v0.4.0

Added

• Geometry-integral intent parser: the parser IS the Construct's inference machinery running in reverse. Phase 1 projects intent onto face centroids (discriminative cosine similarity). Phase 2 projects onto axis direction vectors (high_pole - low_pole). Phase 3 maps scalars to grid via polarity convention.
• GloVe 6B 100d semantic bridge: 15K-word vocabulary with pre-computed face similarity and axis projection matrices. Built at dev time, ships as ~2MB numpy artifacts. Zero ML runtime dependencies.
• Contextual disambiguation table: 8 polysemous trigger words (state, compelled, right, deep, forces, heaven, tragedy, action) with 10 context-aware senses. Overrides face/axis scores when >= 2 context indicator words are present.
• N-gram phrase embeddings: 92 curated phrases (domain replacements, pole pair bigrams, philosophical key phrases). Greedy longest-match tokenizer with surface-to-canonical mapping.
• Contrastive cube-pair dampening: within each complementary pair, transfers 30% of score difference from weaker to stronger face.
• 46 pole synonym clusters (~320 curated words) for GloVe centroid construction.
• GeometricBridge class (replaces SemanticBridge): face_relevance() + axis_projection() with disambiguation and phrase support.
• 39 new tests for GeometricBridge and geometry-integral parser.
• Semantic bridge algorithm specification (docs/specs/semantic-bridge-algorithms.md).
• Build script (scripts/build_semantic_bridge.py) with pole self-test validation (all 24 axes pass).

Changed

• Intent parser no longer uses TF-IDF or keyword matching for face/position selection — replaced by geometric projection.
• Tokenizer is now unstemmed (GloVe needs word forms) with expanded stop-word list (~95 words).
• Face centroids built from authored layers only (core questions + sub-dimension labels + pole synonyms) — NOT from derived question templates.

Removed

• Keyword-based face matching (_FACE_KEYWORDS dictionary).
• TF-IDF dependency in Stage 1 (intent parser no longer queries construct questions).
• Stemming in intent parser tokenizer.

Performance

• Literary text benchmark: 15/20 expected faces in top 6 across 8 benchmark texts (Shakespeare, Bible, Marx, MLK, Newton, Aristotle, Tao Te Ching, Descartes).
• 300 tests passing.