[Bug]: interrupted FileSlice extraction marks an incomplete file as fully cached #4943

What happened

An interrupted extraction can cache one successful FileSlice as the complete parent file.
A later semantic cache check then reports no uncached files, although later slices never completed.
This can silently omit content after the interruption.

Expected: a strict cache check must keep the parent file uncached until every required slice completes successfully.

Environment and scope

• Reproduced twice with the installed graphifyy==0.9.82 package, Python 3.13.11, macOS arm64, installed through uv tool.
• The example uses the native extract_corpus_parallel and check_semantic_cache functions.
• Only extract_files_direct is mocked. No model request, API key, provider response, or private source is required.
• The first call returns a complete result for its slice. The second call raises KeyboardInterrupt.
• This is an installed-package repro. I did not run the current upstream checkout or its full test suite.

Steps to reproduce

Run this script with the Python interpreter that contains Graphify.

import tempfile,json
from pathlib import Path
from unittest.mock import patch
from importlib.metadata import version
from graphify import llm
from graphify.cache import check_semantic_cache
with tempfile.TemporaryDirectory() as directory:
 root=Path(directory);p=root/'guide.md';p.write_text('# Early\nEarly fact.\n'+'filler '*6000+'\n# Tail\nTAIL_ASSERTION must remain visible.\n');calls=[]
 def endpoint(files,**kwargs):
  calls.append(len(files))
  if len(calls)>1:raise KeyboardInterrupt()
  return {'nodes':[{'id':'early','label':'Early fact.','source_file':str(p),'source_location':'Early'}],'edges':[],'hyperedges':[],'input_tokens':1,'output_tokens':1,'finish_reason':'stop','model':'fixture'}
 try:
  with patch.object(llm,'extract_files_direct',endpoint):llm.extract_corpus_parallel([p],backend='openai',root=root,cache_root=root/'out',max_concurrency=1,token_budget=1,max_retry_depth=0)
 except KeyboardInterrupt:pass
 nodes,edges,hyp,miss=check_semantic_cache([str(p)],root=root,cache_root=root/'out',prompt=llm._extraction_system())
 entries=list((root/'out/graphify-out/cache/semantic').rglob('*.json'));data=json.loads(entries[0].read_text()) if entries else {}
 print(json.dumps({'version':version('graphifyy'),'attempted_chunks':len(calls),'successful_chunks':1,'cache_entries':len(entries),'partial_flag':data.get('partial',False),'warm_uncached_files':len(miss),'warm_nodes':len(nodes),'tail_present':any('TAIL_ASSERTION' in str(n) for n in nodes)}))

Actual output

{"version": "0.9.82", "attempted_chunks": 2, "successful_chunks": 1, "cache_entries": 1, "partial_flag": false, "warm_uncached_files": 0, "warm_nodes": 1, "tail_present": false}

The displayed successful_chunks: 1 value is a fixture constant, not a separate measured counter.
This repro exercises only the serial branch (max_concurrency=1).

The cache contains one entry with partial_flag: false.
warm_uncached_files: 0 means the native checker treats the entire file as cached.
The tail assertion is absent because extraction stopped before its slice completed.

Source investigation

The current v8 source at commit 32e5efe935e6844ef39a5e33409e4efe848f2172 appears to retain the relevant mechanism.
This is source inspection, not a runtime reproduction on that commit.

_checkpoint_chunk maps each slice to its parent path with unit_path.
It calls save_semantic_cache with merge_existing=True.
Its partial set comes from _partial_source_files(result), which does not establish completion of sibling slices.
A successful individual slice can therefore produce an authoritative entry for an incomplete file.

A possible fix is to track completion per parent file and keep interim checkpoints partial.
Any fix also needs native API finalization and final-save protection when one required slice fails.
Existing entries from interrupted runs may need targeted invalidation because their metadata can falsely claim completeness.
These are proposed fix requirements. No patch is included.

Related issues checked

• Incremental semantic cache checkpoint fails on sliced docs: _checkpoint_chunk reads FileSlice.rel, which does not exist #1870 fixed a FileSlice.rel access error that prevented checkpoints. Here the checkpoint succeeds but incorrectly represents file completeness.
• Semantic extraction cannot tell a file with no entities from an omitted file, so empty files arm the shrink guard #4547 concerns empty-file acknowledgements and omitted sources. This repro returns a nonempty successful slice before interruption.
• [Bug]: a .graphify_partial.txt left by an --update can hold a cached doc back in a later full build #4814 concerns stale partial-state instructions. This repro demonstrates a native semantic cache entry that lacks the partial marker.

I checked the README Troubleshooting section and its linked documentation before reporting this issue.