feat(fase5): add the long/async verdict file inbox (Steg 7 resumable feedback)

The short loop captured the expert verdict inline into an in-memory store, so a
verdict arriving days/weeks later in a separate run could not influence any future
hypothesis (målbilde §5 row 7). Steg 7 adds the long timescale: run_project gains an
opt-in verdict_dir async inbox that load_verdicts_from_dir -> store.add MERGES into the
store BEFORE the Step-1 ExpeL fold, so a verdict dropped after an earlier run reaches a
separate, later run's hypothesis — fully resumable across runs separated in time.

- verdicts.py: verdict_to_dict / verdict_from_dict (id read verbatim, never re-minted),
  write_verdict (public authoring primitive, NOT wired into run_project — system reads
  the folder, expert/persona writes it, §3 role split), tolerant load_verdicts_from_dir
  (missing/foreign/half-written files skipped, not raised — RAW layer per §10 R2),
  VerdictStore.from_dir.
- run.py: verdict_dir kwarg; ingest-merge block after load_contracts (merge not replace
  keeps run_portfolio's cross-project threading; store.add idempotent on content-hash id;
  no change to the fold). CLI --bundle-dir/--verdict-dir thread the long loop to the
  console entry. No auto-persist of the run's own captured verdict (outbox/Steg 8).
- Load-bearing PAIR (test_step7_async_loop_loadbearing.py): a verdict dropped after run A
  must reach run B's prompt (run B uses a FRESH store -> the transfer is the file loop,
  not in-memory carryover); empty-inbox control proves causality. Marker = a realization
  value absent from the bundle (not the seed's 0.82). Proven RED on ingest detach.

Suite 138 -> 140 passed, 4 skipped; mypy + ruff check clean. Målbilde treated as frozen
(no §3/§5/§7 edit). Step 8 (gated wiki promotion) remains.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MHR8iKxJRxDiDfNw8HZmWE

CLAUDE.md

@@ -56,6 +56,19 @@ Python ≥3.10. MAF (`agent-framework-core` 1.9.0). Pakkehåndtering: `uv`. To b
   her) — så koden påstår ikke mer enn den gjør. Load-bearing:
   `tests/test_step5_refine_loadbearing.py` blir rød når reason-injeksjonen detaches (utfallet
   flipper aldri + reason-verbatim-asserten faller); kontrollen beviser at løkka forblir bundet.
+- **Lang/async fil-løkke (Steg 7, målbilde §3/§7):** `run_project(verdict_dir=...)` er den lange
+  tilbakemeldings-tidsskalaen — en ekspert/persona dropper en verdict-fil (vanlig JSON, RAW-laget
+  per §10 R2) i en inbox-mappe ETTER en kjøring, og en separat, senere kjøring `load_verdicts_from_dir`
+  → `store.add` **merger** den inn FØR Steg-1-folden (ingen endring i folden), så dommen når neste
+  hypotese. **Rolledeling (§3, ufravikelig):** systemet LESER mappa; eksperten/personaen SKRIVER den
+  — `run_project` persisterer ALDRI sin egen fangede dom tilbake (det er outbox/Steg 8). Merge, aldri
+  erstatt (`run_portfolio`-tråding intakt); tolerant last (manglende mappe / fremmede / halvskrevne
+  filer hoppes over, ikke raises — RAW-lag, kontrast `okf.load_ir_projection`s fail-fast); `id` leses
+  verbatim, re-mintes aldri. `write_verdict` er den offentlige authoring-primitiven (persona/test +
+  framtidig Steg 8), men wires IKKE inn i `run_project`. Load-bearing:
+  `tests/test_step7_async_loop_loadbearing.py` — en dom droppet etter Run A MÅ nå Run B's prompt
+  (Run B bruker FERSK store → overføringen er fil-løkka, ikke in-memory-carryover); tom-inbox-kontroll
+  beviser kausalitet. Markør = realiseringsverdi som finnes ingen steder i bundelen (ikke frøets 0.82).
 - **Kostnadsdisiplin:** utvikle primært på lokal profil (gratis); Foundry/Azure (privat tenant finnes) kun til målrettet, minimal verifisering; billigste modeller + små syntetiske data + harde token-tak. Ingen tunge test-kjøringer.
 - **STATE.md er local-only** (gitignored). Voyage session-state er efemert; STATE.md er kanonisk kontinuitet.
 - Prosess: Voyage-plugin (`/trekbrief → /trekplan → /trekexecute → /trekreview`) per større fase.

README.md

@@ -2,7 +2,7 @@
 
 Generic, open framework on **Microsoft Agent Framework (MAF)** for finding cost-savings / efficiency proposals *within* each project of a portfolio of independent projects. Multiple agents collaborate to generate candidate proposals; a mandatory deterministic validator (solver + Monte Carlo) decides the numbers; domain experts review via human-in-the-loop, and the system learns from their verdicts.
 
-> **Status:** Early development. The deterministic backbone is solid; the agentic learning loop is being wired one load-bearing seam at a time — **Step 1 (OKF-navigated context → hypothesis) is wired**, with the verdict layer kept out of the read-context (see below). Not yet end-to-end usable.
+> **Status:** Early development. The deterministic backbone is solid; the agentic learning loop is being wired one load-bearing seam at a time — **Steps 1, 3/4, 5 and 7 are wired** (OKF-navigated context, checker gate, informed refinement, and the async file feedback loop; see below). Step 8 (gated wiki promotion) remains. Not yet end-to-end usable.
 
 > **Disclaimer — technical framework only.** This project is a *technical framework*. Organizations that deploy it are themselves responsible for ensuring a valid processing purpose and for any required assessments (DPIA, risk/ROS, security reviews, etc.). The framework ships technical affordances (local-only mode, provenance/audit logging, no silent data egress) to *enable* compliant use, but makes no compliance guarantees.
 
@@ -20,7 +20,8 @@ The mandatory deterministic backbone (validator + budget meter + provenance) is
   - Two load-bearing tests fail when the seam is detached: the realization signal must reach the prompt via the fold (`tests/test_step1_expel_loadbearing.py`), and must never leak via context (`tests/test_okf.py::test_bundle_context_excludes_verdict_layer`).
 - **Steps 3/4 — checker gates the reasoning (wired).** Two falsifiers now act on the same candidate: the deterministic validator gates the **numbers** (blocking, as before), and the maker-checker's checker gates the **reasoning** (target picture §2/§6). The checker ends its turn with a `VERDICT: APPROVE` / `VERDICT: REJECT — <reason>` line; an explicit reject blocks an otherwise-validated proposal (`run_project` surfaces both debate participants via `output_from=agents` and overrides the outcome to a checker-sourced `Rejection`). The gate is opt-in-reject (fail-open on a missing marker), and `provenance.validator_decision` stays honest — it reflects the validator only, never the checker. Load-bearing: `tests/test_checker_gate_loadbearing.py` goes red on *either* detach (revert `output_from`, or drop the override).
 - **Step 5 — informed refinement (wired).** The proposer's bounded retry is no longer blind: the validator's *previous* `Rejection.reason` is fed into the next attempt's prompt (`generate.py` `generate_via_llm` → `_build_messages(prior_rejection=...)`), so the model corrects against the falsification instead of re-answering identically (target picture §5/§7). It carries only the most-recent reason (never an accumulated history, never the rejected proposal JSON) and runs under the *existing* `max_attempts` + token-meter cap — no new loop, so "improve until good enough" without a ceiling stays impossible. The only per-attempt falsifier here is the validator; seeding generation with the checker's critique is a run-level, separately-scoped concern and is deliberately not done here. Load-bearing: `tests/test_step5_refine_loadbearing.py` goes red when the reason is detached from the prompt (the outcome never flips *and* the verbatim-reason assertion fails), with a bounded control proving the loop still stops at `max_attempts`.
-- **Steps 7–8** (async file feedback, gated wiki promotion) — not yet wired.
+- **Step 7 — async file feedback loop (wired).** `run_project(..., verdict_dir=...)` adds the long feedback timescale (target picture §3/§7): an expert/persona drops a verdict file (plain JSON — the raw output layer, §10 R2) into an inbox folder *after* a run, and a separate, later run ingests it — merged into the store *before* the Step-1 fold — so a verdict that landed out of band reaches the next hypothesis. The loop is fully resumable across runs separated in time; no live session is assumed. The system *reads* the folder, the expert/persona *writes* it (§3 role split), so `run_project` deliberately does **not** persist its own captured verdict back (that is the outbox / Step-8 concern). Ingestion is tolerant (a missing folder, foreign or half-written files are skipped, not raised) and **merges** (never replaces), preserving `run_portfolio`'s cross-project store. Reachable from the CLI via `--bundle-dir --verdict-dir`. Load-bearing: `tests/test_step7_async_loop_loadbearing.py` — a verdict dropped after run A must reach run B's prompt (run B uses a *fresh* store, so the transfer is the file loop, not in-memory carryover), with an empty-inbox control proving causality.
+- **Step 8** (gated wiki promotion) — not yet wired.
 
 ## Docs
 

src/portfolio_optimiser/run.py

@@ -53,6 +53,7 @@ from portfolio_optimiser.verdicts import (
     VerdictStore,
     bundle_candidate_features,
     capture_verdict,
+    load_verdicts_from_dir,
 )
 from portfolio_optimiser.workflow import fresh_workflow
 
@@ -198,6 +199,7 @@ async def run_project(
     verdict_input: dict[str, str],
     bundle_dir: str | None = None,
     store: VerdictStore | None = None,
+    verdict_dir: str | None = None,
     client_factory: Callable[[str], BaseChatClient] | None = None,
     max_rounds: int = 3,
     max_tokens: int = 100_000,
@@ -210,9 +212,13 @@ async def run_project(
     (defaults to the real backend). ``verdict_input`` carries the expert decision/rationale
     (Layer-2). ``bundle_dir`` (Fase 2a) makes the run OKF-bundle-driven: the project is derived
     from the bundle and, before generation, the candidate's prior verdicts in ``store`` are folded
-    into the hypothesis prompt (Step-1 ExpeL wiring, målbilde §5/§7). Raises
-    ``pydantic.ValidationError`` on a bad contract and ``BudgetExceeded`` when the token/round cap
-    is crossed."""
+    into the hypothesis prompt (Step-1 ExpeL wiring, målbilde §5/§7). ``verdict_dir`` (Fase 5,
+    Steg 7, målbilde §3/§7) is the async file inbox: a folder of expert/persona-authored verdict
+    files (plain JSON, R2 raw layer) MERGED into the store BEFORE the Step-1 fold, so a verdict
+    dropped after an earlier run is consumed by this separate, later run — the long feedback loop,
+    fully resumable across runs separated in time. The system READS this folder; it does not write
+    to it (the expert/persona writes, målbilde §3). Raises ``pydantic.ValidationError`` on a bad
+    contract and ``BudgetExceeded`` when the token/round cap is crossed."""
     # 1. Fail-fast: validate ALL contracts (incl. the verdict-feedback shape) before any client.
     load_contracts(
         {"docs_dir": docs_dir, "top_k": top_k},
@@ -220,6 +226,15 @@ async def run_project(
         verdict_input,
     )
 
+    # 1b. Long loop (Steg 7): ingest the async verdict inbox INTO the store before the Step-1 fold.
+    # Merge (not replace) into the passed store so run_portfolio's cross-project threading stays
+    # intact; store.add is idempotent on the content-hash id. A verdict that landed after an earlier
+    # run thus reaches THIS run's hypothesis via the existing fold below — no change to the fold.
+    if verdict_dir is not None:
+        store = store if store is not None else VerdictStore(verdicts=[])
+        for dropped in load_verdicts_from_dir(verdict_dir):
+            store.add(dropped)
+
     # 2-3. Project + agent read-context + first-class citations. A bundle run derives ALL THREE from
     # the navigated OKF bundle via progressive disclosure (verdict layer EXCLUDED — målbilde §2/§4),
     # NOT keyword chunk-stuffing; the road path keeps the chunk-retrieval data source. ``debate_tools``
@@ -406,6 +421,15 @@ def main(argv: list[str] | None = None) -> int:
     parser.add_argument("project_id")
     parser.add_argument("--profile", default="local")
     parser.add_argument("--docs-dir", required=True)
+    parser.add_argument(
+        "--bundle-dir", default=None, help="OKF bundle dir (enables the Step-1 fold)"
+    )
+    parser.add_argument(
+        "--verdict-dir",
+        default=None,
+        help="async verdict inbox: a folder of dropped expert verdicts, ingested before generation "
+        "(the long loop — a verdict that landed after an earlier run is consumed by this run)",
+    )
     parser.add_argument("--decision", default="approved", choices=["approved", "rejected"])
     parser.add_argument("--rationale", default="reviewed by expert")
     args = parser.parse_args(argv)
@@ -415,6 +439,8 @@ def main(argv: list[str] | None = None) -> int:
             args.project_id,
             args.profile,
             docs_dir=args.docs_dir,
+            bundle_dir=args.bundle_dir,
+            verdict_dir=args.verdict_dir,
             verdict_input={"decision": args.decision, "rationale": args.rationale},
         )
     )

src/portfolio_optimiser/verdicts.py

@@ -21,6 +21,7 @@ from __future__ import annotations
 import hashlib
 import json
 from dataclasses import dataclass
+from pathlib import Path
 from typing import Any
 
 from agent_framework import ContextProvider, SessionContext
@@ -104,12 +105,111 @@ def capture_verdict(features: ProposalFeatures, decision: str, rationale: str) -
     )
 
 
+# --- Async file inbox (Fase 5, Steg 7): plain-JSON verdict serialization + tolerant folder load ---
+# The raw output layer is a plain JSON store (R2), NOT OKF — OKF is reserved for the promoted
+# (Step-8) layer. One verdict per file lets an expert/persona drop verdicts into a folder out of
+# band, and a separate later run pick them up (målbilde §3 long loop).
+
+_REQUIRED_VERDICT_KEYS = {"id", "decision", "rationale", "proposal_features"}
+
+
+def verdict_to_dict(verdict: Verdict) -> dict[str, Any]:
+    """Serialize a ``Verdict`` to a JSON-ready dict. ``affected_codes`` (a frozenset) is emitted as
+    a SORTED list — matching ``_mint_id``'s canonical form — so the round-trip is lossless and the
+    on-disk form is deterministic."""
+    f = verdict.proposal_features
+    return {
+        "id": verdict.id,
+        "decision": verdict.decision,
+        "rationale": verdict.rationale,
+        "proposal_features": {
+            "affected_codes": sorted(f.affected_codes),
+            "measure_type": f.measure_type,
+            "claimed_saving_nok": f.claimed_saving_nok,
+            "description": f.description,
+        },
+    }
+
+
+def verdict_from_dict(data: dict[str, Any]) -> Verdict:
+    """Reconstruct a ``Verdict`` from its serialized dict. The ``id`` is read VERBATIM — never
+    re-minted — so a verdict authored elsewhere keeps its identity and the int/float magnitudes in
+    its features survive the JSON round-trip untouched (re-minting would re-hash and could diverge,
+    since ``_mint_id`` hashes the raw value, e.g. ``30000`` vs ``30000.0``)."""
+    pf = data["proposal_features"]
+    return Verdict(
+        id=data["id"],
+        proposal_features=ProposalFeatures(
+            affected_codes=frozenset(pf["affected_codes"]),
+            measure_type=pf["measure_type"],
+            claimed_saving_nok=pf["claimed_saving_nok"],
+            description=pf.get("description", ""),
+        ),
+        decision=data["decision"],
+        rationale=data["rationale"],
+    )
+
+
+def write_verdict(directory: str, verdict: Verdict) -> Path:
+    """Write one verdict as ``{id}.json`` into ``directory`` (created if needed) and return the path.
+
+    This is the public authoring primitive the expert/persona (sim) or human (prod) uses to drop a
+    verdict into the async inbox via the SAME folder interface — and the writer a future Step-8
+    promotion would reuse. It is deliberately NOT called inside ``run_project``: the system READS the
+    inbox, it does not write to it (målbilde §3 role split).
+
+    Limitation: ``id`` is a content hash of the FEATURES only (``_mint_id``), so two verdicts with
+    identical features but different decisions share a filename — last write wins on disk (mirroring
+    ``VerdictStore.add``'s first-wins in memory). Acceptable for the raw MVP layer."""
+    path = Path(directory)
+    path.mkdir(parents=True, exist_ok=True)
+    target = path / f"{verdict.id}.json"
+    target.write_text(
+        json.dumps(verdict_to_dict(verdict), sort_keys=True, indent=2), encoding="utf-8"
+    )
+    return target
+
+
+def load_verdicts_from_dir(directory: str) -> list[Verdict]:
+    """Load every well-formed verdict JSON file from an async inbox folder, TOLERANTLY (OKF §4
+    spirit): a missing folder yields ``[]``; files that are not ``.json``, fail to parse, or lack a
+    required key are SKIPPED, never raised. The folder is written out of band by an external party,
+    so half-written or foreign files (e.g. a bundle's ``golden.json``) are realistic — this is the
+    raw layer, not required input (contrast ``okf.load_ir_projection``'s fail-fast). Order is
+    deterministic (sorted by filename)."""
+    path = Path(directory)
+    if not path.is_dir():
+        return []
+    verdicts: list[Verdict] = []
+    for file in sorted(path.glob("*.json")):
+        try:
+            data = json.loads(file.read_text(encoding="utf-8"))
+        except (OSError, json.JSONDecodeError):
+            continue
+        if not isinstance(data, dict) or not _REQUIRED_VERDICT_KEYS <= data.keys():
+            continue
+        try:
+            verdicts.append(verdict_from_dict(data))
+        except (KeyError, TypeError):
+            continue
+    return verdicts
+
+
 @dataclass
 class VerdictStore:
-    """Minimal in-memory store of historical verdicts (MVP — no file persistence)."""
+    """Minimal store of historical verdicts. In-memory by default; ``from_dir`` / the
+    ``load_verdicts_from_dir`` merge primitive back it with the async file inbox (Steg 7)."""
 
     verdicts: list[Verdict]
 
+    @classmethod
+    def from_dir(cls, directory: str) -> VerdictStore:
+        """Build a store from an async inbox folder (convenience constructor). ``run_project`` uses
+        the merge primitive (``add`` per loaded verdict) instead, to preserve a passed store's
+        existing verdicts (the cross-project threading in ``run_portfolio``); this is for callers
+        that want a fresh store straight from a folder."""
+        return cls(verdicts=load_verdicts_from_dir(directory))
+
     def retrieve(self, query: ProposalFeatures, k: int) -> list[Verdict]:
         """Return the top-``k`` verdicts by structural similarity. Deterministic: ties break
         by verdict id, so ordering is stable across runs."""

tests/test_step7_async_loop_loadbearing.py

@@ -0,0 +1,151 @@
+"""Step-7 load-bearing seam (målbilde §3 / §5 row 7 / §7): the LONG/async feedback loop — a
+verdict file that lands in a folder AFTER one run MUST be consumed by a separate, later run.
+
+The gap (fot-i-bakken, verified in code): the short loop captured the expert verdict inline
+(``verdict_input``) into an IN-MEMORY ``VerdictStore`` — so a verdict that arrives days/weeks
+later, in a separate run, could not influence any future hypothesis. Fase 5 adds a file inbox:
+``run_project(verdict_dir=...)`` ingests expert/persona-authored verdict files (plain JSON, R2 raw
+layer) and MERGES them into the store BEFORE the Step-1 ExpeL fold, so the dropped verdict reaches
+the next run's hypothesis prompt through the existing fold.
+
+These two tests are a load-bearing PAIR (per the Fase-2 green-but-dead trap):
+- the positive test keys on a realization marker that occurs NOWHERE in the bundle (0.57, not the
+  bundle seed's 0.82), carried in the dropped verdict's rationale — proving the learning PAYLOAD
+  round-tripped end to end (write -> file -> ingest -> retrieve -> fold -> prompt), not merely that
+  an object was retrieved. It goes RED the moment the ingest is detached.
+- the control proves causality: the SAME run with an EMPTY inbox carries no marker — so the signal
+  is caused by the file loop, not incidental to the bundle. ``bundle_context`` excludes
+  ``type: verdict`` files (okf.py), so the marker can ONLY arrive via ingest -> store -> fold.
+
+Critical invariant: run B uses a FRESH store (never run A's object) — otherwise the transfer would
+be in-memory carryover, not the file loop.
+"""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+from portfolio_optimiser.run import run_project
+from portfolio_optimiser.validator import ValidatedProposal
+from portfolio_optimiser.verdicts import (
+    Verdict,
+    VerdictStore,
+    bundle_candidate_features,
+    write_verdict,
+)
+
+BUNDLE_DIR = Path(__file__).resolve().parents[1] / "shared" / "examples" / "bygg-energi-mikro"
+
+# A VALID BYGG-KONTOR-NORD proposal (same as the Step-1 test): P90 = 0.30 x 300000 = 90000 >=
+# claimed 30000 -> validates on the first attempt.
+_ENERGY_REPLY = (
+    '{"measure":"LED-retrofit av kontorbelysning","affected_items":'
+    '[{"code":"ENERGI-TOTAL-EL","quantity":300000,"unit_cost":1.0}],"claimed_saving_nok":30000}'
+)
+_VERDICT_INPUT = {"decision": "approved", "rationale": "expert reviewed (sim)"}
+
+# Two load-bearing markers carried by the DROPPED verdict, by design:
+#  - a realization value absent from EVERY bundle file (the bundle seed is 0.82). It can reach a
+#    prompt only via the ingest->fold path -> isolates the file loop AND would catch a regression
+#    where bundle navigation starts leaking verdict files (a 0.82 marker would pass for the wrong
+#    reason and mask that).
+#  - a sentinel verdict id (cannot appear in the bundle's raw text) for object identity.
+_MARKER = "realiseringsgrad=0.57"
+_DROP_ID = "STEG7-DROP"
+
+
+def _generation_prompts(sink: list[str]) -> list[str]:
+    """The generation-call prompts (``generate._build_messages`` embeds 'SavingsProposal'),
+    isolated from the debate-round prompts also captured in the sink."""
+    return [p for p in sink if "SavingsProposal" in p]
+
+
+def _drop_expert_verdict(inbox: Path) -> None:
+    """Simulate the long loop: an expert/persona drops a verdict file into the inbox AFTER a run.
+    Keyed on the bundle candidate's features (the same key the Step-1 fold queries on), carrying the
+    realization marker in its rationale. ``write_verdict`` is the public authoring primitive a
+    persona (sim) or human (prod) uses against the SAME folder interface."""
+    verdict = Verdict(
+        id=_DROP_ID,
+        proposal_features=bundle_candidate_features(str(BUNDLE_DIR)),
+        decision="approved",
+        rationale=f"LED-retrofit godkjent med realiseringskorreksjon ({_MARKER})",
+    )
+    write_verdict(str(inbox), verdict)
+
+
+async def test_dropped_verdict_reaches_later_run(make_recording_client_factory, tmp_path) -> None:
+    """LOAD-BEARING: a verdict dropped into the inbox AFTER run A reaches run B's hypothesis prompt.
+    Goes RED if ``run_project`` stops ingesting ``verdict_dir`` before the Step-1 fold (the marker
+    never arrives -> no fold -> the assertion fails)."""
+    inbox = tmp_path / "verdict-inbox"
+    inbox.mkdir()
+
+    # Run A: an earlier run with an EMPTY store and an EMPTY inbox -> nothing folds in, and (no
+    # auto-persist) run A does NOT write its own captured verdict back to the inbox.
+    factory_a, recorded_a = make_recording_client_factory(_ENERGY_REPLY)
+    await run_project(
+        "BYGG-KONTOR-NORD",
+        "local",
+        docs_dir=str(BUNDLE_DIR),
+        bundle_dir=str(BUNDLE_DIR),
+        verdict_input=_VERDICT_INPUT,
+        store=VerdictStore(verdicts=[]),
+        verdict_dir=str(inbox),
+        client_factory=factory_a,
+    )
+    assert all(_MARKER not in p for p in recorded_a), "precondition: run A carries no marker"
+
+    # Days/weeks later, out of band: the expert drops a verdict into the inbox folder.
+    _drop_expert_verdict(inbox)
+
+    # Run B: a SEPARATE, later run with a FRESH store (never run A's object) -> the dropped verdict
+    # must reach the hypothesis prompt purely via the file loop.
+    factory_b, recorded_b = make_recording_client_factory(_ENERGY_REPLY)
+    result = await run_project(
+        "BYGG-KONTOR-NORD",
+        "local",
+        docs_dir=str(BUNDLE_DIR),
+        bundle_dir=str(BUNDLE_DIR),
+        verdict_input=_VERDICT_INPUT,
+        store=VerdictStore(verdicts=[]),
+        verdict_dir=str(inbox),
+        client_factory=factory_b,
+    )
+
+    assert isinstance(result.outcome, ValidatedProposal)
+    gen_prompts = _generation_prompts(recorded_b)
+    assert gen_prompts, "the generation call must have happened"
+    assert any(_MARKER in p for p in gen_prompts), (
+        "the dropped verdict's realization signal did not reach run B's hypothesis prompt — the "
+        "async file inbox is not ingested before the Step-1 fold"
+    )
+    assert any(_DROP_ID in p for p in gen_prompts), (
+        "the dropped verdict's id did not reach the prompt"
+    )
+
+
+async def test_empty_inbox_leaves_no_signal(make_recording_client_factory, tmp_path) -> None:
+    """CAUSALITY CONTROL: the same run with an EMPTY inbox carries no marker into any prompt —
+    proving the positive test's signal is caused by the file loop, not incidental to the bundle
+    (``bundle_context`` excludes ``type: verdict`` files, so the marker has no other path)."""
+    inbox = tmp_path / "verdict-inbox"
+    inbox.mkdir()
+
+    factory, recorded = make_recording_client_factory(_ENERGY_REPLY)
+    await run_project(
+        "BYGG-KONTOR-NORD",
+        "local",
+        docs_dir=str(BUNDLE_DIR),
+        bundle_dir=str(BUNDLE_DIR),
+        verdict_input=_VERDICT_INPUT,
+        store=VerdictStore(verdicts=[]),
+        verdict_dir=str(inbox),
+        client_factory=factory,
+    )
+
+    assert all(_MARKER not in p for p in recorded), (
+        "the realization marker reached a prompt with an EMPTY inbox — the positive test is not "
+        "load-bearing (the signal is incidental, not caused by the ingest seam)"
+    )
+    assert all(_DROP_ID not in p for p in recorded), "the sentinel id leaked without a dropped file"