fix(tools): parse-harden Do Better Legal ask against leaky fine-tune output

The dbn-legal-agent-v3 fine-tune (Track 1 / family) emits a labelled-prose
template — duplicate `answer:` prefixes, markdown-escaped underscores (`\_`),
and a trailing raw JSON blob — rather than the strict JSON the Azure/gpt-4o
path produces via response_format. decodeJsonObject() returned null on that
invalid JSON, so ask() dumped the entire raw blob into `answer`.

Fix at the parse layer (no upstream response_format change, to avoid fighting
the fine-tune's training):
- dbnToolsRepairJsonText(): strip fences, drop only invalid `\_`/`\*` escapes,
  then balanced-brace scan collecting every top-level {...} (longest first) to
  recover an appended JSON object. Shared by both gateways' decodeJsonObject(),
  so all JSON tools benefit.
- dbnToolsParseLabeledFields(): parse labelled-prose into real fields when no
  JSON decodes, tolerating escaped key names and collapsing duplicate prefixes.
- ask() null-fallback now builds clean structured fields from the parsed prose
  instead of dumping raw; what_remains_uncertain becomes a proper list.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

includes/LegalTools.php

@@ -268,12 +268,31 @@ PROMPT;
 
         $json = $gateway->decodeJsonObject($raw);
         if (!$json) {
+            // Some fine-tuned models emit a labelled-prose template instead of JSON.
+            // Parse those labels into the real fields rather than dumping the raw blob.
+            $fields = dbnToolsParseLabeledFields($raw, [
+                'answer', 'what_we_found', 'evidence_trail', 'what_remains_uncertain', 'next_practical_step',
+            ]);
+            $uncertain = trim((string)($fields['what_remains_uncertain'] ?? ''));
+            $uncertainList = $uncertain !== ''
+                ? array_values(array_filter(array_map(
+                    static fn(string $l): string => trim(ltrim($l, "-*• \t")),
+                    preg_split('/\r?\n/', $uncertain) ?: []
+                ), static fn(string $l): bool => $l !== ''))
+                : ['The response format could not be validated as structured JSON.'];
+            $cleanAnswer = trim((string)($fields['answer'] ?? ''));
+            if ($cleanAnswer === '') {
+                // No usable label — strip the trailing appended JSON blob from raw.
+                $cleanAnswer = trim((string)preg_replace('/\s*\{[\s\S]*$/', '', (string)preg_replace('/\\\\([_*])/', '$1', $raw)));
+            }
             $json = [
-                'answer' => $raw,
-                'what_we_found' => 'Azure returned a plain-text answer based on the retrieved excerpts.',
+                'answer' => $cleanAnswer !== '' ? $cleanAnswer : $raw,
+                'what_we_found' => trim((string)($fields['what_we_found'] ?? ''))
+                    ?: 'The model returned a plain-text answer based on the retrieved excerpts.',
                 'evidence_trail' => [],
-                'what_remains_uncertain' => ['The response format could not be validated as structured JSON.'],
-                'next_practical_step' => 'Review the source excerpts manually before relying on the answer.',
+                'what_remains_uncertain' => $uncertainList,
+                'next_practical_step' => trim((string)($fields['next_practical_step'] ?? ''))
+                    ?: 'Review the source excerpts manually before relying on the answer.',
             ];
         }