fix(llm): pin claude-cli output to a JSON schema when supported (#2076)

[Yyunozor]

Summary

The claude-cli backend puts the extraction schema in the user turn and trusts the model to emit raw JSON (the llm.py:1376 workaround, "verified against Claude Code 2.1.197"). On newer Claude Code the model treats that prompt as an agentic task and reports the result in prose instead — Knowledge graph extracted — 21 nodes, 20 edges…. That parses to zero nodes, _response_is_hollow reads truncation, and adaptive-retry bisects the chunk without ever converging, so the graph comes out empty (#2076).

Why this is version-robust, and how it was checked

We run Claude Code as this backend's harness, so the failing component is directly testable here. On 2.1.207 the current user-turn workaround still returns JSON — the empty-graph break does not reproduce there, so the regression is version-gated, consistent with the 2.1.216 report. But the failure mode is easy to force and to fix on a live CLI: a prompt that invites prose makes claude -p --output-format json put a conversational summary in result, while adding --json-schema makes the same call return the object regardless of framing (also surfaced in a structured_output envelope field). Structured output is a structural guarantee independent of the model's discretion, so it survives the prompt-behaviour drift that broke the prompt-shape workaround.

Fix

Pass --json-schema — a schema pinning the top-level {nodes, edges} shape graphify consumes, item internals left loose so valid extractions are never rejected — when the CLI advertises the flag, probed once via claude --help and cached. Older CLIs that predate the flag keep the user-turn prompt as a fallback, unchanged. The result envelope still carries the JSON string, so the _parse_llm_json path is untouched.

The issue's second suggestion — routing a successful-sounding prose response away from the bisect path — is complementary and left out of scope; with structured output the prose case no longer arises on supported CLIs.

Validation

tests/test_claude_cli_backend.py: 28 passed, 5 new (flag present when supported, absent when not, the --help capability probe, safe fallback on probe error). Full suite: 3317 passed on the branch vs 3312 on the parent e32c9f4, same environment — delta is exactly these 5 tests. (13 failures in unrelated suites — terraform/ollama/manifest/install-references — are pre-existing, identical on the parent, from optional deps absent locally.) ruff check clean.