No reviewers
Labels
No labels
area/accessibility
area/audio
area/input
area/install
area/menu
area/renderer
area/saves
area/tooling
area/upscaling
beanstalk: approved
beanstalk: failed
beanstalk: needs approval
beanstalk: running
beanstalk: tested
ci: approved
good first issue
gpu/amd
gpu/igpu
gpu/intel
gpu/nvidia
help wanted
kind
bug
kind
chore
kind
docs
kind
feature
kind
performance
kind
question
needs measurements
platform/linux
platform/macos
platform/steam-deck
platform/windows
preview
priority
critical
priority
high
priority
low
regression
risk
high
risk
low
risk
medium
status
blocked
status
confirmed
status
duplicate
status
invalid
status
needs-info
status
wontfix
ci: approved
No milestone
No project
No assignees
3 participants
Notifications
Due date
No due date set.
Dependencies
No dependencies set
Reference
MagicBeanDev/Longshot!1466
Loading…
Reference in a new issue
No description provided.
Delete branch "mooki/Longshot:docs/guide-graphics-retention-1309"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Type: docs · Risk: low · Tested on: Linux x86-64, no GPU (the picture is drawn in memory by SDL2's software renderer; the checks need no game) · Needs: any PC with a browser (the site's Settings page), and a maintainer's look at the pictures
TL;DR
The guide-screenshot check is red on
mainitself, so every pull request inherits the failure: #1309, #1362 and #1464 changed seven source files that the Graphics guide picture is recorded against, and no merge carried a review. I did the review: the Graphics view rendered from main and from the last commit the check passed on differ only in the build's version text (976 of 921,600 pixels). So the picture is kept, and this pull request updates its record (seven hashes, ten headers that were never listed, a dated note), which makes the check pass again. Two questions for maintainers at the end would stop it going red for changes that cannot touch the picture.Breakdown
b50f49771(#1383), whose own change advanced this record.manifest.json,settings-graphics.png'sreviewed_source: seven hashes advance (menu_i18n.h,vk_framegen.h,vk_fsr3fi.h,vk_internal.h,vk_menu.cpp,vk_reflex.h,vk_texture.h), ten headers join the list (the onesdevice_policy.hbrings in since #1309), and the review revision (the SHA-256 of the sorted file/hash lines) is recomputed. Insource-review-fullscreen-monitor.md, a dated "Retention after #1309" section says what was compared and found.b50f49771,67efc4d3e(#1309),c14641a6f(#1362) and597c5a4e0(#1464) compared pixel by pixel and looked at; the two other steps of the workflow's check (14 and 4 tests); the whole suite here and on main, by name; the description checker.Why
mainfails its own Guide images check, and every pull request shows the same red status that has nothing to do with it. The Graphics picture in the settings guide is recorded against the source files it is drawn from (reviewed_source.files); when one changes,scripts/check_guide_screenshots.py --review-sourcesprintsREVIEW REQUIREDand exits 1 until someone compares the picture with a rendering of the new source and either retains it or recaptures it (docs/images/guide/README.md, "Capture and maintenance", step 5). #1383 did that at 16:02 UTC today (b50f49771); three minutes later #1309 mergedvk_internal.handvk_reflex.hchanges, at 18:53 UTC #1362 merged changes tovk_internal.h,vk_menu.cppandmenu_i18n.h, and at 20:00 UTC #1464 merged changes tovk_framegen.h,vk_fsr3fi.handvk_texture.h. #1144 (16 Nov) was the same repair for one file and was closed as superseded on 17 Nov ("Reopen if the check goes red for the same reason again"); this is that case. The trackers (issues and pull requests, open and closed) were searched for "guide screenshots", "guide images", "review-sources" and "retention": no open issue or pull request is about this failure; #1144 and #1383 are the earlier ones.What changed
The Graphics picture is retained. Its review record now names the content that was compared, and the review is written down.
docs/images/guide/manifest.json,settings-graphics.png→reviewed_source:revisionis nowsha256:006bc2510d667d862da9190b16a64059c80e8a2b8d824f213cd4e1bd92f1439a(it wassha256:275a1444…);fileshas the new hashes ofgpu/vk/menu_i18n.h,gpu/vk/vk_framegen.h,gpu/vk/vk_fsr3fi.h,gpu/vk/vk_internal.h,gpu/vk/vk_menu.cpp,gpu/vk/vk_reflex.h,gpu/vk/vk_texture.hand ten new entries (astc_bc_policy.h,device_policy.h,fallback_size_policy.h,gpu_clock.h,mesh_mode.h,mirror_budget.h,retired_give_back.h,target_pool.h,texture_budget.h,vram_compact.h, all ingpu/vk/): 77 files instead of 67. The helper's compiler pass finds them because #1309'svk_internal.hnow includesdevice_policy.h; they were not recorded before. Nothing else in the file changes (the JSON round-trips exactly, so the diff is only those lines).docs/images/guide/source-review-fullscreen-monitor.md: a "Retention after #1309 (2026-11-17)" section (+30 lines): what was rendered and compared, what each changed input is (#1309: the include and three members inDevice, a comment that lost a switch name, a comment invk_reflex.h; #1362:swap_slots,kPresentSlotsMax,present_slots_for, and one sentence added to the hints of the two values of the Frame pacing row and their translations, a row on the Debug tab, not on the Graphics tab; #1464: the declarationsframegen_hud_before, ahud_maskmember andtexture_alive), why the ten headers join, and what is not claimed.settings-graphics.png(sha256 3531f6ab…),captured_source,fixture_files,capture, the entry'sreview.notes,docs/graphics-and-controls.mdand its alt text and caption ("development build54d25c79"), the checker, the workflow and the README.Measurements
Verdict: the check fails on main (exit 1, 7 lines) and passes with this PR (exit 0). The Graphics view rendered from main differs from the render of the last green commit (
b50f49771, before #1309) in 976 of 921,600 pixels, all in the banner's version text, so the picture is retained. Nothing else to measure in the game: a record and a note change no code.Hardware: Linux 7.2.15, Intel i9-13900K, 93 GB; an RTX 4090 is in the PC but not used (nothing here opens a GPU)
Scene: none; the Graphics view of the menu fixture, 1280x720, rendered by SDL2's software renderer
Method:
scripts/check_guide_screenshots.py --review-sourceson six commits; the capture helper's renders of five, compared pixel by pixel (results by commit under "How it was tested", 2 and 4); one run each (all deterministic: main rendered twice gave byte-identical PNGs andmenu-capture.json)How to reproduce: below, under "How it was tested"
dcb7b47c29fd31222dscripts/check_guide_screenshots.py --review-sourcesREVIEW REQUIREDformenu_i18n.h,vk_framegen.h,vk_fsr3fi.h,vk_internal.h,vk_menu.cpp,vk_reflex.h,vk_texture.htests/test_guide_screenshots.pytests/test_launcher_capture.pytests/run_all.py597c5a4e0)597c5a4e0); the same 30 fail, by nameb50f49771(94a3b352…) in 976 pixels, all in the banner's version textsettings-graphics.png, 3531f6ab…)The check rows are from
dcb7b47c2and9fd31222d; the whole-suite and render rows are from597c5a4e0and this PR on it (27a147927). Main has moved since (#1454: the PR template, the description checker and their test; #718:vk_draw.cppand a test), with no recorded input and no guide file, and the last rebase left the patch unchanged (git range-diff=): the check and the two guide tests were rerun on the new base, the suite and the renders carried, not rerun.The check on main, whole, as the program printed it (
python3 -I -B scripts/check_guide_screenshots.py --review-sources, exit 1):On this PR's head (exit 0; the six Launcher pictures are
NOT CHECKEDhere and on main alike: there is no Launcher checkout in CI):Pictures
PICTURE: GuideFix-graphics-renders.jpg - Left: the Graphics view rendered by the capture helper from
b50f49771(main before #1309, the last commit the check passed on); middle: rendered from597c5a4e0(main with #1309, #1362 and #1464); right: the middle picture dimmed, with the 976 pixels that differ from the left one in red. The same tabs, rows (Quality preset Custom, Aspect ratio 16:9, Output resolution 2048x1152 (reduced), Custom output size None, Profile Docked, Upscaler Game's own, Render resolution Auto (dynamic), Sharpness 0%, Frame generation Off, Game frame rate cap 60 (the console's)), help panel and button prompts in both; the only red is in the top left banner.PICTURE: GuideFix-graphics-banner.jpg - The banner enlarged, the same three pictures: "Longshot nightly-404-5d8293da-2-gb50f49771" on the left render, "Longshot nightly-407-67efc4d3-7-g597c5a4e0" on the middle one, and the red pixels (x 126 to 260, y 25 to 38) of the difference at the bottom: the build's version text is the whole difference.
Flags and settings
None.
How it was tested
Everything here was run by an AI agent (Claude Sonnet 5.5) working for me; the commands are given so anyone can rerun them. The commit hashes are the ones on the forge. I started on main
c14641a6f(17 Nov 2026, 13:35 US Central time) with four changed inputs; main then moved three more times, to7bb06359c(#1453, no recorded input), to597c5a4e0(#1464, three more recorded inputs: the review was extended and the commit amended) and todcb7b47c2(#1454 and #718, no recorded input). The renders and the whole-suite results below are from597c5a4e0and this PR on it; the check and the two guide tests were rerun ondcb7b47c2and this head9fd31222d(see the history under Testing).1. Environment
Linux 7.2.15 (Arch, x86-64), Intel i9-13900K, 93 GB; the load average was 11 at the start and 20 to 30 later (other jobs on the PC; the checks and renders ran in the idle scheduling class). clang 23.1.3, SDL2 2.32.74 (this is
sdl2-compaton SDL3 3.4.18), libpng 1.6.59, Vulkan headers from/usr/include(headers only, no Vulkan runtime or device), Python 3.14.8 for the checker and 3.12.16 for the suite. The manifest records clang 23.1.6, SDL2 2.32.10 and libpng 1.6.57 for the published PNG: a different toolchain, which matters below. No GPU, window or game was started.2. What the check does, and where it first fails
.forgejo/workflows/guide-screenshots.yml: it runsscripts/check_guide_screenshots.py --review-sources, thentests/test_guide_screenshots.pyandtests/test_launcher_capture.py, on every pull request and every push to main. It never builds the fixture or opens a game file.scripts/check_guide_screenshots.py: for each image it checks the PNG's size and SHA-256, then with--review-sourcesthe SHA-256 of every file inreviewed_source.filesandfixture_filesagainst the working tree. A difference is aREVIEW REQUIREDline; a missing Launcher checkout isNOT CHECKED, which does not fail. It compares hashes only: it never looks at a picture.git fetch forge main,git worktree add --detach <dir> <commit>) at597c5a4e0andc14641a6f, and ofb50f49771,67efc4d3eand7fd3e2ace, and ran the check in each: the table below.b50f49771(git diff b50f49771 597c5a4e0 -- <the seven files>): #1309 invk_internal.h(#include "device_policy.h";Devicegetsdriver,facts,policy; a comment loses a switch name) andvk_reflex.h(a comment); #1362 invk_internal.h(swap_slots,kPresentSlotsMax,present_slots_for),vk_menu.cpp(thehn_pacehint strings) andmenu_i18n.h(those strings' translations); #1464 invk_framegen.h(framegen_hud_before),vk_fsr3fi.h(ahud_maskmember) andvk_texture.h(texture_alive): declarations and comments, which the menu fixture neither calls nor draws.hn_pacebelongs to the Frame pacing row, which is declared withTAB_DEBUG(vk_menu.cppline 1777): not a row of the Graphics tab this picture shows.What the check said on each commit (
python3 -I -B scripts/check_guide_screenshots.py --review-sources, seven worktrees):REVIEW REQUIREDlinesb50f4977167efc4d3evk_internal.h,vk_reflex.h7fd3e2acec14641a6fmenu_i18n.h,vk_internal.h,vk_menu.cpp,vk_reflex.h597c5a4e0menu_i18n.h,vk_framegen.h,vk_fsr3fi.h,vk_internal.h,vk_menu.cpp,vk_reflex.h,vk_texture.hdcb7b47c29fd31222ddcb7b47c2(27a147927on597c5a4e0gave the same)3. The renders: the project's UI-only capture helper
The helper is
scripts/screenshots/capture_menu.py, with the fixturescripts/screenshots/menu.cpp(both on the forge; the guide README's "UI-only settings fixture" describes them). It compiles the realgpu/vk/vk_menu.cpp, the translation code and the bundled Dear ImGui against the fixture, calls the menu's draw function with fixed illustrative settings, and writes the result with SDL2's software renderer: 1280x720, Medium menu scale, English, the bundled Libertinus fonts over a solid RGB 16,18,21 background, explicit 16:9, the compiledOptionsdefaults, no GPU capabilities, no saved options, no game; the environment has everyHOS_*,RX_*andLONGSHOT_*variable removed andHOS_DISPLAY_MODES=1280x720, and it runs in an empty temporary folder. It refuses a source with uncommitted capture inputs. The banner's text isgit describeof the source plus " (UI fixture)", from-DLONGSHOT_VERSION.I ran main's copy of the helper, so the fixture is the same for every render and only the source differs:
once each for
b50f49771,67efc4d3e,c14641a6f,597c5a4e0and7fd3e2ace, and once more forc14641a6finto a second folder: the second render's three PNGs (Graphics, Audio, Controls) andmenu-capture.jsonare byte for byte equal to the first's, so a render is repeatable on this PC. The image's recorded capture source,54d25c79, is no longer on the forge (#1277's head moved: the pull request's head isc0dfffe73now), so the comparison base is the last commit the check passed on;7fd3e2ace, #1277's squash on main, stands in for the capture source below.4. The comparison: every number from the files
I compared the renders with a small script (PIL and NumPy only for decoding and counting; it is attached below): for each pair, the number of pixels where any channel differs, the largest channel difference, and the bounding box of those pixels. Where a pair differs in several places (the published PNG against a render), I grouped the differing pixels into boxes of 8 by 8 pixel blocks within two blocks of each other with a longer version of the same script, to say where.
compare_renders.py
Results (all pairs rendered on this PC with main's helper; PNG sha256:
b50f4977194a3b352…,67efc4d3ecf24f221…,c14641a6f2e25da3d…,597c5a4e0a7d8f675…,7fd3e2ace6268e1f0…, published 3531f6ab…):b50f49771and597c5a4e0b50f49771and67efc4d3e(#1309 alone)nightly-407-67efc4d3, shown as "Nightly 407 ·67efc4d3")67efc4d3eandc14641a6f(#1362 and #1359)c14641a6fand597c5a4e0(#1464)b50f49771andc14641a6f7fd3e2ace(#1277) and597c5a4e0597c5a4e0render7fd3e2acerenderThe last two rows are why the comparison is between renders. The PNG on main differs from a render of #1277's own merge commit on this PC's toolchain in 60,473 pixels (the rows and the help are drawn a pixel differently; the labels, values and layout read the same), so a render of any commit differs from the PNG by about as much, and that difference is not from #1309, #1362 or #1464. I did not find the cause by experiment: the toolchain here differs from the recorded one in the compiler, in SDL2 (an sdl2-compat on SDL3 against 2.32.10) and in libpng, and I have not rendered with the recorded versions. It is labelled as inferred.
Looking at the pictures. I opened each render at full size (
b50f49771,67efc4d3e,c14641a6f,597c5a4e0,7fd3e2ace) and the published PNG, and compared them with the pixel boxes above: tabs Graphics, Audio, Gameplay, Controls, Debug; Quality preset selected on Custom; the rows Aspect ratio (16:9 or Monitor's shape), Output resolution 2048x1152 (reduced), Custom output size None, Profile (Docked or Handheld), Upscaler Game's own, Render resolution Auto (dynamic), Sharpness 0%, then "Frame generation and latency": Frame generation Off, Game frame rate cap 60 (the console's); the help panel "Quality preset: Sets the scaling rows and Shadow resolution for a GPU with that much memory. The Upscaler and Frame generation stay yours." with its note; and the prompts Choose, Change, Compact (RShift), Tabs (Q E) and Close (Esc). They are the same in all of them, and they are what the guide's alt text names ("Quality preset selected, setting rows on the left, help on the right, and keyboard navigation along the bottom"). Only the banner's version text differs.5. Advancing the record
menu-capture.jsonwithsource_files: the repository files the compiler used for that render (-MM) plus the Libertinus fonts, each with its SHA-256. For597c5a4e0that is 77 files: the 67 of the old record and ten more; #1464 adds none.reviewed_source.files(asserting that each hash equals the file's own SHA-256 and that the manifest is injson.dumps(indent=2)form, so nothing else moves) and setrevisionto the SHA-256 of the sortedpath=hashlines, each with a final newline, as the README says. I checked the formula first: applied to the old 67 entries it gives exactly the recordedsha256:275a1444….b50f49771to67efc4d3e:vk_internal.h,vk_reflex.hchanged, the ten added;67efc4d3etoc14641a6f:menu_i18n.h,vk_internal.h,vk_menu.cppchanged, none added;c14641a6fto597c5a4e0:vk_framegen.h,vk_fsr3fi.h,vk_texture.hchanged, none added.source-review-fullscreen-monitor.mdin the form of its earlier sections, dated.git interpret-trailers --parselists both).6. The checks after the change
python3 -I -B scripts/check_guide_screenshots.py --review-sourceson27a147927(on597c5a4e0) and on9fd31222d(ondcb7b47c2): exit 0 both times (lines above).tests/test_guide_screenshots.py(14 tests) andtests/test_launcher_capture.py(4 tests), through the workbench's idle gate (a wrapper that runs CPU-only jobs in the idle scheduling class, in one of three test slots, behind any timing hold; https://github.com/mooki000/ls-workbench): bothOK.python -I -B tests/run_all.py --jobs 8 --junit <file>, through the same gate, in a worktree of this PR on597c5a4e0and in one of main597c5a4e0: 614 passed, 30 failed, 82 skipped on each (108 s and 132 s); the same 30 fail on both by name (below). The 82 skipped are the same on both: 65 need the Vulkan headers or the SPIR-V headers, which this run's tool search did not find (it prints "not found here: vulkan_headers" and "not found here: spirv_headers"); the rest need Windows, the game's dump or a GPU. I compared the two JUnit files by test name: no test changed status. Not rerun after the later rebases (patch identical, no recorded input in main's new commits).python -I -B scripts/check_pr_description.py --file <this body> --diff <git diff --src-prefix=a/ --dst-prefix=b/ main..HEAD> --title <title>: run on this text (result under "Testing").The 30 failures on both (main
597c5a4e0and this PR on it), named:test_bug_report_defaults,test_cpu_report,test_crt_options,test_damage_debug_menu,test_display_startup_failure,test_dlss_feature_size,test_failed_compile_keeps_objects,test_framegen_preset,test_guest_imported,test_hud_layout_row,test_invalidate_vcore,test_linux_preflight,test_linux_pressure,test_linux_update,test_mirror_budget,test_missing_nup,test_mods_size_table,test_pending_guest_import,test_pipeline_play_cache,test_present_thread,test_profiling_mode,test_readback_timeout_source,test_rumble_menu,test_run_all_changed,test_slow_frames,test_target_image_pool,test_upscaler_round_trip,test_upscaler_rows,test_vcore_escape,test_walk_key. None was looked into: a change of one JSON file and one Markdown file cannot reach them, and they fail on main the same way.7. Overlaps, and the tracker
pulls/N/files, 50 a page) were scanned for the guide's files (docs/images/guide/*, the checker, the workflow, the capture helper and fixture) and for the four headers changed when the scan was made (#1309's and #1362's). Of those that carry a guide edit, I fetched the heads (git fetch forge +refs/pull/N/head:refs/remotes/pr/N) and rangit merge-tree --write-tree --name-onlyof each head against forge main and against this branch; the trial merges of the five of ours were repeated on this head; the result and what it means is under Overlaps. One pull request, #768, returned no file list on the first pass; its diff holds none of those files.8. How to rerun it
(
capture_menu.pyis run from the checkout that has this PR, so the fixture is the same for both renders; it needs a C++20 compiler, pkg-config, SDL2 and libpng development files and the Vulkan headers.)9. What was not tested
See "Not tested" under Testing.
Testing
Tested: on Linux x86-64 (kernel 7.2.15), clang 23.1.3, SDL2 2.32.74 (sdl2-compat), libpng 1.6.59, no GPU, on 17 Nov 2026, base main
dcb7b47c2, head9fd31222d. For this docs-only change "tested on my own PC" is the checks I ran, and they are these: the Guide images check on seven commits (main fails with 7 lines, this PR passes); the capture helper's renders fromb50f49771,67efc4d3e,c14641a6f,597c5a4e0and7fd3e2ace, compared by pixel and looked at;tests/test_guide_screenshots.py(14) andtests/test_launcher_capture.py(4) pass; the wholetests/run_all.pyon this PR and on main, both on597c5a4e0: 614 passed, 30 failed, 82 skipped on both, the same 30 by name (listed above; none is this PR's).check_pr_description.py --file <this body> --diff <the diff>: passes (run on the final text, with the checker ofdcb7b47c2, before the handover).History: the commit was first made on
c14641a6f(9db835e58, four inputs), rebased onto7bb06359c(7586502e3, patch identical; #1453 touches no recorded input), then onto597c5a4e0, where #1464 changed three more recorded inputs: the review was extended (a render of597c5a4e0, its comparison, three more hashes, the note) and the commit amended into27a147927. Last, rebased ontodcb7b47c2(#1454 and #718: neither touches a recorded input or a guide file): head9fd31222d, patch identical to27a147927(git range-diff=). Rerun there: the check (exit 0; on maindcb7b47c2it still says the same 7 lines) and the two guide tests (14 and 4 OK). Carried, not rerun: the renders, the pixel comparisons and the whole suite (from597c5a4e0). Windows: not applicable. Game: not applicable (docs only).Not tested: the website build and
website/tests/guide-screenshots.test.ts(the site's packages are not installed on this PC; the PNG is unchanged, so what the site copies is unchanged); Windows and Intel (the picture is the UI-only fixture, not a live Vulkan frame, as its caption says; the Windows-target syntax check does not apply: no C or C++, Python or workflow file changes); a render with the toolchain recorded for the PNG (clang 23.1.6, SDL2 2.32.10, libpng 1.6.57); rows below the first Graphics view and the other tabs (the picture does not show them, and this review does not look at them); no game was run, none is needed for a record and a note.Must be tested:
website/README.mdsays, at desktop width and at 360 to 390 px, in both themes, expecting the Graphics picture to load from the site's own files with its caption "development build54d25c79" unchanged and readable (by hand: look at the picture and its caption, as the guide README's "Completing a refresh" asks; nothing about the picture changed, so this is the README's standing step, not a new risk).GuideFix-graphics-renders.jpgabove and say whether "only the version text differs" is a retention you accept, expecting a yes or a recapture request (a person; the README says a hash match is not a visual review).Risks
What could break: nothing at run time: a JSON record and a Markdown file. The one way to be wrong is the retention itself: if a reader sees a difference I did not (I compared every pixel of the first Graphics view between four commits, and read the rows), the guide would show a stale picture; the answer then is a recapture, which the README describes. A second, smaller one: the list now has 77 files, so a change to any of the ten new headers asks for a review again, as the other 67 do. Costs: none that I know of: nothing the game runs changes, so no speed, latency, memory or setup is affected, and nothing was measured because there is nothing to measure.
Turning it off: revert the commit; the check is red again on main, as now.
Compatibility: Nothing changes for a player or for the website: the PNG, its hash, its alt text and its caption are the same.
Overlaps
From a scan of the file lists of the 338 open pull requests, then a trial merge (
git merge-tree --write-tree --name-only) of each head that edits the guide's manifest against main and against this branch (the five of ours repeated on this commit before the last rebase, which did not change the patch), reading the conflict hunks (only a conflict that shows with this branch and not with main is an overlap):9872740e8,92a6c5950,4221dd097): the same kind of commit as this one (hashes and a "Retention after" section)manifest.jsontherevisionline and the lines ofvk_internal.h,vk_menu.cppandvk_menu.h(the same entry), in the note file both add a section at the end; the other two conflict in the same two filesc80312e60,8b3a36cc4)manifest.jsonconflicts with this branch and not with main: it replaces the whole entry, so it takes its own on rebaseOnce this merges, the five docs commits of #1358, #1411, #1426, #61 and #562 are redone on top of it, and each loses the part that covers main's changes (the seven hashes of #1309, #1362 and #1464 and the ten new headers). They do not disappear: each of the five changes recorded inputs of its own, so the check on each head minus its docs commit, merged onto this branch, still asks for: #1358
fps_stats.h,vk_api.h,vk_menu.cpp,vk_menu.h; #1411oot_cheat_items.h,vk_api.h,vk_menu.cpp; #1426vk_api.h,vk_menu.cpp,vk_menu.h; #61menu_i18n.h,vk_api.h,vk_menu.cpp,vk_menu.h; #562fps_stats.h,menu_i18n.h,vk_menu.cpp,vk_menu.h(checked by extracting each merge's tree and running the checker on it). This PR touches none of the five.Questions for you
The recurring cost is that the check goes red on every API bump or policy header change whose picture is unchanged: since 15 Nov (#264,
d7aeaa269) 240 commits reached main, 56 of them changed one of the files the check records, and the review each time is "render both, look, they differ only in the version text" (this PR; #1144; the five of ours). Two ways to cut it, as questions:-DLONGSHOT_VERSIONso that unchanged UI gives a byte-identical PNG on one toolchain, would turn today's look into a command that an author, a reviewer or a CI job runs, and the review note could quote its output. It would catch: any change in the first Graphics view (a row, a value, a help sentence, the prompts, the layout) between two revisions on the same toolchain. It would not: stop the check going red, because the checker compares file hashes and never looks at a picture (the CI runner has no compiler, SDL2 or libpng, and the workflow says it does not build the fixture): by itself it makes each review cheaper, not rarer; see anything the first view does not show (rows below it, other tabs, a help text of a row that is not selected: #1362's change is of that kind); work against the published PNG (a render here differs from it in about 60,000 pixels on another toolchain, so it has to be two renders on one toolchain); or notice a real change in the masked box (a longer version string that overflows the banner). A fixed version string also means the banner stops naming a build, so the PNG and its caption "development build54d25c79" need one recapture.gpu/vk/vk_api.hfirst, or should the list be the menu's own files only? The ten I mean by "the menu's own files" are, ingpu/vk/:vk_menu.cpp,vk_menu.h(it holdsOptionsand its defaults),menu_i18n.h,vk_menu_i18n.cpp,vk_menu_i18n.h,vk_menu_hit.h,menu_restart.h,upscaler_rows.h,display_modes.handmonitor_choice.h; it is my reading, a starting list, not a decision. Of the 56 commits above, 28 changed none of them;vk_api.his in 13 (every ABI bump; two of the 13 changed nothing else the check records),vk_internal.hin 16, and five of today's seven red lines (vk_internal.h,vk_reflex.h,vk_framegen.h,vk_fsr3fi.h,vk_texture.h) are headers that draw nothing. It would catch: every change to the files that define the rows, labels, hints, defaults and layout, which is where a picture goes stale, and it would stop the check going red for the other 28. It would not: catch a change in a dropped header that alters what the menu draws without touching the menu's files:vk_menu.cppuses types and constants fromvk_api.h(MkwVkMonitor,MkwVkSetting,MkwVkPointer,HOS_TEXT_KEY), so a changed field or key constant there could change a row with no hash asking; it needs someone to own the line between "draws" and "does not draw" and revisit it when a dependency arrives, where the helper's compiler-derived list is complete by construction; and it would still ask for a review when a menu file only changes another tab's text (as #1362 did on the Debug tab). The two combine: (2) makes the check ask less often, (1) makes each remaining review a command.This PR only restores the record; neither change is in it.
Issues
See #1144 (closed, superseded) and #1383, which kept this record green until #1309.
Checklist
Everyone
type(scope): description) of the status line's type.can't be run by hand is in Testing's Not testable by hand, with why. (Docs only, nothing the game runs: the checks I ran are what Testing's Tested says.)
in Risks. (Nothing the game runs changes; Risks says so.)
main. (Based ondcb7b47c2, forge main when this was written.)python -I -B tests/run_all.pypasses (any failure that also fails onmainis named). (614 passed, 30 failed, 82 skipped here and on main, both on597c5a4e0; the same 30 fail on both, named under How it was tested, 6.)third_party/untouched (a new folder with aPROVENANCE.mdinstead).When it applies
Assisted-by:/Co-Authored-By:trailers.b50f49771and from597c5a4e0; the change itself is a record and shows nothing on screen.)New on this forge; not opened on GitHub.
WIP: docs(guide): the Graphics guide image is retained after #1309, #1362 and #1464to docs(guide): the Graphics guide image is retained after #1309, #1362 and #1464Possibly related, for context: