fix(eod): defensive JSON extraction for LLM rationale synthesis (L1248/L2669)

[cipher813]

Summary

`_synthesize_rationales` parsed Haiku's response via bare `json.loads(response.content[0].text)` which silently fell back to the template path whenever Haiku returned anything but a clean top-level JSON object. EOD logs have shown the recurrence (`Expecting value: line 1 column 1 (char 0) — using template fallback`) across multiple runs over the last month — promoted P3 → P2 in the 2026-05-20 curation pass per the entry's own "if recurrence confirmed" trigger.

What changed

New `_extract_json_object` defensively handles three observed LLM anomalies before raising:

1. Markdown code fences (````json\n{...}\n```` or ````\n{...}\n````)
2. Conversational preamble (`"Sure, here's the JSON:\n{...}"`)
3. Trailing text after the closing brace

Clean JSON still goes through the fast path. On total parse failure, log a bounded sample (`raw[:400]`) so the next recurrence surfaces a concrete failure mode instead of the opaque "char 0" message, and re-raise into the existing template-fallback handler so the EOD email still ships.

Tests

9 new tests in `tests/test_eod_reconcile_logic.py::TestExtractJsonObject`:

• clean JSON / clean narratives shape
• markdown fences with + without language tag
• preamble alone / trailing text alone / preamble + fence + trailing combined
• garbage input raises / empty string raises

Suite: 905 → 913, all green.

Closes

L1248 + L2669 P3 → P2 promotions (canonical entry + cross-ref from the Executor side) — same fix lands both checkboxes.

Test plan

• Targeted: pytest tests/test_eod_reconcile_logic.py → 36 passed (was 27 before this PR's 9 new tests)
• Full ae suite: 913 passed
• Next EOD reconcile run on ae-trading exercises the new path; on a Haiku-returns-fenced-JSON day, expect "LLM rationale synthesis failed" log lines to disappear and per-position narratives to land in the EOD email body again.

🤖 Generated with Claude Code

[cipher813]

Pushed SOTA upgrade as follow-up commit 32f3597 per the no-shortcuts directive.

What changed from the initial heuristic:

• Removed _extract_json_object string-fence-stripping helper entirely.
• New _Narrative + _RationalesResponse Pydantic models define the contract; JSON Schema derived via model_json_schema() and registered as _RATIONALES_TOOL.
• _synthesize_rationales calls Haiku with tools=[_RATIONALES_TOOL] + tool_choice={"type": "tool", "name": "emit_rationales"} — forces a typed tool_use response. SDK validates the input shape against the schema before it lands here; Pydantic re-validates at the boundary.
• Parse failures from text anomalies (fences / preamble / trailing text) are now structurally impossible, not handled-after-the-fact.
• Pattern matches alpha-engine-research's with_structured_output discipline (agents/macro_agent.py:380) without adding langchain — uses the raw Anthropic SDK primitive directly.

Tests rewritten: deleted 9 string-extraction tests, added 11 contract tests (Pydantic boundary correctness + tool-use happy/sad paths with SDK-kwargs assertions). Suite 913 → 915. Both commits remain in the PR for review-trail clarity.

Anyone reviewing: feel free to squash on merge — the heuristic commit is obsolete.