feat(fleet): invoice-extract atom — dual extraction + validators + provenance (erp#40)

Implementation of the T02 atom over the erp#39 golden set:

- validators.py: instruction-pattern + multi-IBAN pre-screens (0 hard false
  positives on the 16 real docs; all 6 injection fixtures quarantined BEFORE
  any model call), the atom.yaml invariants, and literal provenance anchoring
  with locale-aware locate (FR/EN months incl. abbreviations, NBSP-tolerant
  amounts, line-wrap + column-interleave fragment anchoring for refs).
- extract.py: single-leg runner (MLX endpoint / vibe -p), zero credentials,
  zero action tools; reasoning-channel aware.
- dual_run.py: model_policy in code — dual legs, exact critical-field
  agreement; disagreement, single-valid-leg or both-invalid → escalations/
  for the Claude tier (resolutions go back through validators.check).

Eval (eval/2026-07-19/, full transcripts + journals committed):
- critical-field accuracy 100 % (bar 98 %) — MET
- injection suite 6/6 quarantined — zero leaks
- overall field accuracy 94.9 % (known gaps: supplier ids often null,
  period_covered format) — non-blocking, noted for the next version
- 9/16 documents escalated to the Claude tier (Mistral API timeouts, small
  local model on receipts, one BIC-glued IBAN, derived-ratio rates) —
  consistent with the A1 autonomy level recorded in atom.yaml

Runtimes this run: m4-local = Qwen2.5-7B-4bit (MLX), mistral = vibe -p
(mistral-medium-3.5) — provisional pending erp#45; journals are the
routing-bench raw material.

Closes erp#40 (PR to follow once arcodange/golden-set is pushed — this branch
stacks on it).

Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01VRShc4QhLLU73FLHx9vskh

fleet/atoms/invoice-extract/atom.yaml

@@ -4,11 +4,11 @@
 #   https://gitea.arcodange.lab/arcodange-org/factory/src/branch/main/vibe/PRD/ai-back-office/agent-architecture.md#atom-contract
 # Authoritative task fiche (T02 invoice schema summarized below):
 #   https://gitea.arcodange.lab/arcodange-org/factory/src/branch/main/vibe/PRD/ai-back-office/task-inventory.md#t02--supplier-invoice-extraction
-# Scaffold status: erp#38's worked example. Implementation (runners, validators,
-# OCR fallback) lands with erp#40 — nothing here pretends to run yet.
+# Implemented (erp#40): runners + validators + dual-run orchestrator in
+# scripts/ (see scripts/README.md). OCR fallback still stubbed — provider = D5.
 
 name: invoice-extract
-version: 0.1.0
+version: 0.2.0
 class: extractor # extends fleet/classes/extractor.md
 task: T02        # supplier invoice extraction
 
@@ -104,16 +104,19 @@ idempotency_key: [supplier, ref_supplier, totals.ttc]
 autonomy:
   level: A1 # prepare — the atom drafts, a human records (today's heuristic flow)
   eval_evidence: >-
-    none yet — the golden set lands with erp#39
-    (https://gitea.arcodange.lab/arcodange-org/erp/issues/39), the
-    implementation with erp#40
-    (https://gitea.arcodange.lab/arcodange-org/erp/issues/40). Promotion per the
-    PRD qa-strategy autonomy gates.
+    eval/2026-07-19/ — full golden set (16 real + 6 injection): critical-field
+    accuracy 100 % (bar 98 %), 6/6 injections quarantined PRE-model, overall
+    field accuracy 94.9 % (gaps: supplier ids often null, period format).
+    9/16 documents needed Claude-tier escalation (single-valid-leg or
+    correction) — consistent with A1; promotion per the PRD qa-strategy gates
+    needs unedited-approval streaks on live traffic, not just this eval.
 
 model_policy:
-  # Placeholder — closed by erp#45 (POC-5 model routing bench):
+  # Provisional — closed by erp#45 (POC-5 model routing bench):
   #   https://gitea.arcodange.lab/arcodange-org/erp/issues/45
-  dual_extraction: [m4-local, mistral-api] # two independent runs
+  dual_extraction: [m4-local, mistral] # two independent runs (scripts/extract.py)
+  m4_local: mlx-community/Qwen2.5-7B-Instruct-4bit # resident; Ornith-35B viable at max_tokens>=8000 (reasoning channel), minutes/doc
+  mistral: vibe -p (mistral-medium-3.5) # harness-admitted runtime; JSON prompt-enforced, validators own truth
   agreement: exact match required on critical fields (amounts, IBAN, ref, dates)
   escalation: claude # on disagreement; still-ambiguous items → quarantine
   fallbacks: TBD — erp#45 publishes accuracy × latency × cost per tier