MVP-B: real CPU inference worker (blocked on the frozen 400-token output budget) #152

MVP-B — Real CPU Inference Worker (mbongo-infer:v1, kaayu-cv-extract:v1)

Child of #148 (milestone MVP v0.1). Baseline 20d46774bdb1890b3f510e0ed8e31552dd41c96a (MVP-A merged, #151).

State: BLOCKED on the frozen MVP-A output budget, not on the worker. Implementation, real start-up verification, real inference and the real live-chain end to end are done; the gate's closing bar (10/10 synthetic CVs structurally valid) is not met, and meeting it requires a change to the frozen kaayu-cv-extract:v1 profile.

Evidence (i7-8850H, 6 threads, CPU only, llama.cpp b11669 = v0.7.0 @ 620d2225, model SHA-256 3605803b…e597 verified exactly):

• 8/10 synthetic CVs → schema-valid canonical JSON; warm P50 48.6 s, P95 55.4 s; cold start 3.5–3.8 s; model hash 10.3 s; peak working set 6.3 GB.
• R08 (long, 12 positions) and R10 (executive, 3 positions) stopped at exactly max_tokens = 400 with finish_reason: "length" → OUTPUT_NOT_JSON → no result, no receipt (correct fail-closed behaviour). With a diagnostic budget of 1,500 both completed as valid JSON: R10 needed 463 tokens (63 s), R08 needed 1,331 tokens (223 s).
• Grammar enforcement works (response_format: json_schema at this revision); every response allowed to finish was schema-shaped. The frozen schema permits ~2,000 tokens of JSON at its maximum, against a 400-token budget.
• Live chain: task accepted, first attempt killed mid-inference (failed closed, nothing persisted), retry after automatic server restart, receipt anchored and verified by the client, result commitment matched, 81 blocks and all logs clean.

Decision needed (change control, MVP_v0.1.md §20): kaayu-cv-extract version 2 —

• (a) tighten output bounds so a maximal valid profile fits the budget (keeps P95 ≤ 120 s; drops older positions / long descriptions) — recommended, with a test that serializes a maximal profile and counts its tokens;
• (b) raise max_tokens with the same schema (completes every case; R08-class CVs take ~220 s on this host, breaking P95 ≤ 120 s);
• (c) both, moderately.

PR: draft, not merged until the profile change lands and the real corpus reaches 10/10.

Refs #148