fix: 4 pipeline bugs — domain knowledge passthrough, torch import guard, bare excepts

[cauchyturing]

---

Problem

Found 4 bugs while working with the pipeline — domain knowledge never reaches the LLM prompts despite being collected, import torch crashes on CPU-only environments, and bare except: blocks silently swallow KeyboardInterrupt / SystemExit.

Changes

1. causal_discovery/filter.py — dead [DOMAIN_KNOWLEDGE] placeholder

The prompt template algo_select_prompt.txt has a [DOMAIN_KNOWLEDGE] placeholder at line 25, but create_prompt() never included it in the replacement dict. Domain knowledge collected from the user was silently dropped during algorithm selection.

Fix: Extract knowledge_docs from global_state.user_data.knowledge_docs, format it, and add "[DOMAIN_KNOWLEDGE]": knowledge to the replacements dict.

Also: import torch at module level crashes in environments without CUDA/torch (e.g., lightweight containers, CI). Wrapped in try/except ImportError and guarded torch.cuda.is_available() with a torch is not None check.

2. postprocess/judge.py — knowledge_docs never forwarded

quality_judge() receives knowledge_docs as a parameter (line 34) and it's passed correctly from forward() (line 146), but the actual call to llm_evaluation_new() at line 94 never forwarded it. The Judge's LLM edge evaluation was always running without domain knowledge, even when the user provided it.

Fix: Add knowledge_docs=knowledge_docs to the llm_evaluation_new() call.

3. postprocess/judge_functions.py — accept + inject domain knowledge, fix bare excepts

llm_evaluation_new() didn't accept a knowledge_docs parameter, so even after fixing the call site in judge.py, there was nowhere for the knowledge to go.

Fix:

• Added knowledge_docs=None parameter to llm_evaluation_new().
• When provided, formats it into a **Domain Knowledge** section and appends to the [RELATIONSHIP] replacement so it reaches the pruning prompt.
• Replaced 5 bare except: with except Exception: — bare excepts catch KeyboardInterrupt, SystemExit, GeneratorExit which makes debugging painful and can prevent clean shutdown.

4. llm/__init__.py — hard crash without openai package

from llm import LLMClient fails with ImportError if openai isn't installed, even in code paths that never use the LLM client (e.g., rule-based algorithm selection). Same for OllamaClient.

Fix: Wrap both imports in try/except ImportError, defaulting to None. Callers that actually need these classes will get a clear error at instantiation time rather than a cryptic import crash at module load.

Testing

Ran targeted verification on each fix:

=== Test 1: torch import guard in filter.py ===
PASS: filter.py loads without torch

=== Test 2: llm/__init__.py import guard ===
PASS: 2 try/except blocks found (expected 2)

=== Test 3: judge.py passes knowledge_docs to llm_evaluation_new ===
PASS: knowledge_docs forwarded to llm_evaluation_new()

=== Test 4: judge_functions.py signature + knowledge injection ===
PASS: llm_evaluation_new() accepts knowledge_docs parameter
PASS: knowledge_docs injected into pruning prompt

=== Test 5: no bare except: in modified functions ===
PASS: no bare except: in judge_functions.py

=== Test 6: [DOMAIN_KNOWLEDGE] placeholder connected in filter.py ===
PASS: [DOMAIN_KNOWLEDGE] placeholder in replacements dict
PASS: knowledge_docs extracted from global_state

Confirmed algo_select_prompt.txt:25 contains [DOMAIN_KNOWLEDGE] — the placeholder was always there, just never replaced.

Confirmed no bare except: remain in any of the 4 modified files.

Impact

• Domain knowledge now flows end-to-end: user input → knowledge_docs → Filter prompt (algorithm selection) → Judge prompt (edge pruning). Previously it was collected but dropped at both stages.
• No behavior change when knowledge_docs is None — all new code paths are guarded with if knowledge_docs checks, so existing pipelines without domain knowledge are unaffected.
• No new dependencies. All changes are backward compatible.