reach godot parse-sweep / cold-parse, reach visual diff / blank-check /
thumbnail. Five scripts retired, and the callers rewired — tests/run-visual
invoked three of them by path at four sites, which is a wider blast radius than
the make targets were.
The godot pair were grep pipelines encoding five hard-won lessons as comments
nobody could test. They are Python filters now, with the reasons attached, and
the engine invocation is a guarded exec. Verified on the real client: 229
scripts, clean.
Their three not-ok states stay distinct, because only one is a verdict about
the code. An engine that crashed or is missing is not a parse failure —
reporting it as one blames the tree for a broken toolchain. A sweep that
emitted no completion marker checked nothing, and zero errors from a check that
never ran reads as clean, which is the false-green the sweep exists to close.
The deliberate asymmetry between the two checks is preserved and documented:
cold-parse filters "Cannot infer the type", the sweep does not, because that
suppression is why cold-parse stayed silent about a helper that genuinely does
not parse.
All three visual scripts carried the same root bug as validate-checklist:
Path(__file__).parent.parent, correct at tooling/ and two levels too deep at
tooling/domains/visual. Fixed during the move rather than after, having learned
that it fails silently — paths resolve to nothing, the work appears to have
nothing to do, and the tool reports success. Three domains now where that would
have shipped a false pass.
Two bugs my own transformation introduced, both found by running rather than
reading. Multi-line print(..., file=sys.stderr) became console.event(...,
file=sys.stderr), and console puts unknown kwargs into the payload — a file
object would have reached json.dumps at the exact moment something was already
being reported as an error. And the replacement script wrote escaped quotes
into three files. Mechanical transformations need mechanical verification.
sys.exit removed from four sites: a service must not end the process.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
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>
run-visual verified only that the captured PNG was non-empty AS A FILE. A
blank screen is a perfectly valid ~19 KB PNG, so it passed — and once a blank
capture had been recorded as a golden, every later blank capture matched it at
0.0% and the scenario PASSED. atlas_GJ338Bd_Block and atlas_GJ445c-m1_Chunk
sat green against blank goldens while the suite's other 30 scenarios failed.
That is the worst kind of test result: indistinguishable from success, and
load-bearing for exactly the work it fails to cover. e024cfb3f recorded this
same failure once already ("the Atlas Global goldens have been measuring
nothing"); it recurred because nothing checked the property, only the file.
tooling/visual-blank-check measures the share of the frame taken by its single
most common colour. On this project's real captures the classes are far apart:
Global (real world map) 38.7% modal
Region (flat colour wash) 7.2% modal <- dither; least uniform of all
District 45.4% modal
Block / Chunk / Quarter 92.9-94.6% modal <- nothing drawn
Nothing falls between 45% and 93%, so the 0.85 default sits in open space
rather than being tuned against a boundary case. Deliberately NOT an aesthetic
judgement: the Region wash is a real product gap (T-1213) and scores 7.2%,
comfortably "content". The question is only whether a world reached the
screen.
Wired into both paths, and the update path is the one that matters — refusing
to RECORD a blank golden is what stops the trap being re-armed. Ad-hoc
--screenshot only warns, since capturing a rung that renders nothing is a
legitimate thing to want to do; that is how the empty deep rungs were found.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Two faults, one fix.
Captures were stealing the desktop. Only the golden path even tried to go
offscreen, via xvfb-run -- which is not installed here, so it took the
"using visible window" fallback; --screenshot and --movie never wrapped at
all. Every capture opened a Godot window on the machine Jeroen is working
and gaming on. Now a single wrapper covers all three paths.
gamescope, not the alternatives, for two independent reasons. It renders on
the real AMD GPU, and the goldens are pinned to this box's Mesa/AMD output
(T-1121 -- they do not port across rendering stacks), so xvfb-run's llvmpipe
would shift every pixel: offscreen must not silently mean a different
renderer. And it is the only installed option that lets the output size be
set. cage is also present and also GPU-backed, but it is a kiosk compositor
and forces its client to the headless output's default -- measured, a
960x540 request produced a 1280x720 PNG. A wrapper that quietly changes
resolution is worse than none here.
Worse, because resolution is not cosmetic on this map. D-255's extent
inversion makes the shorter viewport axis span exactly one cell of the rung,
so the viewport decides how much world a rung shows and at what cell count
-- a small capture is a DIFFERENT map, not a scaled one. 960x540 was also
16:9, so it never exercised the ultrawide aspect added in 21e263d0a, which
is the aspect actually in daily use. Raised to the panel's native 3440x1440.
It paid for itself immediately: at native, Ferrath Global reports
courses=375 drawn=0 -- every river culled, where the same build drew them at
960x540. Filed as T-1239. That is exactly the class of bug a too-small
capture hides.
Goldens are NOT regenerated here. They are stale across 15 commits already,
and blessing the current look before it has been reviewed is the trap this
suite just spent a day proving.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The visual gate had stopped measuring anything: 30 of 32 scenarios
failed, and the two that passed were the worst result of the lot.
Deleted the 15 top-down scenarios (fog, HUD, dialogue, NPC, minimap,
cursor) and their goldens. They all failed at a near-uniform ~12%, and
that uniformity across unrelated scenes is one global cause -- the
ultrawide UI stretch moved every element. They cover the renderer the
cascade freezes until Phase 5, which will need its own tests anyway, so
re-baselining would only have blessed a deprecated layer nobody is
reviewing. Jeroen's call.
The remaining problem was the goldens that PASSED. atlas_GJ338Bd_Block
and atlas_GJ445c-m1_Chunk matched at 0.0% because capture and golden
were both blank -- the same "goldens have been measuring nothing" trap
e024cfb3f caught at Global, still live at the bottom of the ladder. The
cause is that every below-Global golden descends at jump_to(ZERO), and
world-metre zero is merely the origin of the region grid, not anywhere
chosen. So _setup_atlas_golden_shot now takes an optional world_center
(default ZERO -- existing goldens are untouched), and a new
atlas_GJ820Bc_land_* set walks Region through Chunk at ONE land point,
so the rungs can be read as a descent instead of five unrelated frames.
aliveness_probe prints the placement's world metres alongside its pixel
and survey cell, since that is the coordinate the Atlas actually
navigates in.
Recorded because it will be asked again: the ladder is anchored via a
CityPlacement, but that is a match record -- a pixel, an archetype, an
orientation -- not built geography. No settlement exists anywhere yet
and none is due before T-1207, so the empty deep rungs are the expected
state. What the ladder judges is the nature layer.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Add live server lifecycle to tests/run-visual (start/stop server per
scenario, parse LISTENING:{port}). Add MessagePack snapshot replay to
visual_capture.gd via Protocol.decode_snapshot() — exercises the full
client pipeline from wire bytes to rendered fog. Three replay scenarios
(hub_spawn, fog_theater, hub_after_movement) plus one live scenario
(fog_live_hub). Add gen_gauntlet_fixtures.rs to produce .msgpack fixtures
from the Gauntlet test world. Add max_diff_pct threshold to visual-diff.
Makefile: add fixtures-gauntlet target, fix build-client double-import,
preserve .godot cache in clean.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
jpmschweitzerandClaude Opus 5