test(qa): align AI oracles with fixtures and honor QA_OUT_DIR (#686)

The ocr oracle demanded six pangram words when ocr-clean.png prints five; the ocr-pdf oracle expected a searchable PDF when the tool's contract is text extraction; media-30s.mp4 was eight seconds long. Fix all three, regenerate tool-contract.json from current schemas, record the new fixture in the manifest, and add a QA_OUT_DIR override so parallel machines stop clobbering each other's lane output. Fixes #677.

tests/fixtures/manifest.json

@@ -402,11 +402,11 @@
     },
     {
       "path": "video/valid/media-30s.mp4",
-      "sourceUrl": "Big Buck Bunny (Blender Foundation, www.bigbuckbunny.org) via archive.org, 8s segment at 30s, 480x270",
+      "sourceUrl": "Big Buck Bunny (Blender Foundation, www.bigbuckbunny.org) via archive.org, 8s segment at 30s, 480x270, stream-looped to a true 30.1s (#677)",
       "license": "CC-BY",
       "toolsServed": [],
-      "sha256": "6ad37435eeb01cc694b918f8588255fead60eb01f2a06ca0dc904070064e9136",
-      "bytes": 247284
+      "sha256": "19c3256788660ee49766497dee364c38bf508050f505f949b8ad41714d7db4de",
+      "bytes": 941146
     },
     {
       "path": "video/valid/speech-10s.mp4",

tests/helpers/installed-ai-output-oracles.ts

@@ -570,17 +570,6 @@ export function expectJapaneseScript(text: string, minimumChars = 4): void {
   );
 }
 
-/** PDF OCR: a real PDF that now carries a selectable text layer. */
-export function expectSearchablePdf(output: Buffer): void {
-  assertOracle(output.subarray(0, 5).toString("latin1") === "%PDF-", "artifact is not a PDF");
-  assertOracle(output.length > 1000, "PDF artifact is implausibly small");
-  const body = output.toString("latin1");
-  assertOracle(
-    body.includes("/Font") || body.includes("BT\n") || body.includes("Tj"),
-    "PDF carries no text-drawing operators, so no OCR layer was added",
-  );
-}
-
 /** Animated cut-outs must stay animated: a single-frame result is a regression. */
 export async function expectAnimatedFrames(output: Buffer, minimumFrames = 2): Promise<void> {
   const meta = await sharp(output, { pages: -1 }).metadata();

tests/qa/api-sweep.mts

@@ -44,10 +44,15 @@ const PASSWORD = process.env.QA_PASSWORD ?? "";
 const CONCURRENCY = Number(process.env.QA_CONCURRENCY ?? 4);
 const PAIRWISE_CAP = Number(process.env.QA_PAIRWISE_CAP ?? 6);
 const FORMAT_WITNESSES = Number(process.env.QA_FORMAT_WITNESSES ?? 3);
+// QA_OUT_DIR lets several machines sweep in parallel without clobbering each
+// other's lane-<mode>.jsonl (#677).
+const OUT_DIR_OVERRIDE = process.env.QA_OUT_DIR;
 /* biome-ignore-end lint/suspicious/noUndeclaredEnvVars: QA harness runs outside Turbo. */
 
 const REPO = join(import.meta.dirname, "..", "..");
-const OUT_DIR = join(REPO, "docs", "qa", "master-20260724", "evidence", "processing-ai", "final");
+const OUT_DIR =
+  OUT_DIR_OVERRIDE ??
+  join(REPO, "docs", "qa", "master-20260724", "evidence", "processing-ai", "final");
 
 /** Class-aware hard timeouts. Exceeding one is a finding, never a silent skip. */
 const TIMEOUT_MS: Record<string, number> = { fast: 240_000, long: 480_000, ai: 900_000 };

tests/qa/lifecycle-sweep.mts

@@ -31,12 +31,15 @@ const BASE = process.env.QA_BASE_URL ?? "http://localhost:13492";
 const USERNAME = process.env.QA_USERNAME ?? "admin";
 const PASSWORD = process.env.QA_PASSWORD ?? "";
 const PROJECT = process.env.QA_COMPOSE_PROJECT ?? "snapotter-qa-f4c2bde9";
+const OUT_DIR_OVERRIDE = process.env.QA_OUT_DIR;
 /* biome-ignore-end lint/suspicious/noUndeclaredEnvVars: QA harness runs outside Turbo. */
 
 const APP = `${PROJECT}-app`;
 const PG = `${PROJECT}-postgres`;
 const REPO = join(import.meta.dirname, "..", "..");
-const OUT_DIR = join(REPO, "docs", "qa", "master-20260724", "evidence", "processing-ai", "final");
+const OUT_DIR =
+  OUT_DIR_OVERRIDE ??
+  join(REPO, "docs", "qa", "master-20260724", "evidence", "processing-ai", "final");
 
 const PNG = resolveFixture(".png", "image") as string;
 const JPG = resolveFixture(".jpg", "image") as string;

tests/qa/tool-contract.json

@@ -3374,7 +3374,7 @@
       },
       {
         "key": "sobelThreshold",
-        "values": [1, 10.5, 20, null]
+        "values": [1, 11, 20, null]
       },
       {
         "key": "square",

tests/qa/verify-ai-tool-matrix.mts

@@ -37,7 +37,6 @@ import {
   expectRedPixelsReduced,
   expectRegionRewritten,
   expectSameSizeButChanged,
-  expectSearchablePdf,
   expectSrtArtifact,
   expectUpscaled,
 } from "../helpers/installed-ai-output-oracles.js";
@@ -102,7 +101,11 @@ const SPEECH_WAV = `${FIXTURES}/audio/valid/speech-10s.wav`;
 const SPEECH_MP4 = `${FIXTURES}/video/valid/speech-10s.mp4`;
 
 /** The English OCR fixture's known answer, checked word by word. */
-const OCR_ENGLISH_TERMS = ["the", "quick", "brown", "fox", "jumps", "over", "lazy", "dog"] as const;
+// ocr-clean.png literally reads "The quick brown fox 12345"; the oracle must
+// demand the fixture's words, not the full pangram (#677).
+const OCR_ENGLISH_TERMS = ["the", "quick", "brown", "fox", "12345"] as const;
+// ocr-scanned.pdf is a chat screenshot about Redis scaling.
+const OCR_PDF_TERMS = ["redis", "replicas", "cache", "bottleneck"] as const;
 
 const CASES: ToolCase[] = [
   {
@@ -398,10 +401,10 @@ const CASES: ToolCase[] = [
     fixture: OCR_CLEAN,
     contentType: "image/png",
     settings: { quality: "fast", language: "en", enhance: false },
-    oracle: "recognizes >=6 of the 8 known English fixture words",
+    oracle: "recognizes >=4 of the 5 words printed in the fixture",
     verify: (out) => {
       const text = out.toString("utf8");
-      expectRecognizedTerms(text, OCR_ENGLISH_TERMS, 6);
+      expectRecognizedTerms(text, OCR_ENGLISH_TERMS, 4);
     },
   },
   {
@@ -409,8 +412,13 @@ const CASES: ToolCase[] = [
     fixture: `${FIXTURES}/document/valid/ocr-scanned.pdf`,
     contentType: "application/pdf",
     settings: { quality: "fast", language: "en", enhance: false },
-    oracle: "returns a PDF carrying text-drawing operators",
-    verify: (out) => expectSearchablePdf(out),
+    // The tool's contract is text extraction (it writes <name>_ocr.txt); a
+    // searchable-PDF output would be a feature, not this oracle (#677).
+    oracle: "extracts the known scanned text as plain text",
+    verify: (out) => {
+      const text = out.toString("utf8");
+      expectRecognizedTerms(text, OCR_PDF_TERMS, 3);
+    },
   },
   {
     toolId: "transcribe-audio",