hyperedge-only doc output never stamped in manifest — perpetual re-extraction

[Yashasm18]

Bug

_stamped_manifest_files() (added in #1897 to fix zero-node docs being stamped incorrectly) decides whether a semantic document "produced output" by checking only the nodes and edges collections of sem_result. It never checks hyperedges.

sem_extracted: set[Path] = set()
for coll in ("nodes", "edges"):
    for item in sem_result.get(coll, []):
        sf = item.get("source_file", "")
        if sf:
            sem_extracted.add(_resolve(sf))

If a chunk's only output for a given document is a hyperedge (a coherent group of 3+ existing nodes/concepts spanning that doc — the extraction skill prompt explicitly supports this: "if 3 or more nodes clearly participate together in a shared concept... add a hyperedge"), that document is left out of sem_extracted.

This means:

• _response_is_hollow() in llm.py correctly treats hyperedges as valid output (not nodes and not edges and not hyperedges), so the chunk is NOT flagged as a failure/truncation.
• But _stamped_manifest_files() disagrees — it excludes the doc from the manifest anyway.
• semantic_hash never gets stamped for that file.
• detect_incremental() re-queues the file as "changed" on every subsequent graphify extract run, even though it was successfully extracted.

Observed with

• graphifyy 0.9.16/0.9.17, code reading (not yet reproduced against a live LLM backend — filing from source-level analysis, repro below is a direct unit-level trigger)

Reproduction

from graphify.cli import _stamped_manifest_files

files_by_type = {"document": [str(doc_path)]}
sem_result = {
    "nodes": [],
    "edges": [],
    "hyperedges": [{"id": "h1", "label": "...", "nodes": ["a", "b", "c"], "source_file": "doc.md"}],
}
out = _stamped_manifest_files(files_by_type, sem_result, root)
# out["document"] == [] — doc.md is dropped even though it produced a hyperedge

Expected behavior

A document whose only extraction output is a hyperedge should be treated the same as one with a node or edge: included in sem_extracted, stamped in the manifest, and not re-queued on the next incremental run.

Suggested fix

Add "hyperedges" to the iterated collections in _stamped_manifest_files, matching _response_is_hollow's definition of "produced output":

for coll in ("nodes", "edges", "hyperedges"):
    for item in sem_result.get(coll, []):
        ...

Impact

Documents whose semantic value is purely relational (a synthesis/overview doc describing how several existing entities relate) never get a stable semantic_hash, so they're silently re-sent to the LLM backend on every graphify extract run — wasted tokens/cost with no visible symptom other than persistently non-zero "changed files" counts.

[safishamsi]

Fixed on v8 in cb96bda (ships in 0.9.17). _stamped_manifest_files now counts hyperedge output, so a doc whose only chunk output is a hyperedge is stamped instead of re-extracted on every run. Mirrors the per-source_file keying the semantic cache already uses. Regression tests added.