[S2] F metric double-defined — operator and behavioral layers measure incomparable quantities

[erinepshovel-code]

The letter F is assigned two incompatible operationalizations:

File	F means
backend/src/edcmbone/metrics/compute.py::_compute_F	Fixation risk — repetition rate + n-gram overlap + non-novelty
core/behavioral/behavioral_metrics.py	Feature cosine similarity — cosine similarity to previous window's feature vector

core/bridge/bridge_engine.py then correlates operator-side outputs against behavioral-side outputs. The F↔F correlation entry compares a repetition measure against a vector-similarity measure — two quantities with no shared semantics. Any bridge output that uses this correlation is misleading.

Fix options:

1. Rename the behavioral-side key (e.g. F_sim, F_cos, feat_cos).
2. Rename the operator-side key (e.g. F_fix, fixation).
3. At minimum, add a comment in bridge_engine.py noting the naming collision so readers do not interpret the F↔F bridge correlation as comparing like with like.

The rename needs to propagate through RoundMetrics, bridge_engine correlation tables, and any downstream consumers of the metric keys.

[erinepshovel-code]

@copilot fix this