feat(simulation): K4 — closed-loop two-run simulation binds §11 'Closed loop' (closes R-1)

Scripted two-run driver over the run.py composition: run A -> persona verdict
(shared skill artifact) -> §6 promotion gate -> run B on a fresh store. The
marker crosses runs via the promoted wiki layer ONLY - run B reads no inbox,
a rejected verdict is refused fail-closed and its marker never crosses.
Two detach proofs delivered (promotion step removed -> red; verdict exclusion
in bundle_context removed -> red via the '## verdict' section anchor).

Known-limitation note (C-F5, deferred to C3.2): a persona verdict over the
bundle seed's own candidate mints the seed's §4.2 id and is silently shadowed
by first-write-wins; the test has run A propose a distinct candidate.

395 -> 400 tests; README synced (test count + the S10 section now reflects
that D7 has its own scripted closed-loop proof).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

README.md

@@ -13,7 +13,7 @@ human-in-the-loop, and the system learns from the verdicts.
 > **Status:** the D7 build (S5–S10) is complete, and the deterministic **ingest layer**
 > (CSV and SQL source types) has since been added in front of the loop. The deterministic
 > backbone, the agentic loop, the learning loop, and the ingest connectors are wired seam by
-> seam, each proven by load-bearing tests (395 tests, all running offline without an API
+> seam, each proven by load-bearing tests (400 tests, all running offline without an API
 > key). The programme's single budgeted **live model run has been executed and validated** —
 > its artifacts are committed under [`runs/s10/`](runs/s10/) (see below).
 
@@ -130,9 +130,11 @@ layer works and how one would extend it is documented in [`docs/extending.md`](d
 
 ## The live run — S10, executed and validated
 
-Where the MAF sibling proves its loop end-to-end with a scripted offline simulation, this
-repo's end-to-end proof is the programme's single budgeted **real** run (D6: exactly one
-live API run in the whole programme), executed 2026-07-03 against the micro bundle
+The loop's closure is proven offline by a scripted two-run simulation
+(`simulation.py`: run A → persona verdict → §6 promotion gate → run B on a fresh store —
+the marker crosses runs via the gate, and never without it). On top of that offline proof
+sits the programme's single budgeted **real** run (D6: exactly one live API run in the
+whole programme), executed 2026-07-03 against the micro bundle
 [`shared/examples/bygg-energi-mikro/`](shared/examples/bygg-energi-mikro/):
 
 - exit 0 · validator `validated` · checker `approve` on the first attempt · 2 of 12
@@ -156,7 +158,7 @@ Python ≥3.10 · [`claude-agent-sdk`](https://pypi.org/project/claude-agent-sdk
 
 ```bash
 uv sync          # install dependencies
-uv run pytest    # 395 tests — run without any API key and without network
+uv run pytest    # 400 tests — run without any API key and without network
 uv run ruff check . && uv run ruff format --check .
 uv run mypy src  # strict
 ```

src/portfolio_optimiser_claude/simulation.py

@@ -0,0 +1,152 @@
+"""Closed-loop two-run simulation driver (method-spec §11 «Closed loop»; §3 Steps 7–8, §6).
+
+Honesty rule (§1): this is a SCRIPTED stand-in, not a live experiment. The
+driver makes no model calls of its own — both runs use injected ``ModelClient``
+instances (scripted in the suite), and the persona plays the human expert from
+the shared skill artifact. What it proves is the loop's closure: run A →
+persona judgement → §6 promotion gate → run B with a FRESH store, where the
+marker reaches run B's hypothesis prompt through the promoted wiki layer and
+the gated fold ONLY. Run B deliberately reads NO inbox: the persona's raw
+verdict is authored into the inbox (§3 Step 7 write side), but the sole
+sanctioned crossing channel between runs is the gate — a rejected verdict is
+refused there (fail-closed, recorded, nothing written) and its marker must
+never cross. ``timestamp`` is an explicit required argument (no wall-clock
+default) — the whole simulation is deterministic and reproducible.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+from pathlib import Path
+
+from portfolio_optimiser_claude.budget import BudgetMeter
+from portfolio_optimiser_claude.contracts import Contracts, load_contracts
+from portfolio_optimiser_claude.experience import CandidateFeatures
+from portfolio_optimiser_claude.inbox import VerdictDocument
+from portfolio_optimiser_claude.loop import ModelClient, RunResult, run_project
+from portfolio_optimiser_claude.persona import drop_persona_verdict
+from portfolio_optimiser_claude.promotion import PromotionError, promote
+from portfolio_optimiser_claude.run import ComposedRunContext, compose_run_context
+
+_PERSONA_APPROVER = "persona:expert-reviewer"
+_VERDICT_DESCRIPTION = "closed-loop simulation: persona judgement of run A's candidate"
+
+
+@dataclass(frozen=True)
+class SimulationRun:
+    """One simulated run: the §5 composition that fed it plus the §3 result."""
+
+    composed: ComposedRunContext
+    result: RunResult
+
+
+@dataclass(frozen=True)
+class ClosedLoopResult:
+    """The two-run outcome: what run A produced, what the gate did, what run B saw.
+
+    Exactly one of ``promoted_path`` / ``promotion_refusal`` is set: the §6
+    gate either promoted the persona's verdict into the bundle or refused it
+    fail-closed (the refusal reason recorded, nothing written).
+    """
+
+    run_a: SimulationRun
+    verdict: VerdictDocument
+    promoted_path: Path | None
+    promotion_refusal: str | None
+    run_b: SimulationRun
+
+
+def _run_once(
+    client: ModelClient,
+    bundle_dir: Path,
+    contracts: Contracts,
+    *,
+    max_debate_rounds: int,
+    max_attempts: int,
+) -> SimulationRun:
+    # Fresh store (inside compose) and fresh §8 meter per run — nothing crosses
+    # runs except what the promotion gate wrote into the bundle.
+    composed = compose_run_context(bundle_dir, None, k=contracts.data_source.top_k)
+    result = run_project(
+        client,
+        composed.context,
+        meter=BudgetMeter(contracts.termination),
+        max_debate_rounds=max_debate_rounds,
+        max_attempts=max_attempts,
+        default_project_id=composed.ir_projection.project_id,
+    )
+    return SimulationRun(composed=composed, result=result)
+
+
+def simulate_closed_loop(
+    bundle_dir: Path,
+    persona_artifact: Path,
+    inbox_dir: Path,
+    *,
+    client_a: ModelClient,
+    client_b: ModelClient,
+    timestamp: str,
+    experiment: str,
+    k: int = 3,
+    max_rounds: int = 12,
+    max_tokens: int = 150_000,
+    max_debate_rounds: int = 3,
+    max_attempts: int = 3,
+) -> ClosedLoopResult:
+    """Drive the closed loop: run A → persona verdict → §6 gate → run B (fresh store).
+
+    The persona judges run A's actual candidate (§4.2: the verdict id is
+    minted from the produced proposal's features) and authors the verdict
+    through the inbox primitive. The gate then decides: an approved verdict is
+    promoted into the bundle's wiki layer; anything else is refused fail-closed
+    and recorded. Run B recomposes from the bundle alone — the promoted layer
+    is the ONLY channel a judgement may cross runs on.
+    """
+    contracts = load_contracts(
+        data_source={"docs_dir": str(bundle_dir), "top_k": k},
+        termination={"max_rounds": max_rounds, "max_tokens": max_tokens},
+        feedback={"decision": "approved", "rationale": "startup shape check (§10)"},
+    )
+    # The navigated dir is the CONTRACT's, so the validated config is load-bearing.
+    bundle = Path(contracts.data_source.docs_dir)
+    run_a = _run_once(
+        client_a,
+        bundle,
+        contracts,
+        max_debate_rounds=max_debate_rounds,
+        max_attempts=max_attempts,
+    )
+    verdict = drop_persona_verdict(
+        inbox_dir,
+        persona_artifact,
+        CandidateFeatures.from_proposal(run_a.result.proposal),
+        description=_VERDICT_DESCRIPTION,
+    )
+    promoted_path: Path | None
+    promotion_refusal: str | None
+    try:
+        promoted_path = promote(
+            verdict,
+            bundle,
+            approved_by=_PERSONA_APPROVER,
+            experiment=experiment,
+            timestamp=timestamp,
+        )
+        promotion_refusal = None
+    except PromotionError as refusal:
+        promoted_path = None
+        promotion_refusal = str(refusal)
+    run_b = _run_once(
+        client_b,
+        bundle,
+        contracts,
+        max_debate_rounds=max_debate_rounds,
+        max_attempts=max_attempts,
+    )
+    return ClosedLoopResult(
+        run_a=run_a,
+        verdict=verdict,
+        promoted_path=promoted_path,
+        promotion_refusal=promotion_refusal,
+        run_b=run_b,
+    )

tests/test_simulation_loadbearing.py

@@ -0,0 +1,182 @@
+"""Closed-loop two-run simulation — LOAD-BEARING (K4/R-1; method-spec §11 «Closed loop»).
+
+The seam this file keeps alive: the persona's marker crosses from run A to
+run B's hypothesis prompt VIA the §6 promotion gate — and NEVER crosses
+without it. Both directions are bound: detach the promotion step from the
+driver → the marker is absent from run B → red; render the verdict layer
+into the read-context → the marker leaks around the gated fold → red (the
+verdict-exclusion invariant, re-proven at simulation level).
+"""
+
+from __future__ import annotations
+
+import json
+from pathlib import Path
+
+import pytest
+
+from _scripted import ScriptedClient, reply
+
+from portfolio_optimiser_claude.ir import load_validator_input
+from portfolio_optimiser_claude.okf import bundle_context
+from portfolio_optimiser_claude.persona import load_persona_example
+from portfolio_optimiser_claude.simulation import ClosedLoopResult, simulate_closed_loop
+
+SHARED_BUNDLE = Path(__file__).resolve().parents[1] / "shared" / "examples" / "bygg-energi-mikro"
+SHARED_ARTIFACT = (
+    Path(__file__).resolve().parents[1]
+    / "shared"
+    / "skills"
+    / "expert-reviewer"
+    / "references"
+    / "example-verdict.json"
+)
+# The shared artifact's own marker (realiseringsgrad=0.79) — distinct from the
+# bundle seed's 0.82, so its presence in run B proves the PROMOTION crossing,
+# not the pre-existing seeding path.
+MARKER = load_persona_example(SHARED_ARTIFACT).marker
+
+TIMESTAMP = "2026-07-17T04:00:00Z"
+EXPERIMENT = "K4-closed-loop-sim"
+
+
+@pytest.fixture()
+def bundle(tmp_path: Path) -> Path:
+    import shutil
+
+    target = tmp_path / "bundle"
+    shutil.copytree(SHARED_BUNDLE, target)
+    return target
+
+
+def _proposal_json(bundle: Path, *, measure_suffix: str = "") -> str:
+    proposal = load_validator_input(bundle).model_dump()
+    if measure_suffix:
+        proposal["measure"] = f"{proposal['measure']}{measure_suffix}"
+    return json.dumps(proposal)
+
+
+def _scripted_client(proposal_json: str) -> ScriptedClient:
+    # One full run: debate round (proposer + approving checker) + generation.
+    return ScriptedClient(
+        replies=[reply("debate reasoning"), reply("VERDICT: APPROVE"), reply(proposal_json)]
+    )
+
+
+def _simulate(
+    bundle: Path, artifact: Path, inbox: Path
+) -> tuple[ClosedLoopResult, ScriptedClient, ScriptedClient]:
+    # Run A proposes a DISTINCT candidate (measure variant): the persona's
+    # verdict then gets its own §4.2 id. Judging the bundle seed's own
+    # candidate would mint the seed's id, and the store's first-write-wins
+    # would silently shadow the promoted rationale behind the seed (the known
+    # C-F5 feature-keying limitation — deferred to C3.2, not worked around
+    # in src).
+    client_a = _scripted_client(_proposal_json(bundle, measure_suffix="-simulert-variant"))
+    client_b = _scripted_client(_proposal_json(bundle))
+    result = simulate_closed_loop(
+        bundle,
+        artifact,
+        inbox,
+        client_a=client_a,
+        client_b=client_b,
+        timestamp=TIMESTAMP,
+        experiment=EXPERIMENT,
+    )
+    return result, client_a, client_b
+
+
+def _rejecting_artifact(tmp_path: Path, marker: str) -> Path:
+    path = tmp_path / "rejecting-example-verdict.json"
+    path.write_text(
+        json.dumps(
+            {
+                "decision": "rejected",
+                "marker": marker,
+                "rationale": f"Avvist: driftstimene er overestimert ({marker}).",
+            }
+        ),
+        encoding="utf-8",
+    )
+    return path
+
+
+class TestMarkerCrossesViaPromotion:
+    """§11 «Closed loop», direction 1: WITH promotion the marker MUST cross."""
+
+    def test_the_persona_marker_reaches_run_b_via_promotion(
+        self, bundle: Path, tmp_path: Path
+    ) -> None:
+        result, _, client_b = _simulate(bundle, SHARED_ARTIFACT, tmp_path / "inbox")
+        # The marker did not exist anywhere before the persona judged run A.
+        assert MARKER not in result.run_a.composed.context
+        # The gate promoted the approved verdict into the bundle...
+        assert result.promoted_path is not None and result.promoted_path.is_file()
+        assert result.promotion_refusal is None
+        # ...and run B's FRESH store picked it up through seeding into the fold.
+        assert result.run_b.composed.seeded == result.run_a.composed.seeded + 1
+        assert MARKER in result.run_b.composed.context
+        # Detach proof 1 anchor: the marker reaches run B's actual prompts.
+        assert any(MARKER in prompt for prompt in client_b.prompts("proposer"))
+
+    def test_both_runs_complete_against_the_scripted_clients(
+        self, bundle: Path, tmp_path: Path
+    ) -> None:
+        result, _, _ = _simulate(bundle, SHARED_ARTIFACT, tmp_path / "inbox")
+        assert result.run_a.result.validator_decision == "validated"
+        assert result.run_b.result.validator_decision == "validated"
+        # The persona judged run A's candidate — the verdict carries its rationale.
+        assert MARKER in result.verdict.rationale
+
+
+class TestMarkerNeverCrossesWithoutPromotion:
+    """§11 «Closed loop», direction 2: WITHOUT promotion the marker NEVER crosses.
+
+    The rejected persona verdict IS authored into the inbox — if run B read the
+    inbox (or any channel other than the promoted wiki layer), the marker would
+    cross anyway and this control goes red.
+    """
+
+    def test_a_rejected_verdict_is_refused_and_the_marker_stays_out(
+        self, bundle: Path, tmp_path: Path
+    ) -> None:
+        marker = "realiseringsgrad=0.11"
+        artifact = _rejecting_artifact(tmp_path, marker)
+        result, _, client_b = _simulate(bundle, artifact, tmp_path / "inbox")
+        assert result.promoted_path is None
+        assert result.promotion_refusal is not None
+        assert "rejected" in result.promotion_refusal
+        assert marker not in result.run_b.composed.context
+        assert all(marker not in prompt for prompt in client_b.prompts("proposer"))
+
+    def test_a_refused_promotion_leaves_the_bundle_untouched(
+        self, bundle: Path, tmp_path: Path
+    ) -> None:
+        before = {p.name: p.read_bytes() for p in bundle.iterdir()}
+        _simulate(
+            bundle, _rejecting_artifact(tmp_path, "realiseringsgrad=0.11"), tmp_path / "inbox"
+        )
+        after = {p.name: p.read_bytes() for p in bundle.iterdir()}
+        assert after == before
+
+
+class TestGatedFoldIsTheOnlyChannel:
+    """LOAD-BEARING (§11): the marker crosses ONLY via the gated experience fold.
+
+    Detach proof 2 anchor: render the verdict layer into the read-context and
+    the marker leaks into the base context around the fold — red.
+    """
+
+    def test_the_verdict_layer_is_never_rendered_into_either_context(
+        self, bundle: Path, tmp_path: Path
+    ) -> None:
+        result, _, _ = _simulate(bundle, SHARED_ARTIFACT, tmp_path / "inbox")
+        assert MARKER in result.run_b.composed.context  # crossing happened...
+        # ...but NO verdict-typed concept is rendered into the read-context —
+        # a `## verdict` section in either run's context means the seed's (or
+        # a promoted file's) body leaked around the gated fold (§3 Step 1).
+        assert "## verdict" not in result.run_a.composed.context
+        assert "## verdict" not in result.run_b.composed.context
+        # Belt and braces: the promoted marker itself never surfaces via
+        # rendering (promoted files are frontmatter-only by design, §6).
+        assert MARKER not in bundle_context(bundle)