feat(compute): MVP-A inference execution contract (mbongo-infer:v1, kaayu-cv-extract:v1)

.gitattributes

@@ -0,0 +1,2 @@
+crates/mbongo-compute/profiles/** -text
+test-vectors/inference/** -text

.github/workflows/ci.yml

@@ -86,6 +86,12 @@ jobs:
       # GitHub-hosted runners have no SEV-SNP guest. See the gated job below.
       - name: Mbongo Compute Hardware Attestation
         run: cargo test -p mbongo-compute --test hardware_attestation --locked
+      # MVP-A: the mbongo-infer:v1 execution contract and the kaayu-cv-extract:v1
+      # profile (docs/product/MVP_INFERENCE_CONTRACT_v1.md). Pins the golden
+      # vector in test-vectors/inference/ and the worker validation order. No
+      # model runs here; MVP-B adds execution.
+      - name: Mbongo MVP Inference Contract
+        run: cargo test -p mbongo-compute --test inference_contract --locked
 
   # ---------------------------------------------------------
   # Mbongo Compute v0.4 Vertical (run on PR + push)

crates/mbongo-compute/profiles/kaayu-cv-extract-v1/job-spec.json

@@ -0,0 +1 @@
+{"contract":"mbongo-infer:v1","generation":{"max_tokens":400,"seed":42,"temperature_milli":0,"top_p_milli":1000},"limits":{"max_input_bytes":10000,"max_output_bytes":8000},"model":{"artifact_name":"Qwen3-4B-Instruct-2507-Q4_K_M.gguf","artifact_sha256":"3605803b982cb64aead44f6c1b2ae36e3acdb41d8e46c8a94c6533bc4c67e597","artifact_size_bytes":2497281120,"family":"Qwen3","format":"gguf","license":"Apache-2.0","name":"Qwen3-4B-Instruct-2507","quantization":"Q4_K_M","source":"https://huggingface.co/unsloth/Qwen3-4B-Instruct-2507-GGUF","source_revision":"a06e946bb6b655725eafa393f4a9745d460374c9"},"profile":{"id":"kaayu-cv-extract","input_contract":"kaayu-cv-extract-input:v1","output_contract":"kaayu-cv-profile:v1","output_schema_sha256":"81521447e7b7b45c670ff1717d608fc2304765ab530587aba2370ba0a9191af5","prompt_template_sha256":"6c950b4863bc339af94d904dbb609b3dd9098259740f01a6539406e641276d63","version":1},"runtime":{"commit":"620d2225cc436ab0b001e3200e590e3b286ef9ab","family":"llama.cpp","license":"MIT","mode":"llama-server","revision":"v0.7.0"}}

crates/mbongo-compute/profiles/kaayu-cv-extract-v1/output-schema.json

@@ -0,0 +1,52 @@
+{
+  "$schema": "https://json-schema.org/draft/2020-12/schema",
+  "$id": "kaayu-cv-profile:v1",
+  "title": "kaayu-cv-profile:v1",
+  "type": "object",
+  "additionalProperties": false,
+  "required": ["summary", "seniority", "location", "skills", "experience", "education", "languages", "certifications"],
+  "properties": {
+    "summary": { "type": "string", "maxLength": 600 },
+    "seniority": { "type": "string", "enum": ["unknown", "junior", "mid", "senior", "lead", "executive"] },
+    "location": { "type": "string", "maxLength": 120 },
+    "skills": { "type": "array", "maxItems": 40, "items": { "type": "string", "maxLength": 60 } },
+    "experience": {
+      "type": "array", "maxItems": 15,
+      "items": {
+        "type": "object", "additionalProperties": false,
+        "required": ["title", "organization", "start", "end", "description"],
+        "properties": {
+          "title": { "type": "string", "maxLength": 120 },
+          "organization": { "type": "string", "maxLength": 120 },
+          "start": { "type": "string", "maxLength": 10 },
+          "end": { "type": "string", "maxLength": 10 },
+          "description": { "type": "string", "maxLength": 300 }
+        }
+      }
+    },
+    "education": {
+      "type": "array", "maxItems": 10,
+      "items": {
+        "type": "object", "additionalProperties": false,
+        "required": ["degree", "institution", "year"],
+        "properties": {
+          "degree": { "type": "string", "maxLength": 120 },
+          "institution": { "type": "string", "maxLength": 120 },
+          "year": { "type": "string", "maxLength": 10 }
+        }
+      }
+    },
+    "languages": {
+      "type": "array", "maxItems": 10,
+      "items": {
+        "type": "object", "additionalProperties": false,
+        "required": ["language", "level"],
+        "properties": {
+          "language": { "type": "string", "maxLength": 40 },
+          "level": { "type": "string", "enum": ["unknown", "basic", "intermediate", "fluent", "native"] }
+        }
+      }
+    },
+    "certifications": { "type": "array", "maxItems": 15, "items": { "type": "string", "maxLength": 100 } }
+  }
+}

crates/mbongo-compute/profiles/kaayu-cv-extract-v1/prompt-template.txt

@@ -0,0 +1,15 @@
+You extract a structured, identity-free candidate profile from the text of one CV.
+
+Rules:
+1. Read only the CV text supplied in the user message. Do not use outside knowledge about any person or company.
+2. Produce exactly one JSON object that conforms to the kaayu-cv-profile:v1 schema. Output nothing before or after the JSON object. No markdown, no comments, no explanations.
+3. Never include the candidate's name, email address, phone number, postal address, date of birth, photo description, or any government or social identifier. "location" is a city, region or country only.
+4. Never invent facts. When a field cannot be determined from the CV, use an empty string for a text field and an empty array for a list field.
+5. "summary": at most three sentences describing the professional profile, written in the third person without a name.
+6. "seniority": one of "unknown", "junior", "mid", "senior", "lead", "executive", judged from years of experience and titles; "unknown" when unclear.
+7. "skills": distinct skill names as they appear in the CV, each a short noun phrase, most relevant first, at most 40.
+8. "experience": one entry per position, most recent first, with "title", "organization", "start", "end" (use "present" for a current position, "" when unknown; dates as written, e.g. "2021-03" or "2021"), and a "description" of at most two sentences.
+9. "education": one entry per degree or programme with "degree", "institution" and "year" ("" when unknown).
+10. "languages": each with "language" and "level" from "unknown", "basic", "intermediate", "fluent", "native".
+11. "certifications": certification or licence names only.
+12. Keep every string in the language of the CV. Do not translate.