fix(compiler): unwrap JSON-array LLM responses so concept/entity pages aren't silently dropped

[cnndabbler]

Problem

The four page generators in compiler.py (_gen_create, _gen_update, _gen_entity_create, _gen_entity_update) do:

parsed = _parse_json(raw)
brief = parsed.get("brief", "")

When the model returns a JSON array instead of an object, _parse_json returns a list, and parsed.get(...) raises AttributeError: 'list' object has no attribute 'get'. That error is not caught by the surrounding except (json.JSONDecodeError, ValueError), so the whole coroutine dies and the page is silently dropped:

openkb.agent.compiler WARNING: Entity generation failed: 'list' object has no attribute 'get'
[WARN] N entity(ies) planned but only M written ... (AttributeError).

This is the 'list' object has no attribute 'get' failure reported in #71. It still reproduces on current main (observed across a 165-doc ingest with several models — deepseek, qwen — affecting ~20% of docs, each losing one or more pages).

Fix

Add _as_obj() next to _parse_json(): if the parsed value is a list, unwrap the first dict element (the common "model wrapped the object in a one-element array" case); otherwise raise ValueError so the callers' existing raw-body fallback applies. Wrap all four call sites as parsed = _as_obj(_parse_json(raw)).

No behavior change for the normal object case.

Tests

Adds TestAsObj covering dict passthrough, single-element unwrap, first-dict-in-list, and the list-without-dict / empty-list ValueError paths. Full suite green locally.

Relates to #71.

[KylinMountain]

Nice, clean fix and good test coverage — _as_obj is correct and the four call sites correctly fall back via the existing except (json.JSONDecodeError, ValueError).

One thing before merge: the fix is incomplete — the same .get-on-a-list crash still lives in the summary step of compile_short_doc, which this PR does not wrap:

OpenKB/openkb/agent/compiler.py

Lines 1992 to 1999 in 1ad8189

	try:
	summary_parsed = _parse_json(summary_raw)
	doc_brief = summary_parsed.get("brief", "")
	summary = summary_parsed.get("content", summary_raw)
	except (json.JSONDecodeError, ValueError):
	doc_brief = ""
	summary = summary_raw

summary_parsed = _parse_json(summary_raw)
doc_brief = summary_parsed.get("brief", "")          # AttributeError when the model returns a JSON array
summary = summary_parsed.get("content", summary_raw)
except (json.JSONDecodeError, ValueError):           # does not catch AttributeError

It uses the same _parse_json, the same models, and the same _JSON_RESPONSE_FORMAT, so the array-wrapping case from #71 is just as likely here. And the blast radius is larger: the uncaught AttributeError propagates out of compile_short_doc, killing the whole short-doc compile rather than dropping a single page. Suggest applying the same _as_obj(_parse_json(summary_raw)) here.

(For reference, the plan-parse site is already safe — it guards with an isinstance(parsed, (list, dict)) check before calling .get.)

Minor / optional:

• _as_obj silently discards all but the first dict when the list has multiple objects. Fine for the one-element wrap case, just noting it.
• Since five sites now want "parse to a dict", a small _parse_json_obj(raw) wrapper would dedupe the call sites and naturally cover the summary path so it cant be missed again.