Invalid-confidence warning is permanent: _fold_edge_aliases normalizes numeric legacy confidences but not string ones (high/medium/confirmed) #4495

Summary

Any graph whose semantic (LLM) pass emitted a string confidence outside the enum re-warns forever on every command:

[graphify] Extraction warning (259 issues): 259x other schema issue
(e.g. Edge 7335 has invalid confidence 'confirmed' - must be one of ['AMBIGUOUS', 'EXTRACTED', 'INFERRED'])

_fold_edge_aliases() in graphify/build.py normalizes only numeric legacy confidences; free-form strings pass through untouched and are never repaired by any CLI command.

Environment

• graphify / graphifyy 0.9.91, Windows, Python 3.15
• Corpus: mixed code + docs (OpenSpec markdown), 2777 nodes / 6648 edges, 153 communities
• The graph was built by an earlier full build (semantic pass = host agent); graphify update . only re-extracts code, so the offending edges persist.

Impact

• 259 edges carry a non-enum confidence: high = 154, confirmed = 93, medium = 6, inferred = 3, medium-high = 2, medium-low = 1.
• All 259 have _origin: semantic, all have confidence_score, and all come from documents (253 .md, 6 .yaml).
• graph["hyperedges"] carries the same class of bad values: floats (e.g. 0.75, 0.85, 0.9) and strings (high, medium, inferred, medium-high).
• The warning is re-emitted on every graphify update, query, explain, label, … and is not fixable from the CLI.

Root cause

graphify/build.py:

def _fold_edge_aliases(edge: dict) -> None:
    ...
    _conf = edge.get("confidence")
    if isinstance(_conf, (int, float)) and not isinstance(_conf, bool):
        if edge.get("confidence_score") is None:
            edge["confidence_score"] = float(_conf)
        edge["confidence"] = "INFERRED"
    if not edge.get("confidence") and edge.get("confidence_score") is not None:
        edge["confidence"] = "INFERRED"

The docstring already reasons: "numeric confidences only ever came from the LLM semantic pass, and LLM-derived edges are INFERRED by definition" and "Without this fold, every reload of a pre-enum graph re-warns once per legacy edge, forever." That reasoning applies verbatim to string legacy values (high / medium / medium-high / medium-low / confirmed / lowercase inferred), but they are not folded — so they re-warn forever, exactly as the docstring warns for numerics.

The extraction prompt (references/extraction-spec.md) asks for "confidence":"EXTRACTED|INFERRED|AMBIGUOUS" plus a separate confidence_score float, so this string vocabulary is model drift, not a documented format.

Suggested fix

Extend the fold to non-enum strings, case-insensitively, mapping to INFERRED (LLM-derived by definition), moving a parseable value into confidence_score:

_ALLOWED = {"AMBIGUOUS", "EXTRACTED", "INFERRED"}
_conf = edge.get("confidence")
if isinstance(_conf, (int, float)) and not isinstance(_conf, bool):
    if edge.get("confidence_score") is None:
        edge["confidence_score"] = float(_conf)
    edge["confidence"] = "INFERRED"
elif isinstance(_conf, str) and _conf.upper() not in _ALLOWED:
    try:
        if edge.get("confidence_score") is None:
            edge["confidence_score"] = float(_conf)
    except ValueError:
        pass
    edge["confidence"] = "INFERRED"
if not edge.get("confidence") and edge.get("confidence_score") is not None:
    edge["confidence"] = "INFERRED"

The same fold should be applied to graph["hyperedges"], which currently is not normalized at all.

Steps to reproduce

1. Build a graph whose semantic pass emits a non-enum string confidence (e.g. {"confidence": "high", "confidence_score": 0.85}).
2. Run any graphify command (graphify query "...", graphify update .).
3. Observe the "invalid confidence" schema warning; re-running never clears it and the edge values stay unchanged.

Workaround

Post-process graph.json: map every non-enum confidence to INFERRED (and mirror the same mapping into graph.hyperedges).