fix(llm): distinguish failed chunks from model omissions

[GREYGROUPJP]

Problem

When a semantic chunk raises before returning a usable result, the final reconciliation warning says the model returned a response but omitted the affected files. This contradicts the preceding chunk error and points operators toward model-output debugging instead of the dependency or provider failure that actually occurred.

Change

Track the files assigned to failed chunks and split uncovered files into two groups:

• failed before returning a usable result; and
• omitted from a successful model response.

Each group now receives an accurate warning. The existing omission wording remains unchanged for successful responses that omit files.

Fixes #3616.

Validation

• Added a regression test for a missing optional dependency/backend exception.
• Kept coverage for genuine model omissions and asserted the original warning remains.
• pytest tests/ -q --tb=short: 5,643 passed, 96 skipped.
• ruff check graphify/llm.py tests/test_chunking.py: passed.
• git diff --check: passed.

[graphify-labs]

Graphify reviewed this change.

Looks safe to merge — no coupling regressions and no blocking issues, checked against the code graph (not a self-assessment).

Formal verification. No changes could be formally verified in this run.

---

Graphify review — findings

Splits the corpus-parallel uncovered-files warning into two distinct messages: files whose semantic chunk raised before returning a result now get a "failed before returning a usable result" warning that points at the chunk errors above, while files the model genuinely returned but omitted keep the "returned a response but omitted them" message. Tracks which paths belonged to failed chunks via a new failed_files set so the two cases are classified correctly at reconciliation time.

No blocking issues surfaced. 2 lower-confidence candidates did not survive cross-model review.

Analysis details — impact, health, verification

Impact & health

Graphify review

Impact — 913 functions depend on the 270 functions this change touches.

Health — this change adds coupling hotspots:

• new: deduplicate_entities() — 77 callers, 24 callees
• new: build_merge() — 76 callers, 14 callees
• new: extract_files_direct() — 17 callers, 20 callees
• new: build() — 52 callers, 6 callees
• new: extract_corpus_parallel() — 27 callers, 11 callees
• new: _call_claude_cli() — 33 callers, 9 callees
• new: dispatch_command() — 2 callers, 125 callees
• new: _extract_with_adaptive_retry() — 22 callers, 10 callees
• …and 16 more — each is listed as a finding

Verification — 913 functions in the blast radius were not formally verified this run (proofs are advisory here).

Gate & verification

graphify gate

PASS — objectively clean (no health regressions, tests not run — proofs not run this pass (advisory)). Grounded, not self-assessed.

Advisory (not blocking):

• verification_scope: 578 function(s) in the blast radius were not formally verified this run

Test selection

Test selection

41 of 286 test file(s) selected (14%) via static blast radius.

• tests/test_backend_env_isolation.py — impact
• tests/test_backend_extras.py — impact
• tests/test_build.py — impact
• tests/test_build_merge_dedup_scope.py — impact
• tests/test_build_merge_hyperedges_and_prune.py — impact
• tests/test_build_merge_shrink_guard.py — impact
• tests/test_carried_hyperedge_remap.py — impact
• tests/test_charmap_encoding.py — impact
• tests/test_chunking.py — impact, changed-test
• tests/test_claude_cli_backend.py — impact
• tests/test_corrupt_graph_json.py — impact
• tests/test_cross_extension_reexport_self_cycle.py — impact
• tests/test_dedup.py — impact
• tests/test_dedup_remaps_hyperedges.py — impact
• tests/test_dedup_survivor_richness.py — impact
• tests/test_evidence_binding.py — impact
• tests/test_file_slice.py — impact
• tests/test_global_graph.py — impact
• tests/test_go_qualified_resolution.py — impact
• tests/test_hyperedge_member_shapes.py — impact
• tests/test_image_vision.py — impact
• tests/test_injection_sentinel_coverage.py — impact
• tests/test_issue_3472_source_file_collision.py — impact
• tests/test_label_retry.py — impact
• tests/test_labeling.py — impact
• tests/test_llm_backends.py — impact
• tests/test_llm_parser.py — impact
• tests/test_llm_parser_reasoning.py — impact
• tests/test_no_dedup_flag.py — impact
• tests/test_non_string_node_ids.py — impact
• tests/test_ollama.py — impact
• tests/test_ollama_retry_cap.py — impact
• tests/test_oversized_document_slicing.py — impact
• tests/test_partial_cache.py — impact
• tests/test_pdf_slicing.py — impact
• tests/test_pdf_token_estimate.py — impact
• tests/test_provider_registry.py — impact
• tests/test_prs.py — impact
• tests/test_prune_sweeps_orphans.py — impact
• tests/test_semantic_fragment_sanitize.py — impact
• tests/test_unverified_semantic_shrink.py — impact

Selection is safe under the controlled-regression assumption; always-run tests + a periodic full run are the backstops. Advisory — it never changes the check verdict.

Formal verification

Could not verify: Could not verify extract\_corpus\_parallel.

The verifier did not have enough to check extract\_corpus\_parallel, so it is saying so rather than guessing. No false assurance is the whole point.

Guarantee: No guarantee either way, this is an honest abstention, not a pass.

Note: Reason: no capturable inputs from the test suite; property tier: parameter `root` is annotated `Path` — outside the synthesizable primitive/collection set

· 24 more finding(s) on lines outside this diff (see the check run).