feat(prisma): eligible --max-chars, deterministic unavailable-fulltext path, qwen silence fix (refs P43)

src/cli/prisma.py

@@ -168,6 +168,12 @@ def screen(
 def eligible(
     name: str = typer.Argument(..., help="Project slug."),
     limit: int = typer.Option(0, "--limit", "-n"),
+    max_chars: int = typer.Option(
+        0,
+        "--max-chars",
+        help="Cap per-item markdown (full text) at N chars so bounded-"
+        "context local models keep the criteria in window. 0 = no cap.",
+    ),
     db: Path = typer.Option(_DEFAULT_ZOT, "--db"),
     storage: Path = typer.Option(_DEFAULT_STORAGE, "--storage"),
     hold: bool = typer.Option(True, "--hold/--no-hold"),
@@ -188,6 +194,7 @@ def eligible(
                 project,
                 storage_dir=storage,
                 limit=limit or None,
+                max_fulltext_chars=max_chars or None,
             )
 
     stats = _hold(_go, hold=hold)

src/prisma/eligibility.py

@@ -28,10 +28,17 @@ if TYPE_CHECKING:
 log = logging.getLogger(__name__)
 
 
-_SYSTEM = """\
-You are a systematic-review assistant running PRISMA stage 3
-(full-text eligibility). The paper's full text is attached; apply the
-criteria to the concrete methods, comparator, and outcomes reported.
+# One-sentence system message: longer system prompts make qwen models
+# served by Ollama return an EMPTY response (no text, no tool call,
+# finish=stop) — same quirk screen.py hit; the working guidance lives in
+# the cached first user message instead.
+_SYSTEM = "You are a systematic-review assistant running PRISMA stage 3 (full-text eligibility)."
+
+_GUIDANCE = """\
+## Task
+
+The paper's full text follows in the next message. Apply the criteria
+to the concrete methods, comparator, and outcomes reported.
 
 Call `record_eligibility_decision` exactly once. Cite the page or
 section you relied on in `rationale`. If the full text is cut off
@@ -47,7 +54,8 @@ def _build_call(project: Project, item_markdown: str) -> LLMCall:
             LLMMessage(
                 role="user",
                 content=(
-                    "## Inclusion / exclusion criteria\n\n"
+                    _GUIDANCE
+                    + "\n## Inclusion / exclusion criteria\n\n"
                     + project.criteria
                     + "\n\n## Exclusion reason codebook\n\n"
                     + project.reasons
@@ -92,6 +100,9 @@ def _queue(db: Db, project: str, limit: int | None) -> list[int]:
     return [r[0] for r in rows]
 
 
+_TRUNCATION_NOTE = "\n\n[... full text truncated to fit the model context ...]\n"
+
+
 def run(
     db: Db,
     provider: Provider,
@@ -100,8 +111,16 @@ def run(
     storage_dir: Path,
     limit: int | None = None,
     progress: bool = True,
+    max_fulltext_chars: int | None = None,
 ) -> dict[str, int]:
-    """Run eligibility screening over the queue."""
+    """Run eligibility screening over the queue.
+
+    ``max_fulltext_chars`` caps the per-item markdown so local models
+    with bounded context windows see the criteria and the front of the
+    paper rather than silently dropping the oldest tokens (Ollama
+    truncates from the head, which would discard the cached criteria
+    message before the paper's methods section).
+    """
     ids = _queue(db, project.name, limit)
     stats = {"assessed": 0, "include": 0, "exclude": 0, "uncertain": 0, "errors": 0}
 
@@ -109,6 +128,28 @@ def run(
         try:
             snap = load_item(db, zot_id, storage_dir)
             md = to_markdown(snap, include_fulltext=True)
+            if "\n# Full text\n" not in md:
+                # No retrievable full text — a deterministic PRISMA
+                # outcome, not an LLM question. (Small no-fulltext
+                # prompts also trip the qwen/Ollama empty-response
+                # quirk, so this path must not touch the model.)
+                apply_eligibility_decision(
+                    db,
+                    zot_id,
+                    project.name,
+                    {
+                        "decision": "uncertain",
+                        "reasons": ["unavailable-fulltext"],
+                        "rationale": "No full text retrieved by the fetch "
+                        "cascade; cannot assess eligibility.",
+                    },
+                )
+                db.commit()
+                stats["uncertain"] += 1
+                stats["assessed"] += 1
+                continue
+            if max_fulltext_chars and len(md) > max_fulltext_chars:
+                md = md[:max_fulltext_chars] + _TRUNCATION_NOTE
             call = _build_call(project, md)
             result = provider.complete(call)
             if not result.tool_calls: