bug(adapters): hallucination_detection output isn't checked against its own contract #1719

flag_hallucinated_content can return records that break the adapter's published output contract, with no error. Callers then have to guess what a record means.

• A missing field comes back as None. The io.yaml project step builds each record with .get() (mellea/formatters/granite/intrinsics/output.py:908-911), so a record the model wrote without f arrives as faithfulness: None. The required-key check in _ListContract never fires, because every key is always present.
• Any faithfulness string passes. The io.yaml declares four values (faithful, partial, unfaithful, NA); the contract checks keys only (mellea/backends/adapters/_core.py:293-326). Constrained decoding may enforce the values on some backends, but nothing in mellea checks them, and feat(adapters)!: prompt fallback for adapter functions without a trained adapter #1654's prompt fallback will parse plain prompted output through the same contract.
• A bad sentence index crashes or gives wrong data. DecodeSentences._transform (output.py:718-735) checks only isinstance(value, int). An index past the end raises a bare IndexError at output.py:727. A negative index silently picks a sentence from the end of the list, and True/False pass the type check.

The docstring doesn't help: it gives the values as "e.g. "faithful", "hallucinated"" (mellea/stdlib/components/intrinsic/rag.py:257-259). There's no hallucinated, and partial and NA aren't mentioned.

Proposed fix

1. A dedicated contract for hallucination_detection. Add _HallucinationDetectionContract to mellea/backends/adapters/io_contracts.py and register it in _INTRINSIC_IO_CONTRACTS in place of the _ListContract entry (:209-220). _RequirementCheckContract (:134) is the model, since it already does value-level checks. Keep the existing required-key behaviour, including "explanation" in expected_keys (test/stdlib/components/intrinsic/test_rag_contracts.py:141 checks it). Then, in this order:

   1. Per record, raise AdapterSchemaMismatchError when faithfulness isn't one of the four io.yaml values (this also catches None), response_begin or response_end isn't an int or is a bool, begin > end, or explanation isn't a string (the io.yaml marks e as required).
   2. Duplicates. Group records by (response_begin, response_end). If a group disagrees on faithfulness, raise, since there's no defined way to pick one. If it agrees, keep the first and drop the rest; the io.yaml has no drop_duplicates and the model can repeat itself.
   3. Overlap. Raise if any two remaining records overlap, comparing half-open ranges because adjacent sentences share a boundary. Decoded sentences are contiguous and non-overlapping (_desplit_sentences, output.py:435), so after step 2 this is a guard rather than a common case.

   What parse() returns is what callers receive (mellea/stdlib/components/intrinsic/_util.py:255), so the contract can do the dropping. No existing contract changes the data it returns, so say so in the class docstring.

   Leave _ListContract alone. citations and context-attribution use it too, and their output legitimately repeats response spans (one record per pair of response sentence and document sentence) and carries nulls. A generic check would reject valid citations, and GroundednessRequirement would then fail grounded responses.

   The value set is closed on purpose: an unknown label has no defined meaning, so it raises. On the local-file and Ollama paths a new upstream label can only arrive with a catalogue pin bump, which should update the contract. On Granite Switch the io.yaml comes from the served model's own io_configs/ (mellea/backends/adapters/adapter.py:1348-1376), while the contract comes from this registry (adapter.py:1339). A Switch release with a new label would therefore raise for Switch users with no pin bump. That's the intended fail-closed behaviour, but note it in the docstring, and make the error say what to do: name the unknown label, list the four this version understands, and suggest upgrading mellea.

2. Bounds-check DecodeSentences. In _transform (output.py:718-735), raise TypeError for a bool (matching the existing TypeError for non-integers at :718-722), and ValueError for negative and out-of-range indices, naming the index and the sentence count.

   • ValueError matches the direction in bug(intrinsics): malformed output raises bare Exception, not the documented ValueError (HF, Ollama, OpenAI) #1688, and keeps mellea/formatters/granite/ from importing anything from mellea/backends/.
   • Once this is in place, the offsets come from real sentences, so no separate "offset within the response" check is needed. That check couldn't go in the contract anyway: parse(raw) never sees the response text.
   • The transform is shared with citations and context-attribution. For them, an index past the end already raises, so only the exception type and message change. A negative index goes from silently wrong to an error. For context-attribution the bound is the combined documents-plus-history sentence list (output.py:608-708).

3. Fix the docstrings. In flag_hallucinated_content, list the four values with their io.yaml meanings, update the output-contract paragraph (rag.py:261-265) for the new checks, and add ValueError and TypeError to Raises: (rag.py:285-289). find_citations and find_context_attributions go through the same DecodeSentences, so add the new ValueError to their Raises: too. This adds raise statements, so the docstring quality gate in AGENTS.md §10 applies.

Tests (no GPU)

• Contract, in test/backends/test_adapters/test_io_contracts.py. Each of these raises AdapterSchemaMismatchError: an unknown faithfulness, faithfulness: None, a non-int offset, a bool offset, begin > end, two records for one sentence with different faithfulness, and overlapping records. Adjacent records, extra fields and an empty list pass, and a repeated record with the same faithfulness is dropped, keeping the first.
• DecodeSentences. A bool raises TypeError; negative and out-of-range indices raise ValueError. Valid indices are unchanged, and one citations and one context-attribution case still decode as before.
• Existing tests. Everything in test/stdlib/components/intrinsic/test_rag_contracts.py keeps passing. _GOOD_SPAN (:125-131) has response_end=10 against a 9-character response_text, so don't add a check that the text length matches the offsets.

Out of scope

• Truncated output, which today reaches the caller as a bare Exception (bug(intrinsics): malformed output raises bare Exception, not the documented ValueError (HF, Ollama, OpenAI) #1688).
• Telemetry. A contract violation is raised in call_intrinsic (_util.py:255), after the backend has already recorded the adapter invocation as a success (for example mellea/backends/huggingface.py:1651-1655), so adapter telemetry won't show it. Fixing that touches the handlers feat(adapters)!: prompt fallback for adapter functions without a trained adapter #1654 rewrites.
• Checking that the records cover the whole response. That only matters if GroundednessRequirement starts using this adapter, which feat(rag): replace groundedness judgment prompts with the hallucination_detection adapter #1669 has on hold. An empty list stays valid.

Who notices

• Callers of flag_hallucinated_content. Output that used to come back quietly wrong (a None label, an unknown label, a record pinned to the wrong sentence) now raises. Call this out in the PR description, since the changelog is generated from PR titles.
• GroundednessRequirement. Its first stage calls find_citations, which shares DecodeSentences. A citation record with a negative or bool index used to give wrong offsets and possibly a wrong verdict; now it raises, and bug(rag): GroundednessRequirement reports its own failures as "not grounded" #1718 turns that into an errored result.

#1654 doesn't touch io_contracts.py or output.py. Its prompt fallback does register adapters with get_io_contract(name), so the new contract checks prompt-fallback output too.

Line references are against 46ad3b3a5.