paths || defaultPaths();
   const findings = [];
   const scanned = [];
+  const skipped = [];
 
   for (const p of paths) {
     try {
@@ -356,14 +363,21 @@ function main() {
       const content = fs.readFileSync(p, "utf8");
       scanned.push(p);
       findings.push(...scanContent(p, content));
-    } catch {}
+    } catch (e) { skipped.push({ file: p, error: e.message }); }
   }
 
   const score = categoryScore(findings);
+  // An unreadable file must not silently pass as clean. Degrade the score
+  // (min 1) and force a non-clean exit when any target could not be scanned.
+  if (skipped.length > 0) score.score = Math.max(1, score.score - 1);
+
+  // A scan that couldn't read every target must not exit clean — surface at
+  // least HIGH (1) so an unreadable file can't slip through a CI gate as 0.
+  const exitCode = skipped.length > 0 ? Math.max(1, exitCodeFor(findings)) : exitCodeFor(findings);
 
   if (args.json) {
-    process.stdout.write(JSON.stringify({ findings, scanned, score }, null, 2) + "\n");
-    process.exit(exitCodeFor(findings));
+    process.stdout.write(JSON.stringify({ findings, scanned, skipped, score }, null, 2) + "\n");
+    process.exit(exitCode);
   }
 
   // --deep emits a prompt pack for the /qualia-secure skill to spawn agents.
@@ -373,17 +387,17 @@ function main() {
     try { fs.mkdirSync(planningDir, { recursive: true }); } catch {}
     const staticPath = path.join(planningDir, "security-scan.md");
     const promptPath = path.join(planningDir, "security-deep-prompt.md");
-    fs.writeFileSync(staticPath, renderMarkdown({ findings, paths: scanned, score }));
+    fs.writeFileSync(staticPath, renderMarkdown({ findings, paths: scanned, score, skipped }));
     fs.writeFileSync(promptPath, renderDeepPromptPack({ findings, scanned, score }));
     console.log(`Wrote ${staticPath}`);
     console.log(`Wrote ${promptPath}`);
     console.log("");
     console.log("Now run `/qualia-secure` in a Claude session — it will read the prompt pack");
     console.log("and spawn the red/blue/auditor agents in parallel.");
-    process.exit(exitCodeFor(findings));
+    process.exit(exitCode);
   }
 
-  const md = renderMarkdown({ findings, paths: scanned, score });
+  const md = renderMarkdown({ findings, paths: scanned, score, skipped });
   if (args.write) {
     const outDefault = path.join(process.cwd(), ".planning", "security-scan.md");
     const out = args.writePath || outDefault;
@@ -395,7 +409,7 @@ function main() {
     process.stdout.write(md);
   }
 
-  process.exit(exitCodeFor(findings));
+  process.exit(exitCode);
 }
 
 module.exports = { main, scanContent, defaultPaths, categoryScore, renderDeepPromptPack, SECRET_PATTERNS, CONFIG_SMELLS };

package/bin/trust-score.js

@@ -4,6 +4,7 @@
 const fs = require("fs");
 const path = require("path");
 const os = require("os");
+const { spawnSync } = require("child_process");
 const pc = require("./plan-contract.js");
 const ledger = require("./state-ledger.js");
 const { binFiles } = require("./runtime-manifest.js");
@@ -16,6 +17,12 @@ const HOMES = [
 
 const REQUIRED_BIN = binFiles();
 
+// Spine scripts: the load-bearing executables that, if they fail to even parse,
+// render the install dead. binFiles() does not include these, so probe them
+// explicitly — existence AND Node loadability (`--check`). A syntax-broken
+// cli.js (the C1 regression) must force overall status off PASS.
+const SPINE_SCRIPTS = ["cli.js", "install.js", "state.js"];
+
 const REQUIRED_HOOKS = [
   "session-start.js", "auto-update.js", "branch-guard.js", "pre-push.js",
   "pre-deploy-gate.js", "migration-guard.js", "git-guardrails.js",
@@ -54,10 +61,27 @@ function inspectInstall(homes) {
     return { status: "fail", score: 0, issues: ["no Qualia install config found"], targets: [] };
   }
   const issues = [];
+  let spineBroken = false;
   for (const home of homes) {
     for (const f of REQUIRED_BIN) {
       if (!exists(path.join(home.dir, "bin", f))) issues.push(`${home.name}: missing bin/${f}`);
     }
+    // Spine probe: Node loadability of spine scripts that ARE installed.
+    // cli.js / install.js run via npx / the installer and are not part of the
+    // runtime manifest, so their ABSENCE is not a failure (state.js absence is
+    // already reported via REQUIRED_BIN). But any spine script that IS present
+    // must pass `--check` — a syntax-broken installed spine forces status off
+    // PASS (this is the check that catches a dead cli.js).
+    for (const f of SPINE_SCRIPTS) {
+      const fp = path.join(home.dir, "bin", f);
+      if (!exists(fp)) continue;
+      const r = spawnSync(process.execPath, ["--check", fp], { encoding: "utf8" });
+      if (r.status !== 0) {
+        const detail = (r.stderr || r.error?.message || "").split("\n")[0].trim();
+        issues.push(`${home.name}: spine bin/${f} fails --check${detail ? ` (${detail})` : ""}`);
+        spineBroken = true;
+      }
+    }
     if (home.name === "Claude") {
       if (!exists(path.join(home.dir, "CLAUDE.md"))) issues.push("Claude: missing CLAUDE.md");
       if (!exists(path.join(home.dir, "settings.json"))) issues.push("Claude: missing settings.json");
@@ -67,6 +91,16 @@ function inspectInstall(homes) {
       if (!exists(path.join(home.dir, "config.toml"))) issues.push("Codex: missing config.toml");
     }
   }
+  // A broken spine is a hard failure: the install cannot run. Force fail status
+  // (drives overall status to FAIL) and zero the score regardless of other issues.
+  if (spineBroken) {
+    return {
+      status: "fail",
+      score: 0,
+      issues,
+      targets: homes.map((h) => h.name),
+    };
+  }
   return {
     status: issues.length ? "degraded" : "pass",
     score: issues.length ? Math.max(6, 20 - issues.length * 2) : 20,

package/docs/onboarding.html

@@ -546,7 +546,7 @@
       <div class="kit-group">
         <h4>Lock decisions before code</h4>
         <dl>
-          <dt>/qualia-discuss</dt><dd>Without args: project-level non-technical discovery (mandatory at kickoff). With <code>N</code>: phase-level technical alignment with locked decisions and ADRs.</dd>
+          <dt>/qualia-scope</dt><dd>Without args: project-level non-technical discovery (mandatory at kickoff). With <code>N</code>: phase-level archetype-aware grill with locked decisions and ADRs, gated on Definition-of-Done.</dd>
           <dt>/qualia-research</dt><dd>Deep-research a niche library or domain before planning a phase.</dd>
         </dl>
       </div>

package/hooks/migration-guard.js

@@ -161,12 +161,12 @@ if (realCreateTables.length > 0 && !/ENABLE\s+ROW\s+LEVEL\s+SECURITY/i.test(scan
 }
 
 if (errors.length > 0) {
-  console.log("⬢ Migration guard — dangerous patterns found:");
+  console.error("⬢ Migration guard — dangerous patterns found:");
   for (const e of errors) {
-    console.log(`  ✗ ${e}`);
+    console.error(`  ✗ ${e}`);
   }
-  console.log("");
-  console.log("Fix these before proceeding. If intentional, ask Fawzi to approve.");
+  console.error("");
+  console.error("Fix these before proceeding. If intentional, ask Fawzi to approve.");
   _trace("migration-guard", "block", { errors });
   process.exit(2);
 }

package/hooks/pre-deploy-gate.js

@@ -73,7 +73,7 @@ function runGate(label, cmd, args, { required = true } = {}) {
   const r = spawnSync(cmd, args, {
     stdio: ["ignore", "pipe", "pipe"],
     encoding: "utf8",
-    timeout: 180000,
+    timeout: 120000,
     shell: process.platform === "win32",
   });
   if (r.status === 0) {
@@ -170,7 +170,10 @@ function enforceShipPolicy() {
 
   // If this is not a Qualia-managed project, keep the legacy behavior: run the
   // quality/security gates but do not invent state.
-  if (!state || !status) return;
+  if (!state || !status) {
+    console.error("⬢ pre-deploy-gate: no ship-state (brownfield) — skipping ship-policy gate, quality gates still run");
+    return;
+  }
 
   const shippable = status === "polished" || (status === "verified" && verification === "pass");
   if (!shippable && !force) {
@@ -281,9 +284,16 @@ if (fs.existsSync("tsconfig.json")) {
 }
 
 // Lint (with QUALIA_SKIP_LINT=1 escape — for the documented "lint blocks
-// ship" friction when the lint failures are pre-existing or auto-fixer-broken)
+// ship" friction when the lint failures are pre-existing or auto-fixer-broken).
+// The skip is OWNER-only, mirroring the QUALIA_SHIP_FORCE role-gate above.
 if (hasScript("lint")) {
-  if (process.env.QUALIA_SKIP_LINT === "1") {
+  const skipLint = process.env.QUALIA_SKIP_LINT === "1";
+  const lintRole = String(readConfig().role || "").toUpperCase();
+  if (skipLint && lintRole !== "OWNER") {
+    const lintState = readState();
+    blockDeploy("QUALIA_SKIP_LINT is OWNER-only.", (lintState && lintState.next_command) || "/qualia");
+  }
+  if (skipLint) {
     console.log("  ⚠ Lint skipped (QUALIA_SKIP_LINT=1)");
     _trace("pre-deploy-gate", "skip-lint", { reason: "QUALIA_SKIP_LINT=1" });
   } else {

package/hooks/stop-session-log.js

@@ -91,13 +91,16 @@ try {
   // dedupe marker, so it's a cheap no-op on every turn except the one right
   // after a ship. Wrapped + unref'd so it never blocks or breaks the session.
   try {
-    const { spawn } = require("child_process");
-    const child = spawn(
-      process.execPath,
-      [path.join(__dirname, "..", "bin", "auto-report.js")],
-      { cwd: process.cwd(), detached: true, stdio: "ignore" },
-    );
-    child.unref();
+    const autoReportPath = path.join(__dirname, "..", "bin", "auto-report.js");
+    if (fs.existsSync(autoReportPath)) {
+      const { spawn } = require("child_process");
+      const child = spawn(
+        process.execPath,
+        [autoReportPath],
+        { cwd: process.cwd(), detached: true, stdio: "ignore" },
+      );
+      child.unref();
+    }
   } catch {}
 
   // ── Skip if too soon since last write ────────────────────

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "qualia-framework",
-  "version": "6.7.0",
+  "version": "6.8.0",
   "description": "Claude Code and Codex workflow framework by Qualia Solutions. Plan, build, verify, ship.",
   "bin": {
     "qualia-framework": "./bin/cli.js"
@@ -54,7 +54,11 @@
     "docs/onboarding.html",
     "CLAUDE.md",
     "AGENTS.md",
-    "guide.md"
+    "guide.md",
+    "SOUL.md",
+    "FLAGS.md",
+    "TROUBLESHOOTING.md",
+    "CHANGELOG.md"
   ],
   "engines": {
     "node": ">=18"

package/qualia-design/design-rubric.md

@@ -33,7 +33,7 @@ A whole-app `/qualia-polish` scores all 9 dimensions across multiple representat
 
 | Score | Criteria |
 |---|---|
-| 1 | Inter / Roboto / Arial / system-ui / Space Grotesk as primary. Or single weight throughout. |
+| 1 | Banned font as primary — Inter / Roboto / Arial / Helvetica / system-ui / Space Grotesk / Montserrat / Poppins / Lato / Open Sans. Or single weight throughout. |
 | 2 | Project font loaded but only one weight, no scale, no letter-spacing variation. |
 | 3 | Project font with 2-3 weights, basic scale (h1/h2/body), readable. |
 | 4 | Distinctive display + refined body pair. Letter-spacing varied semantically (tight headlines, open labels). Tabular numerals on numeric data. |
@@ -41,6 +41,8 @@ A whole-app `/qualia-polish` scores all 9 dimensions across multiple representat
 
 Evidence format: `font-family: "<name>" used at line N, scale 14/16/24/40 visible, weights 400/500/700`
 
+The banned-font list above is the machine-enforced counterpart of `bin/slop-detect.mjs` (the static anti-pattern scanner) and `agents/visual-evaluator.md` (the vision evaluator). All three are kept in lockstep — Inter, Roboto, Arial, Helvetica, system-ui, Space Grotesk, Montserrat, Poppins, Lato, Open Sans.
+
 ### 2. Color cohesion
 
 | Score | Criteria |

package/rules/architecture.md

@@ -119,7 +119,7 @@ Read this file (auto-load via skill or `@rules/architecture.md`) when:
 
 - Planning a new module or feature with multiple components.
 - The user requests `/qualia-optimize --deepen` or `--alignment`.
-- A verifier is scoring "Container depth & nesting" (per `rules/design-rubric.md` dimension 8).
+- A verifier is scoring "Container depth & nesting" (per `qualia-design/design-rubric.md` dimension 8).
 - An ADR is being drafted for an architectural fork.
 
 Do **not** auto-load this on quick fixes, copy edits, single-component touch-ups — that wastes instruction budget. Use judgment.

package/rules/grounding.md

@@ -102,7 +102,9 @@ Every skill that spawns an agent must order the prompt from most-stable to most-
 - Per-task content goes LAST. Mixing task-specific files into `<phase_context>` breaks cache on every spawn within the same wave.
 - Reference files via `@path` when the harness auto-expands, OR inline the content — but pick one and stick with it per section (switching styles breaks prefix match).
 
-## Design Quality Rubric (any dimension < 3 = mandatory fix before commit)
+## Design Quality Rubric (pre-commit SUBSET — any dimension < 3 = mandatory fix before commit)
+
+> This 6-dimension table (Typography, Color, Spacing, States, Responsiveness, Accessibility) is the fast pre-commit gate every agent runs inline. It is a **deliberate SUBSET** of the authoritative 9-dimension rubric at `@qualia-design/design-rubric.md` (Typography, Color cohesion, Spatial rhythm, Layout originality, Shadow & depth, Motion intent, Microcopy specificity, Container depth & nesting, Visual system & graphics). For verifier scoring and `/qualia-polish --critique`, the 9-dimension rubric is authoritative — read it, don't score from this subset.
 
 | Dimension       | 1 (Fail)                                | 3 (Acceptable)                           | 5 (Excellent)                              |
 |-----------------|-----------------------------------------|------------------------------------------|--------------------------------------------|

package/rules/speed.md

@@ -26,14 +26,14 @@ MCP servers impose a **token tax**: their tool definitions consume context-windo
 - The MCP returns structured JSON that would be expensive to parse from CLI text output.
 - The MCP enforces governance (RLS-aware queries, scoped DB credentials) that the CLI doesn't.
 
-If a `/skill-name` exists that wraps a CLI, prefer the skill — it's been hardened. Canonical example: `/supabase` skill replaces 32 supabase MCP tools with `supabase` CLI calls and saves substantial token budget.
+If a `/skill-name` exists that wraps a CLI, prefer the skill — it's been hardened. Canonical example: drive Supabase through `npx supabase` (migrations, type generation, local dev, SQL) instead of the Supabase MCP — the CLI hits the same API at zero token cost and replaces the bulk of the MCP tool surface (see `infrastructure.md`).
 
 ### MCP tier-list (when each is justified)
 
 | Server | Always-on? | Justification |
 |---|---|---|
 | `claude-in-chrome` | On-demand | Browser automation has no CLI equivalent; use for QA flows only |
-| `supabase` MCP | **Off** in favor of `/supabase` skill | CLI covers 95% of operations; MCP only if you need branch management interactively |
+| `supabase` MCP | **Off** in favor of `npx supabase` CLI | CLI covers 95% of operations; MCP only if you need branch management interactively |
 | `context7` | On-demand | Library docs at runtime — no CLI alternative for Context7 itself |
 | `notebooklm-mcp` | On-demand | NotebookLM has no CLI; only loaded when researching against existing notebooks |
 | `firecrawl-mcp` | On-demand | Web scraping; only loaded when feature requires it |

package/skills/qualia-idk/SKILL.md

@@ -207,13 +207,13 @@ With all three reports + `<user_confusion>` + `<session_context>` in hand, produ
 |---|---|
 | Plan says built but code has stubs | `/qualia-plan {N} --gaps` → `/qualia-build` → `/qualia-verify` |
 | Verify FAILed and no postmortem ran | `/qualia-postmortem` → `/qualia-plan {N} --gaps` → `/qualia-build` |
-| Stale `.continue-here.md`, ongoing context | `/qualia-resume` → `/qualia` |
+| Stale `.continue-here.md`, ongoing context | `/qualia` (restores from `.continue-here.md` / STATE.md) |
 | Brownfield drift (plan and code diverged hard) | `/qualia-map` → `/qualia-plan {N} --gaps` |
 | Phase context missing (no `/qualia-scope` ran) | `/qualia-scope {N}` → `/qualia-plan {N}` |
 | Specific error, scope clear | `/qualia-fix '<symptom>'` |
 | Performance feels off, no profile | `/qualia-fix --perf '<route>'` or `/qualia-optimize --perf` |
 | Design feels off | `/qualia-polish --critique` then `/qualia-polish` |
-| User is overwhelmed | `/qualia-pause` (save handoff), come back later |
+| User is overwhelmed | `/qualia-handoff` (save handoff), come back later |
 | Truly nothing actionable found | Ask one specific question; don't invent a sequence |
 
 Pick the sequence that fits the actual evidence. Substitute real `{N}` from the Plan-view scan.
@@ -240,6 +240,6 @@ node ${QUALIA_BIN}/qualia-ui.js end "DIAGNOSED" "{first command in the sequence,
 - User knows what they're doing and just wants the next command → `/qualia`
 - User has a specific error message they want fixed → `/qualia-fix '<symptom>'`
 - User wants to review code quality → `/qualia-review`
-- User wants to pause and come back → `/qualia-pause`
+- User wants to pause and come back → `/qualia-handoff`
 
 `/qualia-idk` is specifically for **"I'm not sure what I'm even looking at"** situations. Route to sharper tools when the question is sharper.

package/skills/qualia-polish/REFERENCE.md

@@ -32,6 +32,8 @@ The capture script (`scripts/playwright-capture.mjs`) auto-selects in this order
 3. `~/.cache/ms-playwright/chromium-{version}/chrome-{linux64,linux,mac,win}/chrome` — if Playwright was ever installed for browsers but the package isn't import-resolvable
 4. `which google-chrome` / `chromium` / `chromium-browser` / `chrome` — system browser fallback
 
+All four backends capture **full-page** (top-to-bottom, including below-fold content), not just the first viewport-height — the evaluator scores card grids, CTAs, and footers, so the capture must contain them. Playwright uses `fullPage: true`; the binary-direct backends omit any fixed `--window-size` height clamp so the headless render extends to the full document height.
+
 For backends 3 and 4 (binary-direct), the script uses `--headless=new --screenshot --virtual-time-budget`. Less precise than Playwright's `networkidle` waiting but works without any npm dependency.
 
 Setup hints if all four fail:
@@ -70,6 +72,7 @@ Role: @${QUALIA_AGENTS}/visual-evaluator.md
 </product>
 
 <screenshots>
+One full-page PNG per viewport (captured top-to-bottom, includes below-fold content):
 - mobile (375px):  /tmp/qpl-{ts}/iter-{N}/mobile-375.png
 - tablet (768px):  /tmp/qpl-{ts}/iter-{N}/tablet-768.png
 - desktop (1440px): /tmp/qpl-{ts}/iter-{N}/desktop-1440.png
@@ -150,6 +153,8 @@ the {dim} dimension at {score}. Your single task: fix that one dimension.
 
 ## Iteration log entry (what `loop.mjs record` writes to state.json.iterations[])
 
+`tokens_used` is a **deterministic fixed charge** (`PER_ITER_TOKENS = 14500`) added by `loop.mjs` per iteration — NOT a number reported by the evaluator. The evaluator's JSON no longer carries a `tokens_used` field; trusting a model-guessed count to drive the budget kill-switch was unsound, so the loop charges the constant below for every iteration regardless of what any agent claims.
+
 ```json
 {
   "iteration": 1,
@@ -189,20 +194,20 @@ When the kill trigger fires, the verdict becomes `killed_regression` and `state.
 | 6 | ~90K  | default |
 | 8 | ~120K | hard cap; pass `--budget 150000` to allow |
 
-Per-iteration cost (rough):
-- 3 screenshot reads ≈ 9K
+Per-iteration cost is charged as a **fixed constant** — `loop.mjs` adds `PER_ITER_TOKENS = 14500` per iteration to `state.tokens_used` and the kill-switch trips off that deterministic sum, never off an agent's self-estimate. The breakdown below is the rationale for the constant, not a per-run measurement:
+- 3 full-page screenshot reads ≈ 9K (taller PNGs than viewport-height crops, but still 3 reads)
 - rubric + brief inlined ≈ 2K (cached after iter 1)
 - previous-iteration delta ≈ 0.5K
 - 3 fix-builder spawns × (file read + edit + commit-fix call) ≈ 3K
-- **per-iteration ≈ 14.5K**
+- **per-iteration constant = 14.5K** (deterministic — same charge every iteration)
 
 ## Self-test scenarios (mapping to spec)
 
 | # | Fixture | Expected | Verifier |
 |---|---|---|---|
-| 1 | `fixtures/clean.html` | SUCCESS in 1-2 iterations, all dims ≥ 4 | run capture, run evaluator inline, assert pass |
-| 2 | `fixtures/broken.html` | SUCCESS in 4-6 iters; identifies banned font + gradient + 3-card grid + side-stripe + generic CTA | each fix-builder commits a `qpl-N:` change; final eval all dims ≥ 3 |
-| 3 | Kill-switch | KILL at iter ≤ 4 with `LOOP_REGRESSION_DETECTED` | call `loop.mjs record` 3× with the same fingerprint; assert exit 3 + correct verdict |
+| 1 | `fixtures/clean.html` | Each viewport PNG is full-page (height > viewport-height when the fixture scrolls); SUCCESS in 1-2 iterations, all dims ≥ 4 | run capture, assert PNG is full-page, run evaluator inline, assert pass |
+| 2 | `fixtures/broken.html` | SUCCESS in 4-6 iters; identifies banned font + gradient + 3-card grid + side-stripe + generic CTA, INCLUDING below-fold offenders only visible in a full-page shot | each fix-builder commits a `qpl-N:` change; final eval all dims ≥ 3 |
+| 3 | Kill-switch | KILL at iter ≤ 4 with `LOOP_REGRESSION_DETECTED`; `state.tokens_used` increments by exactly `PER_ITER_TOKENS` (14500) per recorded iteration | call `loop.mjs record` 3× with the same fingerprint; assert exit 3 + correct verdict + deterministic token charge |
 
 The pilot-results doc at `docs/playwright-loop-pilot-results.md` records the actual outcome from `bash scripts/_self-tests.sh` (Scenario 3 is exercised by a deterministic unit-style invocation; Scenarios 1+2 require a real vision pass and are run by Claude when the loop ships).
 

package/skills/qualia-polish/SKILL.md

@@ -32,7 +32,7 @@ The first argument selects the scope. Stage selection follows from scope.
 | `/qualia-polish --redesign` | **Redesign** | ~30m | all + Stage 1 mandatory + 2 vision iterations |
 | `/qualia-polish --critique` | **Critique** | read-only | 0, 4, 5 (no edits) |
 | `/qualia-polish --quick` | **Quick** | ~1m | 0, 2, 7 (gates only, no vision loop) |
-| `/qualia-polish --vibe` | **Aesthetic pivot** | ~3m | token-only direction swap |
+| `/qualia-polish --vibe` | **Aesthetic pivot** | ~3m | token-only direction swap, then 6, 7 (mandatory verify) |
 | `/qualia-polish --loop {url}` | **Loop** | ~5-15m | autonomous see/fix/verify, max 8 iterations |
 
 Other flags: `--register=brand|product` overrides register inference. Vibe-specific flags: `--variants N`, `--extract URL|image`, `--sync`, `--write`. Loop-specific flags: `--brief PATH`, `--max N`, `--viewports 375,768,1440`, `--ref PATH`, `--budget 100000`.
@@ -73,6 +73,21 @@ node ${QUALIA_SKILLS}/qualia-polish/scripts/vibe-extract.mjs --source https://ex
 
 Default flow proposes one opinionated direction per `rules/one-opinion.md`. `--variants` is opt-in only. `--extract` stages a reference screenshot for the visual evaluator to reverse-engineer into DESIGN.md tokens; the user must review low-confidence extracted values before application. `--sync --write` patches DESIGN.md to reflect tokens already present in code.
 
+**Mandatory post-vibe verify (the token swap does NOT terminate the run).** After applying the new tokens, a `--vibe` run MUST NOT stop at the swap — it routes back through **Stage 6 — Verify**. Before claiming done, run:
+
+```bash
+# 1. TypeScript still compiles (a token rename can break a typed theme map)
+npx tsc --noEmit 2>&1 | tail -10
+
+# 2. Anti-pattern scan on the files the swap touched (new tokens can reintroduce slop)
+node bin/slop-detect.mjs {changed_files}
+[ $? -eq 0 ] || { echo "vibe failed — slop-detect found critical issues in the new tokens"; exit 1; }
+```
+
+3. **Scoped vision check (best-effort, only if a dev server is reachable).** Capture one screenshot per viewport via `scripts/playwright-capture.mjs` and spawn a single `agents/visual-evaluator.md` pass scored ONLY on the dimensions a token swap can move: **color, typography, shadow, motion**. (Layout/container/microcopy/graphics are out of scope for `--vibe` — do not score or fix them here.) If no dev server and no resolvable browser backend, skip this step and note it in the report; never block the vibe on missing optional tooling. Any in-scope dimension scoring < 3 is a regression in the new direction — re-tune the offending tokens and re-verify, do not ship.
+
+Only after Stage 6 passes does the `--vibe` run reach Stage 7 (commit & state).
+
 ## Setup gates (non-optional, every scope)
 
 Before any work — design or otherwise — pass these gates. Skipping them produces generic output that ignores the project.
@@ -163,6 +178,8 @@ Each agent receives:
 
 For Component scope: do the work in main context. Read, fix, verify.
 
+**Optional best-effort vision check (Component scope).** If a dev server is reachable (`http://localhost:3000` or a port detected via `lsof`) AND `scripts/playwright-capture.mjs` can resolve a browser backend, capture ONE screenshot at a single viewport (desktop 1440 is sufficient for an isolated component) and spawn one `agents/visual-evaluator.md` pass scored on component-relevant dimensions ONLY — Typography, Color, Shadow, Motion, Microcopy (skip Layout/Container/Graphics per the rubric's scope guard). This is opt-in by environment: if there is no dev server or no resolvable backend, **omit the vision check without error and record that it was unavailable** — never block a component touch-up on missing optional tooling.
+
 **Apply fixes scoped by what's missing:**
 
 | If the file scores low on… | Apply fix from… |
@@ -216,7 +233,7 @@ If a11y &lt; 90 OR axe critical/serious violations: **fix programmatically** (th
 
 ### Stage 4 — Vision loop (Redesign scope only — max 2 iterations)
 
-Use `skills/qualia-polish/scripts/playwright-capture.mjs` (Playwright backend, with Chrome-binary fallback). Capture screenshots at 3 viewports: 375 / 768 / 1440 — these match what `agents/visual-evaluator.md` expects, and what the `--loop` mode captures. Single browser (Chromium) is fine in 2026 — cross-browser CSS rendering differences are vanishingly rare.
+Use `skills/qualia-polish/scripts/playwright-capture.mjs` (Playwright backend, with Chrome-binary fallback). Capture one **full-page** screenshot at each of 3 viewports: 375 / 768 / 1440 — these match what `agents/visual-evaluator.md` expects, and what the `--loop` mode captures. The `height` below sets the viewport window, not the screenshot height — each PNG runs top-to-bottom so the evaluator scores below-fold sections too. Single browser (Chromium) is fine in 2026 — cross-browser CSS rendering differences are vanishingly rare.
 
 ```
 viewports: [
@@ -325,6 +342,6 @@ node ${QUALIA_BIN}/qualia-ui.js end "POLISHED" "{next command — depends on con
 | `bin/slop-detect.mjs not found` | Framework not installed in project | Run `npx qualia install` or pull script from framework repo |
 | `PRODUCT.md missing` | Project predates the PRODUCT.md substrate | Run setup (ask 5 questions, generate). For Component scope, can proceed with nudge. |
 | `Lighthouse not installed` | Optional tool | Skip Stage 3 numeric gates, note in report. Don't fail. |
-| `webapp-testing skill not present` | Optional Anthropic skill | Skip Stage 4 vision loop on Redesign scope. Note in report. Recommend installing. |
+| `playwright-capture.mjs cannot resolve a browser backend` | None of the four backends resolve: no `playwright`/`playwright-core` package, no cached Chromium under `~/.cache/ms-playwright`, no system Chrome/Chromium on PATH | Skip Stage 4 vision loop (and any screenshot-dependent verify). Note in report. Install a backend: `npm i -D playwright && npx playwright install chromium` (or put `google-chrome` on PATH). |
 | Vision loop oscillates between iterations | Rubric not anchored properly | Verify rubric prompt instructs "default to 3, only 1-2 = fix". Hard cap at 2 iterations. |
 | User says "you missed X" after polish completes | Scope was too narrow | Re-run with wider scope (`/qualia-polish` whole-app). Don't argue scope. |

package/skills/qualia-polish/scripts/loop.mjs

@@ -14,12 +14,13 @@
  *
  * Eval JSON contract (what the vision evaluator returns):
  *   {
- *     "iteration": 1,
+ *     "iteration": 1,            // optional; if present must equal state.iteration+1 (monotonicity assertion)
  *     "viewport_results": [{ viewport, scores, top_issues }],
  *     "aggregate_scores": { typography, color, spatial, layout, shadow, motion, microcopy, container, graphics },
- *     "top_issues": [{ dim, severity, description, likely_file, fix }],
- *     "tokens_used": 14500
+ *     "top_issues": [{ dim, severity, description, likely_file, fix }]
  *   }
+ *   NOTE: any ev.tokens_used is IGNORED — the budget charges a fixed
+ *   PER_ITER_TOKENS per iteration (deterministic kill-switch).
  *
  * State JSON shape (written by this script, read by Claude):
  *   {
@@ -35,6 +36,12 @@ import { spawnSync } from "node:child_process";
 
 const DIMS = ["typography", "color", "spatial", "layout", "shadow", "motion", "microcopy", "container", "graphics"];
 
+// Deterministic per-iteration token charge (evaluator + 3 builders ≈ 14500,
+// per REFERENCE.md:197). The budget kill-switch must NOT trust ev.tokens_used —
+// that is a model-guessed number and an under-report would defeat the budget
+// guard. Charge a fixed, known cost per iteration instead.
+const PER_ITER_TOKENS = 14500;
+
 // ── helpers ──────────────────────────────────────────────────────────────
 function loadState(p) {
   if (!existsSync(p)) { console.error(`state file not found: ${p}`); exit(2); }
@@ -98,7 +105,17 @@ function cmdRecord() {
   const state = loadState(statePath);
   const ev = JSON.parse(readFileSync(evalPath, "utf8"));
 
-  state.iteration = (ev.iteration ?? state.iteration + 1);
+  // Monotonicity assertion: each record() advances exactly one iteration. If the
+  // evaluator reports an iteration that disagrees with the script's own counter,
+  // the loop is desynced — fail loudly instead of trusting the eval's number,
+  // which would corrupt regression fingerprinting (state.iteration is the key
+  // used at :108-134) and budget accounting.
+  const expected = state.iteration + 1;
+  if (ev.iteration != null && ev.iteration !== expected) {
+    console.error(`loop desync: eval.iteration=${ev.iteration} expected=${expected}`);
+    exit(2);
+  }
+  state.iteration = expected;
   const scores = ev.aggregate_scores || {};
   const failingDims = DIMS.filter((d) => (scores[d] ?? 3) < 3);
   const total = DIMS.reduce((a, d) => a + (scores[d] ?? 3), 0);
@@ -133,7 +150,8 @@ function cmdRecord() {
     return its.length >= 3 && its.every((n, i, arr) => i === 0 || n === arr[i - 1] + 1);
   });
 
-  state.tokens_used += parseInt(ev.tokens_used || 0, 10) || 0;
+  // Charge the fixed per-iteration cost — do NOT trust ev.tokens_used.
+  state.tokens_used += PER_ITER_TOKENS;
 
   state.iterations.push({
     iteration: state.iteration,
@@ -142,7 +160,7 @@ function cmdRecord() {
     pass,
     failing_dims: failingDims,
     top_issues: issues,
-    tokens_used: ev.tokens_used || 0,
+    tokens_used: PER_ITER_TOKENS,
     timestamp: new Date().toISOString(),
   });