Fix 4 bugs: ZIP fallback, Emax bound, NCA terminal min-points, log-rank tolerance

drug_synergy_tool.py — Bug #7 (CRITICAL): ZIP fallback was Bliss, not ZIP
- ZIP requires Hill curve fitting; without it a Bliss independence surface was
  silently returned labelled as a ZIP delta score — scientifically wrong
- Now returns a clear error naming the drug(s) that failed Hill fitting and
  directing the user to calculate_bliss for a non-parametric alternative

drug_synergy_tool.py — Bug #8: Emax upper bound 1.5 in Loewe/CI Hill fitter
- _fit_hill_for_loewe uses fractional inhibition (0-1); Emax > 1 is physically
  impossible, yet the scipy curve_fit upper bound allowed 1.5
- Corrected to 1.0 (Loewe & Muischnek 1926; Chou & Talalay 1984)

nca_tool.py — Bug #4: terminal slope accepted 2 instead of 3 post-Tmax points
- FDA NCA guidance requires 3+ points for lambda_z: a 2-point linear regression
  is a perfect fit by construction (R2=1 trivially), giving no quality signal
- Now requires 3+ post-Tmax positive concentrations before using the full set;
  falls back to last n_points otherwise

survival_tool.py — Bug #6: log-rank zero-variance check used absolute tolerance
- Threshold 1e-10 too tight for large studies: float rounding errors over many
  event times scale as N x 2.2e-16, so 1e-10 would erroneously reject valid
  proportional distributions in datasets with thousands of events
- Replaced with relative tolerance: 1e-8 x total_events (safe up to 10^7 events)

Author: gasvn

src/tooluniverse/drug_synergy_tool.py

@@ -289,12 +289,24 @@ def fit_hill(doses, effects):
         params_b = fit_hill(db, effects_b_marginal)
 
         if params_a is None or params_b is None:
-            # Fallback: simplified ZIP using means
-            expected_zip = (
-                np.outer(effects_a_marginal, np.ones(len(db)))
-                + np.outer(np.ones(len(da)), effects_b_marginal)
-                - np.outer(effects_a_marginal, effects_b_marginal)
-            )
+            # ZIP fundamentally requires Hill curve fits for each drug to compute
+            # the expected interaction surface.  Without fitted Hill parameters we
+            # cannot compute a valid ZIP score — returning a Bliss-independence
+            # surface here would silently report the wrong model.
+            failed = []
+            if params_a is None:
+                failed.append("drug A")
+            if params_b is None:
+                failed.append("drug B")
+            return {
+                "status": "error",
+                "error": (
+                    f"ZIP model requires Hill curve fitting for each drug, but fitting "
+                    f"failed for {' and '.join(failed)}. "
+                    "Ensure each drug has ≥3 non-zero dose points with measurable inhibition. "
+                    "Use calculate_bliss for a simpler non-parametric synergy score."
+                ),
+            }
         else:
             # ZIP expected using Hill fits
             pred_a = np.array([hill_curve(d, *params_a) for d in da])
@@ -348,7 +360,7 @@ def _fit_hill_for_loewe(self, doses, effects):
             dm_init = float(np.median(d))
             m_init = 1.0
             p0 = [dm_init, m_init, emax_init]
-            bounds = ([1e-15, 0.1, 1e-6], [np.inf, 10.0, 1.5])
+            bounds = ([1e-15, 0.1, 1e-6], [np.inf, 10.0, 1.0])
 
             def hill(x, dm, m, emax):
                 x = np.maximum(x, 1e-15)

src/tooluniverse/nca_tool.py

@@ -153,15 +153,19 @@ def _estimate_terminal_slope(
             return None, None, None
 
         # Select terminal phase points
+        # FDA NCA guidance requires ≥3 points for reliable λz: a 2-point regression
+        # is a perfect fit by construction (R²=1 always), giving no quality signal.
         if tmax is not None:
             post_tmax = t_valid > tmax
-            if np.sum(post_tmax) >= 2:
+            if np.sum(post_tmax) >= 3:
                 # FDA NCA: use all post-Tmax points for λz estimation
                 t_term = t_valid[post_tmax]
                 c_term = c_valid[post_tmax]
             else:
                 # Fallback: not enough post-Tmax points, use last n_points
                 n_use = min(n_points, len(t_valid))
+                if n_use < 2:
+                    return None, None, None
                 t_term = t_valid[-n_use:]
                 c_term = c_valid[-n_use:]
         else:

src/tooluniverse/survival_tool.py

@@ -355,7 +355,16 @@ def _log_rank_test(self, arguments: Dict[str, Any]) -> Dict[str, Any]:
             # split is proportional (O==E for every stratum).  In this degenerate case
             # chi2 = 0 and p = 1.0 is the correct, well-defined answer: there is no
             # evidence of a difference between the two groups.
-            if abs(O1_total - E1_total) < 1e-10 and abs(O2_total - E2_total) < 1e-10:
+            # Use relative tolerance: floating-point sums over many events accumulate
+            # rounding errors proportional to total_events × machine epsilon (~2.2e-16),
+            # so an absolute threshold of 1e-10 is too tight for studies with thousands
+            # of events.  Relative tolerance 1e-8 is safe up to ~10^7 events.
+            total_events = O1_total + O2_total
+            rel_tol = 1e-8 * max(total_events, 1.0)
+            if (
+                abs(O1_total - E1_total) <= rel_tol
+                and abs(O2_total - E2_total) <= rel_tol
+            ):
                 chi2_stat = 0.0
                 p_value = 1.0
             else: