Engine Phase 2: fix the pre-Opus funnel (lifts ALL matches)

The sift was feeding Opus a magnitude-blind, under-trusted, truncated shortlist — only 68% of the
genuinely-best jobs reached the curator. Fixes:
- embed.py: embed the FULL description (was desc[:600]) + role_category + industry; richer profile
  text (current_role + experience summary + seniority). The best semantic signal, fed real content.
- sift.py: magnitude-preserving min-max fusion (was rank-position, which flattened cosine 0.95 vs
  0.72) and embedding-LED weights (W_VIBE 0.6 / W_WHITEBOX 0.4). Embed runs FIRST so the white-box
  f_semantic reuses the real cosine (threaded as job["_vibe_cosine"]) instead of token overlap.
- config.py: SIFT_TOP_K 18 → 28 (the gate was cutting good jobs before Opus).
- curate.py: richer Opus briefs — full responsibilities (was desc[:500]) + role_category/industry/
  seniority/pay, so even the shortlist is fully described.

Regression (120 pairs): sift membership recall 0.68 → 0.92, score MAE 24.7 → 18.4. 31 tests pass.

tests/run_regression.py

@@ -20,6 +20,7 @@ from pathlib import Path
 
 from app.engine import rank as _rank
 from app.engine import sift as _sift
+from app.engine import embed as _embed
 from app.config import get_settings
 
 FIXTURE = Path(__file__).parent / "fixtures" / "regression_set.json"
@@ -39,9 +40,17 @@ def run(emit_json: bool = False):
     for c in cases:
         prefs, ctx, jobs = c["prefs"], c.get("user_context"), c["jobs"]
         targets = {str(j): int(s) for j, s in c["targets"].items()}
-        # engine: white-box scores + the sift top-K the curator would see
-        ranked = _rank.rank(prefs, ctx, [dict(j) for j in jobs])["opportunities"]
+        # engine score = the funnel's white-box AFTER the embedding cosine is threaded in (what the
+        # fallback shows + feeds the sift) — replicate embed→thread→rank so MAE reflects the real funnel.
+        jc = [dict(j) for j in jobs]
+        vibe = _embed.vibe_scores(prefs, ctx, jc)
+        if vibe:
+            for j in jc:
+                if j["id"] in vibe:
+                    j["_vibe_cosine"] = vibe[j["id"]]
+        ranked = _rank.rank(prefs, ctx, jc)["opportunities"]
         eng = {j["id"]: j.get("matchScore", 0) for j in ranked}
+        # the sift top-K the curator actually sees
         top, _ = _sift.sift(prefs, ctx, [dict(j) for j in jobs], k=k)
         top_ids = {j["id"] for j in top}
 

tests/test_sift.py

@@ -1,12 +1,13 @@
-"""A1 — sift: rank-fusion math + graceful degradation when the vibe-engineer is unavailable.
+"""A1 — sift: magnitude-fusion math + graceful degradation when the vibe-engineer is unavailable.
 (Offline — no network; the embedding path is exercised in the cached e2e verification.)"""
 from app.engine import sift as S
 
 
-def test_ranks_normalization():
-    r = S._ranks({"a": 0.9, "b": 0.5, "c": 0.1})
-    assert r["a"] == 1.0 and r["c"] == 0.0 and 0.0 < r["b"] < 1.0
-    assert S._ranks({"x": 0.5}) == {"x": 1.0}                 # single item → top
+def test_minmax_normalization():
+    # magnitude-preserving (min-max), NOT rank-position: a much-better score stays much better.
+    r = S._minmax({"a": 0.9, "b": 0.5, "c": 0.1})
+    assert r["a"] == 1.0 and r["c"] == 0.0 and abs(r["b"] - 0.5) < 1e-9
+    assert S._minmax({"x": 0.5}) == {"x": 1.0}                # single item / no range → top
 
 
 def test_sift_degrades_to_whitebox_when_no_vibe(monkeypatch):