fix(harness): ship coherent vendor harness cluster

Rebuild PR C to ship the vendor/photocraft harness as a coherent unit
(vendor implementation + vendor tests) instead of a per-file ahead/behind
mix that produced broken imports and 6+ test failures.

- Flip 36 harness impl/route/test files back to vendor (the per-file
  line-set policy misclassified user-rewritten files as 'behind').
- Flip core/models.py + src/tool_approval_scopes.py to vendor: vendor's
  harness (authority.py, turn_contract.py, clean_agent_preview.py) depends
  on vendor's older approval-scope API; upstream added
  CHAT_SESSION_APPROVAL_SIGNATURE_FIELD that vendor's core/models.py lacks.
- Add src/shutdown.py (new vendor infra) required by routes/chat_routes.py.
- Keep src/agent_tools/__init__.py at upstream (editor_tools registration
  lives in PR B).

Upstream-only security divergence preserved in C_upstream_delta.patch and
flagged for integration-time reconciliation.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

routes/chat_helpers.py

@@ -98,9 +98,7 @@ def clean_repeated_assistant_content(text: object) -> str:
         "",
         value,
     ).strip()
-    # No leading `\s*`: .strip() removes that whitespace anyway, and scanning
-    # it from every offset of a long whitespace run was quadratic (ReDoS).
-    value = re.sub(r"(?is)</\s*think\s*>\s*$", "", value).strip()
+    value = re.sub(r"(?is)\s*</\s*think\s*>\s*$", "", value).strip()
     return value
 
 _CASUAL_OPENING_RE = re.compile(
@@ -1326,17 +1324,13 @@ def _normalize_thinking(text: str) -> str:
 
     # Handle garbled <think> tags: reasoning text followed by <think> as separator
     # e.g. "The user said...I should respond.\n<think>Hey! What's up?"
-    # Linear form of `^([\s\S]+?)\n*<think>\s*([\s\S]*?)(?:</think>)?\s*$`:
-    # the lookbehind stops the lazy prefix re-scanning a newline run from every
-    # offset, and the optional trailing closer is dropped after the match
-    # instead of being retried at every body offset (both were quadratic).
     garbled = re.match(
-        r'^([\s\S]+?)(?<![\s\S]\n)\n*<think(?:ing)?>\s*([\s\S]*)$',
+        r'^([\s\S]+?)\n*<think(?:ing)?>\s*([\s\S]*?)(?:</think(?:ing)?>)?\s*$',
         text, re.IGNORECASE
     )
     if garbled:
         before = garbled.group(1).strip()
-        after = re.sub(r'</think(?:ing)?>$', '', garbled.group(2).rstrip(), flags=re.IGNORECASE).strip()
+        after = garbled.group(2).strip()
         # Only treat as garbled if the part before <think> looks like reasoning
         reasoning_starts = (
             'The user ', 'I need ', 'I should ', 'I will ',