fix(client): the Atlas was replaying a cache from a build that no longer existed (T-1239)

Ferrath's Global map drew no rivers at native resolution: 375 courses arrived
and 0 were drawn. The report suspected the D-261 length cull or the water
truncation. Both were innocent, and so was the renderer.

The client served the canvas from its own disk cache (T-1183). Every payload
for GJ820Bc predated T-1237 (4e503c356), which replaced one-course-per-D8-hop
with one-course-per-river -- so the map was drawing 375 hop fragments whose
longest run was 106 km, all of them under D-261's read-as-a-line floor. Same
build, same scenario, same 3440x1440, cache the only difference:

  stale   courses=375  runs=180  longest=6.0px  (~106 km)   drawn=0
  cold    courses=73   runs=23   longest=93.2px (~1,644 km) drawn=18

It looked resolution-dependent because it wasn't a resolution at all: 960x540
resolves to an 814x407 canvas, a key never cached, so it missed and re-derived
correctly. 3440x1440 resolves to 1080x540, which had an entry from 2026-08-06.
During the stale capture the server logged no course production whatsoever --
the canvas never came from it.

The cache's only invalidation signal is project.yaml's version, and 4e503c356
changed how canvases are generated without touching it, so hop-shaped entries
stayed valid. All 13 stale entries are stamped 0.4.5. 0.4.6 forces them to miss;
that, not clearing a local directory, is what repairs a player's Atlas.

The harness let this hide for eight days, in two ways now fixed. It ran against
the developer's persistent user:// cache, so a capture could render a canvas
built by a build that no longer existed -- and any golden shot in that window
silently inherited it; user:// is now isolated per run. And it sent server
stderr to /dev/null via an already-unlinked mktemp file, so no tracing from a
capture was ever reachable; the log now lives at .cache/visual-server.log.

The capture readout gained runs= and longest= between courses= and drawn=,
because "375 arrived, 0 drawn" is not one fact but three stages, and telling
them apart is what turned a guess between two suspects into a measurement.

Follow-ups filed: T-1241 (current_schema_version() returns its ?.?.? fallback in
an exported build, so a shipped game never invalidates on version at all) and
T-1242 (nothing enforces the generation-change/version-bump pairing -- this is
the fourth bump forced after the fact).

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

server/src/atlas/step_canvas.rs

@@ -781,6 +781,21 @@ fn invent_courses_for_canvas(
         });
     }
 
+    // T-1239: the network is built once per body on the fixed GRID_W x GRID_H
+    // working grid, so these counts MUST NOT vary with canvas extent. They were
+    // observed to (73 courses at one viewport, 375 at another), which is only
+    // visible if the pre-crop network size is logged next to the post-crop wire
+    // size — the client can only ever see the latter.
+    tracing::info!(
+        river_cells = river_network.river_cells.len(),
+        paths = paths.len(),
+        courses = courses.len(),
+        longest_path_cells = paths.iter().map(|p| p.cells.len()).max().unwrap_or(0),
+        ta_w = ta.w,
+        ta_h = ta.h,
+        "step canvas: river course production"
+    );
+
     courses
 }