fix(run): S10 del 2 — post-mortem: stopp-artefakt, SDK-isolasjon, raw-JSON-direktiv

Transkript-analyse av den stoppede live-kjøringen (10 kall, 162 250 tokens,
$0.331506): konfig-lekkasjen (setting_sources=None laster ALLE filsystem-
settings) injiserte operatørens Claude-konfig i hvert kall — ~10-15k uncachede
tokens, en påtvunget bekreftelses-preamble som gjorde ren-JSON-svar umulige,
og en checker kapret av lekkede instrukser (debatt konvergerte aldri).

- persist_stop_artifacts: stopp-event verbatim + usage/kost persisteres ALLTID
  ved BudgetExceeded (delt usage-shape med fullført-run-stien)
- build_call_options: setting_sources=[] (SDK isolation mode, verifisert mot
  installert 0.2.110-kilde), system_prompt=None → tom system-prompt; detach-
  bevis via monkeypatchet query
- _generation_prompt: krever ONLY the raw JSON object (fence-innpakning ga
  4 fullpris parse-retries)

187/187 uten nøkkel · ruff + mypy --strict rene · tre detach-bevis RØDE → grønn

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QdSfQdND84oeq2mbjueLTS

src/portfolio_optimiser_claude/sdk_client.py

@@ -1,11 +1,13 @@
-"""The Claude Agent SDK model client — RUN-PATH ONLY (honesty rule §1, §11).
+"""The Claude Agent SDK model client (honesty rule §1, §11).
 
-This module is the ONE place the programme touches a real model (S10, D6). It
-is never imported by the test suite — the offline suite proves the loop with
-the scripted stand-in, and this client slots into the same ``ModelClient``
-protocol seam. Each ``complete()`` is one bounded ``query()`` call: no tools,
-one turn, a first-class USD cap (``ClaudeAgentOptions.max_budget_usd``) ON TOP
-of the §8 token/round meter that the loop already charges.
+This module is the ONE place the programme touches a real model (S10, D6).
+The network path (``query()``) is run-path-only; the offline suite proves the
+loop with the scripted stand-in, imports this module WITHOUT touching a key
+or the network, and pins the call options (``tests/test_sdk_isolation.py``).
+Each ``complete()`` is one bounded, ISOLATED ``query()`` call: no tools, one
+turn, a first-class USD cap (``ClaudeAgentOptions.max_budget_usd``) ON TOP of
+the §8 token/round meter that the loop already charges, and NO filesystem
+settings (``setting_sources=[]``).
 
 Verified against claude-agent-sdk 0.2.110: ``query()`` yields
 ``AssistantMessage`` (text blocks + real model id) and a closing
@@ -39,6 +41,28 @@ _USAGE_TOKEN_FIELDS = (
 )
 
 
+def build_call_options(model_id: str, *, max_budget_usd: float) -> ClaudeAgentOptions:
+    """One bounded, ISOLATED completion call (§8 + S10 post-mortem).
+
+    ``setting_sources=[]`` is the SDK's documented isolation mode (verified
+    against 0.2.110): the spawned CLI loads NO filesystem settings — no
+    session hooks, no CLAUDE.md, no operator instructions. The default
+    (``None``) loads ALL sources: in the S10 live run that injected the
+    operator's config into every call (~10-15k uncached tokens each) and
+    mandated a confirmation preamble that made pure-JSON replies impossible.
+    ``system_prompt=None`` serializes to an EMPTY system prompt, not the
+    Claude Code preset.
+    """
+    return ClaudeAgentOptions(
+        model=model_id,
+        max_turns=1,  # a single completion — the agentic loop lives in loop.py, not here
+        tools=[],  # pure text completion: no tool surface, no silent egress
+        max_budget_usd=max_budget_usd,
+        setting_sources=[],
+        system_prompt=None,
+    )
+
+
 def _total_tokens(usage: dict[str, Any] | None) -> int | None:
     """Sum the provider-reported token fields; no usage stays ``None`` (§8)."""
     if usage is None:
@@ -77,12 +101,7 @@ class SdkModelClient:
         return asyncio.run(self._complete_async(prompt, model_id))
 
     async def _complete_async(self, prompt: str, model_id: str) -> ModelReply:
-        options = ClaudeAgentOptions(
-            model=model_id,
-            max_turns=1,  # a single completion — the agentic loop lives in loop.py, not here
-            tools=[],  # pure text completion: no tool surface, no silent egress
-            max_budget_usd=self._max_budget_usd_per_call,
-        )
+        options = build_call_options(model_id, max_budget_usd=self._max_budget_usd_per_call)
         text_parts: list[str] = []
         reply_model: str | None = None
         usage_tokens: int | None = None