spec R39: pin VHH thresholds from TNP source assign_flag()

Spec R39 said VHH PSH/PPC/PNC/totalCdrLength thresholds were "pinned
at M1 from TNP source `bin_*` functions". The values live in
oxpig/TNP `bin/TNP`'s `assign_flag()` function (not `bin_*`, but the
function the spec refers to). Transcribed verbatim with cohortSize 36:

  totalCdrLength: bidirectional, green 25-36, amber 20-24 OR 37-39, red <20 OR >39
  psh:            bidirectional, green 79.60-126.82, amber 73.40-79.59 OR 126.83-155.47, red <73.40 OR >155.47
  ppc:            high_bad, amber 0.39-1.18, red >1.18
  pnc:            high_bad, amber 1.47-1.88, red >1.88
  cdrh3Compactness: bidirectional, green 0.82-1.56, amber 0.56-0.81 OR 1.57-1.61, red <0.56 OR >1.61

`scoring.py` was falling back to Fv values for the four shared metrics
when mode == TNP. Now uses the VHH dict end-to-end. Citation in the
JSON points to the TNP repo + assign_flag() function name so the
source is traceable.

Minor edge-case correction to the cdrh3Compactness bands: spec text
listed green 0.82-1.57 but TNP source has amber starting at 1.57
(green effectively 0.82-1.56). Source-authoritative.

Author: xnacly

software/liabilities-script/data/thresholds.json

@@ -1,5 +1,5 @@
 {
-  "_comment": "Spec R39 amber/red threshold bands per Raybould 2019 Table 2 (Fv mode) and Gordon 2025 (VHH mode CDRH3 compactness). cohortSize is the calibration-cohort size from the source paper; preserved here so downstream users can weigh the threshold's statistical confidence. VHH PSH/PPC/PNC/totalCdrLength thresholds are deliberately omitted — they need M1 calibration from the TNP source paper; the scoring code falls back to Fv values until then.",
+  "_comment": "Spec R39 amber/red threshold bands. Fv mode: Raybould 2019 Table 2 (cohortSize 242). VHH mode: Gordon 2025 TNP `assign_flag()` function in oxpig/TNP repo (cohortSize 36). cohortSize preserved so downstream users can weigh statistical confidence.",
   "fv": {
     "cohortSize": 242,
     "citation": "Raybould et al., PNAS 2019, vol. 116 no. 10, 4025-4030 (Table 2)",
@@ -30,12 +30,32 @@
   },
   "vhh": {
     "cohortSize": 36,
-    "citation": "Gordon et al., bioRxiv 2025, doi:10.1101/2025.08.11.669635 (TNP)",
+    "citation": "Gordon et al., bioRxiv 2025, doi:10.1101/2025.08.11.669635 (TNP); thresholds verbatim from oxpig/TNP `bin/TNP` assign_flag()",
     "thresholds": {
+      "totalCdrLength": {
+        "bidirectional": true,
+        "green": [25, 36],
+        "amber_lo": [20, 24],
+        "amber_hi": [37, 39]
+      },
+      "psh": {
+        "bidirectional": true,
+        "green": [79.60, 126.82],
+        "amber_lo": [73.40, 79.59],
+        "amber_hi": [126.83, 155.47]
+      },
+      "ppc": {
+        "direction": "high_bad",
+        "amber": [0.39, 1.18]
+      },
+      "pnc": {
+        "direction": "high_bad",
+        "amber": [1.47, 1.88]
+      },
       "cdrh3Compactness": {
         "bidirectional": true,
-        "green": [0.82, 1.57],
-        "amber_lo": [0.56, 0.82],
+        "green": [0.82, 1.56],
+        "amber_lo": [0.56, 0.81],
         "amber_hi": [1.57, 1.61]
       }
     }

software/liabilities-script/scoring.py

@@ -3,15 +3,13 @@
 Thresholds live in `data/thresholds.json` per spec line 142 (`R39`) so the
 calibration source + cohortSize travel with the values. Loaded once at
 import time; runtime threshold checks read from the parsed dict.
-Composite scoring (R41) mirrors `compute_developability_score` in the
-sequence-liabilities block: fixability_weight × region_weight × exposure
-for motifs, plus per-mode flag bumps (red=8, amber=3, green=0), plus the
-Cys contributions (8 × exposed_extra + 20 × broken_canonical + 20 ×
-missing_canonical).
-
-VHH PSH/PPC/PNC/totalCdrLength thresholds are deliberately omitted from
-the JSON — they need M1 calibration against the TNP source paper; the
-code below falls back to Fv values for those four metrics until then.
+Fv values come from Raybould 2019 Table 2 (cohortSize 242). VHH values
+come from the TNP source paper's `assign_flag()` function in
+oxpig/TNP `bin/TNP` (cohortSize 36). Composite scoring (R41) mirrors
+`compute_developability_score` in the sequence-liabilities block:
+fixability_weight × region_weight × exposure for motifs, plus per-mode
+flag bumps (red=8, amber=3, green=0), plus the Cys contributions
+(8 × exposed_extra + 20 × broken_canonical + 20 × missing_canonical).
 """
 
 import json
@@ -103,8 +101,9 @@ def compute_flags(surface_metrics: dict) -> dict[str, str]:
     if mode == "TAP":
         thresholds = _FV_THRESHOLDS
     elif mode == "TNP":
-        # Fall back to Fv thresholds for the four shared metrics; add VHH-specific.
-        thresholds = {**_FV_THRESHOLDS, **_VHH_THRESHOLDS}
+        # VHH is now fully populated from TNP source `assign_flag()`; use
+        # its own thresholds end-to-end (no Fv fallback needed).
+        thresholds = _VHH_THRESHOLDS
     else:
         return {}