fix(scan): itemise change context only in the files a change touched

The PHP transfer replay (pre-push 13.3) spent its whole 7,200 s case in the
head side's evidence pass. A probe on one PMPro graph named the cause: not
compute -- context adds ~20% engine time -- but OUTPUT. Each evidence request
listed every method its scope reached, up to 1,845 rows; 200 requests into
the bundled Braintree SDK returned 52,831 rows, 9.4 MB, against 211 rows
without context, over 22,125 requests in a 3 GB container.

`affected_context` only ever attaches a location in a changed file, and
every other row was checked (real file under the root, usable extent) and
discarded. The scan now sends the changed files; the engine itemises those
and folds the rest into one `context_elsewhere` row carrying exactly what
that check reads. With no safe file list the engine itemises everything,
as before.

Equivalence: unit controls over valid, missing-file, no-extent and
outside-root cases, with and without a changed-file row; a real-Joern
control; and on PMPro itself, 2,440 relationships and zero gaps decided
identically from both forms, 52,831 rows down to 3,241 with the most
referenced file as the change. Probe: benchmarks/push/context_cost_probe.py.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01EmM1j46qgKMh3zqD1QMwEo

src/openultrasast/model/regions.py

@@ -194,6 +194,37 @@ AMBIGUOUS_CORRESPONDENCE = "ambiguous_line_correspondence"
 REACHABILITY_BOUNDED = "dynamic_external_or_depth_context_unresolved"
 
 
+def _elsewhere_unusable(row: Mapping[str, object], root: Path, checked: dict[str, bool]) -> bool:
+    """Whether a summarised set of locations holds one the itemised check would have rejected.
+
+    The same rejections as an itemised `context_method` row: a file that is not a real file under the root,
+    or an extent that is missing or inverted. Each file is checked once per call, because the summary for
+    every family of a region names the same files.
+    """
+    try:
+        if int(str(row.get("unusableExtents", 0))) > 0:
+            return True
+    except ValueError:
+        return True
+    paths = row.get("paths")
+    if not isinstance(paths, list):
+        return True
+    resolved_root = root.resolve()
+    for raw in paths:
+        text = str(raw)
+        if text not in checked:
+            try:
+                path = Path(text)
+                absolute = path if path.is_absolute() else root / path
+                absolute.resolve().relative_to(resolved_root)
+                checked[text] = absolute.is_file()
+            except (ValueError, OSError):
+                checked[text] = False
+        if not checked[text]:
+            return True
+    return False
+
+
 def affected_context(
     context: ChangeContext,
     questions: Sequence[QuestionIdentity],
@@ -224,6 +255,7 @@ def affected_context(
     # ambiguous anchor remains uncomparable.
     inherited_gap = any(not gap.endswith(":" + AMBIGUOUS_CORRESPONDENCE) for gap in gaps)
     renamed = {context.decode_path(r.base_path): context.decode_path(r.head_path) for r in context.renames}
+    checked: dict[str, bool] = {}
     for identity in questions:
         if execution_budget is not None and time.monotonic() >= execution_budget.deadline_monotonic:
             gaps.append("change_context_deadline_exhausted")
@@ -233,7 +265,9 @@ def affected_context(
         local_gaps = []
         if not any(r.get("kind") == "context_summary" for r in rows):
             local_gaps.append("context_projection_unavailable:contributor-scan")
-        elif not any(r.get("kind") == "context_method" for r in rows) and not any(r.get("kind") == "context_boundary" for r in rows):
+        elif not any(r.get("kind") in ("context_method", "context_elsewhere") for r in rows) and not any(
+            r.get("kind") == "context_boundary" for r in rows
+        ):
             # Only when the query offered no account of its own. A query that already said why its
             # scope is empty -- a named function this file does not define, most often -- has given
             # the specific reason, and adding a generic one on top buries it.
@@ -243,6 +277,13 @@ def affected_context(
             if row.get("kind") == "context_boundary":
                 reason = str(row.get("reason", "context_unresolved"))
                 (recorded_only if reason.startswith(REACHABILITY_BOUNDED) else local_gaps).append(reason)
+            if row.get("kind") == "context_elsewhere":
+                # Locations in files the change did not touch, summarised by the engine. None of them can be
+                # attached to a change, so the only thing to do with them is what was always done before they
+                # were discarded: check each names a real file under the root and had a usable extent.
+                if _elsewhere_unusable(row, root, checked):
+                    local_gaps.append("context_location_unavailable")
+                continue
             if row.get("kind") != "context_method":
                 continue
             path = Path(str(row.get("path", "")))

src/openultrasast/model/scan.py

@@ -340,6 +340,7 @@ def _scan_repository_impl(
             hooks,
             degradations=degradations,
             context_rows=context_rows if change_context is not None else None,
+            context_files=_changed_files(change_context) if change_context is not None else None,
             unit=unit,
         )
         per_kind["evidence"] = round(time.monotonic() - tier_started, 2)
@@ -878,6 +879,17 @@ def _sink_weights(
     return weights
 
 
+def _changed_files(context: ChangeContext) -> tuple[str, ...] | None:
+    """The head-side files a context location can be attached to, exactly the set `affected_context` reads.
+
+    ``None`` when one cannot be sent intact -- a newline in a path -- so the engine itemises everything.
+    """
+    changed = {context.decode_path(p) for p in context.changed_paths} | {context.decode_path(r.head_path) for r in context.renames}
+    if any("\n" in path or not path for path in changed):
+        return None
+    return tuple(sorted(changed))
+
+
 def _request_families(work: Sequence[tuple[str, ScanRegion, ArbiterSpec]]) -> dict[str, str]:
     """The family of each request id: the unit the sizer learns a cost rate for."""
     return {rid: str(getattr(spec, "family", "")) for rid, _region, spec in work}
@@ -1224,6 +1236,7 @@ def _evidence_pass(
     *,
     degradations: list[Mapping[str, object]] | None = None,
     context_rows: dict[QuestionIdentity, list[Mapping[str, object]]] | None = None,
+    context_files: tuple[str, ...] | None = None,
     unit: str = "repository",
 ) -> tuple[dict[tuple[str, str, str], Evidence], int]:
     """The evidence vector of every taint pair in ``work``, keyed by (path, function, family).
@@ -1234,10 +1247,13 @@ def _evidence_pass(
     grouped_taint = _grouped(list(work), hook_callbacks=hooks).get("taint", {})
     if not grouped_taint:
         return {}, 0
-    requests = {
-        rid: {**dict(params), "evidenceOnly": "true", **({"contextEvidence": "true"} if context_rows is not None else {})}
-        for rid, params in grouped_taint.items()
-    }
+    # With change context, the engine itemises locations only in the files the change touched and summarises
+    # the rest, because those are the only locations `affected_context` can attach. `None` means the change's
+    # files cannot be named safely, and then every location is itemised as before.
+    context = {"contextEvidence": "true"} if context_rows is not None else {}
+    if context and context_files is not None:
+        context.update({"contextFilter": "true", "contextPaths": "\n".join(context_files)})
+    requests = {rid: {**dict(params), "evidenceOnly": "true", **context} for rid, params in grouped_taint.items()}
     # Portioned by request count, because evidence has no per-request weight yet -- it is the pass that
     # PRODUCES the weights. It is cheap per request, 17,700 in 44 s, and it was still one all-or-nothing
     # batch: a 22,125-request evidence pass on a plugin's push transaction exceeded its ceiling under load and