AI progress: honest milestone bar from closed JSON contract units

Replace the char-count estimate (EST_OUTPUT_CHARS) with ContractProgress:
points are earned only when real output units complete — first token (5),
each closed contract key / array item (5-90), root closed (95), stored (100).
Reasoning/thinking streams no longer fake progress: the first thinking
delta fires the first-token milestone ('the model is alive') and holds.
Surface the job's tone through ProcessingJobOut/JobInfo so the bar can
label itself 'Summarizing dry wit…'. Test rewritten to pin the new
honest-thinking semantics.

backend/shonar/services/ai/openai_compat.py

@@ -21,9 +21,12 @@ from shonar.services.ai._llm import (
 async def _collect_reply(resp: httpx.Response, ticker) -> str:
     """Reassemble the assistant reply from a (possibly streamed) response.
 
-    Feeds the running character count to [ticker] as deltas arrive so the
-    UI can show progress. Servers that ignored "stream": true answer with
-    a plain JSON body — that path is handled too."""
+    Feeds the accumulated content *prefix* to [ticker] as deltas arrive;
+    the ticker derives progress from completed JSON contract units only
+    (see ContractProgress). Reasoning deltas are real work but not
+    contract output — they stream before the JSON and intentionally do
+    not move the bar. Servers that ignored "stream": true answer with a
+    plain JSON body — that path is handled too."""
     import json as _json
 
     ctype = resp.headers.get("content-type", "")
@@ -33,11 +36,10 @@ async def _collect_reply(resp: httpx.Response, ticker) -> str:
             content = _json.loads(body)["choices"][0]["message"]["content"]
         except (ValueError, KeyError, IndexError, TypeError) as e:
             raise ProviderTransientError("LLM sent an unreadable reply.") from e
-        ticker(len(content or ""))
+        ticker(content or "")
         return content or ""
 
     parts: list[str] = []
-    total = 0
     async for line in resp.aiter_lines():
         if not line.startswith("data:"):
             continue
@@ -49,18 +51,23 @@ async def _collect_reply(resp: httpx.Response, ticker) -> str:
             delta = chunk["choices"][0].get("delta") or {}
             piece = delta.get("content") or ""
             # Thinking models (qwen3 on llama.cpp) stream a long
-            # reasoning_content channel BEFORE any content: counting only
-            # content left the progress bar frozen at "waiting for model"
-            # for the entire run, then jumping straight to done. Reasoning
-            # is real work — count it toward progress (never into the
-            # reply itself).
+            # reasoning_content channel BEFORE any content. It is real
+            # work but not contract output: counting it made the bar
+            # rocket to 99 while the JSON had not even started. The
+            # honest signal is the JSON itself completing.
             thinking = delta.get("reasoning_content") or ""
         except (ValueError, KeyError, IndexError, TypeError):
             continue  # keep-alives / usage chunks / odd frames: not content
-        if piece or thinking:
+        if piece:
             parts.append(piece)
-            total += len(piece) + len(thinking)
-            ticker(total)
+            ticker("".join(parts))
+        elif thinking:
+            # First thinking delta proves the model started working:
+            # the ticker's 5% "first token" milestone fires once (an
+            # empty prefix is a no-op after that). The thinking phase
+            # itself earns no further points — it is not contract
+            # output — but "the model is alive" is real information.
+            ticker("")
     content = "".join(parts)
     if not content:
         raise ProviderTransientError("LLM sent an unreadable reply.")