Thief II: The Metal Age demo
Source: https://archive.org/details/ThiefIiTheMetalAgeDemo Metadata artifact thief2.exe,136880529 bytes, SHA1 e812b6d8a41c7755f87e1e17bf8e7ef11a57de5a. Metadata identifies Looking Glass Studios/Eidos demo,2000; listing contains a1999 demo payload. Private local compatibility fixture only, no redistribution/deployment authorized. Not in current corpus/public registry or existing task list when selected.
Source preflight 2026-10-12 07:06Z
Evidence scratch/runs/20261012T0706Z-thief2-preflight/. Standard endpoint initially appeared stalled. The original completed at07:10:51Z after the last status check. An unchecked mirror retry started at07:11:03Z and overwrote that full file with partial bytes. Original download.json and listing.txt record the earlier complete download, not current payload validity. Current-payload.json is authoritative:59440936 bytes, SHA1 0ea8dbb7247fa7c4a022eb846dc06cee4e72d660, NOT the required complete hash. It was renamed /home/user/thief2-demo-preflight/thief2.exe.part on bx_gqubeuhs before stop. Check the final snapshot; a pre-rename snapshot may retain it as thief2.exe. Rehash either file before using it. Next resume the partial from https://ia801501.us.archive.org/21/items/ThiefIiTheMetalAgeDemo/thief2.exe with a unique destination and no overlapping download, then verify full hash. Use curl -C - with a sufficiently long bounded timeout after checking range support. Never extract or qualify the partial file.
Complete-download listing shows61 files/168447241 expanded bytes, including thf2demo/MOVIES, mission miss11.mis, readme.wri and installer components. Next extract verified payload with7z, inspect original executable/imports and readme, then bounded remote-only launch. No game executable extracted or launch/input/browser/audio/FPS qualification yet. Exact missing prepared path: /home/user/thief2-demo-preflight/game/ (not created).
Verified resume and startup (2026-10-12 07:23Z–07:28Z)
Remote bx_vujsz54x resumes the single partial with curl -C -; no overlapping writer. Full 136880529-byte SHA1 e812b6d8a41c7755f87e1e17bf8e7ef11a57de5a verified, then 7z extraction succeeds. Evidence 20261012T0723Z-thief2-resume. Original thf2demo.exe SHA256: 5e5f831ba549081c35e592afaa540f336f76c828a996b4fcb74edef46beba942. Test build is main0304a8549 plus pending Z16 correction3a4fcf883, module3c3a4efe77335ac8af1719f435d222fc142b330f23ae234171ad42512588fb4d. Build gates pass. No Thief-specific emulator or executable patch.
First direct launch0725Z lacked installer-generated DARKINST.CFG, did not open the resource archives, and called NULL from538d87 at batch36. Adding install_path, resname_base, load_path, script_module_path all C:, plus movie_path C:\MOVIES, loads intrface.crf but hits integer divide by zero (C0000094, reported block44b375) at batch93. Actual divide44b37a uses animation frame count at object+4. Main animation list MAINM.STR names maina/mainb/mainc/main; the last animation's MAIN_1.PCX resides under english/. Original SETUP.ins contains all these path keys and literal language english. Add language english to DARKINST.CFG:0728Z renders normal main menu, no divide fault. This is installer reconstruction, not a game/code workaround. Keep the original CRFs, MOVIES and configuration files.
Ordinary click310,240 at batch400 reaches difficulty in0729Z. Click330,238 at800 selects Play/Normal;0730Z renders The Unwelcome Guest objectives. Menu/mission objectives are not gameplay qualification.
Registered local app thief2_demo/candidate thief2-demo is prepared remotely (40 data files plus original executable). CLI0731Z Continue392,444 at1400 reaches loadout. A PlayMission click at2600 is too early during loading; 0735Z delays it to6000 at530,456 and reaches the textured first-person mission with HUD, reviewed final.png. No movement in that capture. The probe inherited a60000ms spawnSync timeout despite a90s game guard: exit.json records ETIMEDOUT, but the child continued and ended0 at90.018s with final PNG. Preserve that caveat; do not treat this as clean timing data. Future probes must set the external timeout above the internal guard. Browser0737Z uses registered local launch on main+Z16 and Chrome155, waits for each menu state before proceeding; validation ongoing.
Registered browser movement and mobile limitation
0737Z-thief2-browser (started07:36:23Z) on Chrome155/WebGL/workers: ordinary menu clicks at240394,249394,290538; wait35s for loadout; 389546 PlayMission, wait45s. ArrowUp2s moves forward beyond the starting arch; ArrowRight500ms visibly turns right. Before/walk/turn reviewed. This is actual ordinary player-controlled gameplay. No FPS/audio assessment.
0742Z-thief2-touch displays all six controls at844x390 and dispatches actual CDP dpad touch, but the mobile menu clicks used whole-canvas coordinates without compensating for game letterboxing. Captures remain at main menu: FAILED mobile navigation, not touch gameplay. Next map guest coordinates with renderer._unmapExclusiveInputPoint plus its canvas CSS bounding rect, then verify actual dpad displacement. Shorten Objectives label to Goals to avoid the top-right button clipping in phone landscape. Physical device, Safari, full actions and audio/FPS remain unverified. Registration30014ef63 is pushed, ops main integration pending.
0750Z-touch-mapped repeats with the renderer inverse presentation transform; still main menu. Therefore correcting coordinate mapping alone has NOT established the cause. Next compare actual p.touchscreen taps against mouse events under mobile emulation, capturing each menu checkpoint. No further blind route replay. WinePerf.snapshot counters all0 in this run; instrumentation was not enabled, so this is unavailable measurement, NOT measured0FPS. Core keyboard gameplay remains reviewable; mobile/audio/FPS have a separate ready follow-up. Goals label change e2c8bd4cd pushed; debug close overlay still overlaps top-right control in debug mode.
Integration reconciliation, 2026-10-13
Recovered from origin/codex/wsasocket-capability at 15b37458e during the
release merge audit. The local-only registration and its required runtime fixes
are now on main. Earlier statements that integration is pending describe the
historical run, not the current branch. This recovery adds no fresh gameplay,
audio, performance, Safari, or physical-device qualification; retain the
recorded run IDs, module hashes, limitations, and failed attempts above.
Native interaction attempt, 2026-11-01
Run scratch/runs/20261101-thief2-native-interaction/ did not establish
opening a door, taking loot, or visible objective progress. One native original
reached Normal / The Unwelcome Guest, then stopped at the first attempted-route
navigation blocker: the nearby torch-side barred window above a raised stone
sill. No reachable door or loot item was identified; no use/attack was attempted
against the window. Other routes remain untested; no emulator-defect finding.
No second original was launched.
Original readme.wri and unchanged DARK.BND / user.bnd document right mouse
as +use_item, left mouse as +use_weapon, S walking, Up running, M map, and
F4/F5 lockpick selection. The readme says to approach a highlighted door and
hold right mouse to pick its lock. Highlight alone would not establish opening.
The reconstructed DARKINST.CFG and registered defaults were preserved.
The coordinator explicitly overrode the interrupted Quake II launcher gate at
12:53:34Z after original exit 0 and four absent PIDs. ACK at 12:54:36Z set the
13:19:36Z runtime deadline. Actual gate/source main was
90aa4059d818f50185ae5724a945f3b3de34f77b, rather than dispatch f581e849f.
All 4,392 accepted source pins and 141 accepted read pins matched; reused module
SHA256 dd43e4a6624d2449d7133b731d687c32530f7fcf0041a69ba71a09f78c09aaf4
also matched actual runtime-compiled bytes. Before original use the observer
passed actual sync/async execution, whole shared 512 MB read-only comparison,
clock identity and emitted late-10001 retention. The run's 4,444 source/module
pins, 41 fixture files plus registered manifest and recovery inventory, full
fixture tree and 142 observed runtime reads remained unchanged. EXE SHA256
matched 5e5f831ba549081c35e592afaa540f336f76c828a996b4fcb74edef46beba942.
Observer coverage is string-path readFileSync/readFile/promises.readFile,
not exhaustive descriptor/OS I/O or compiler-source provenance.
Native CLI used 100,000-block batches, registered-default 200 ms/batch guest clock, software rendering, frozen stdin control, a 1,300-second internal guard and longer 1,380-second outer allowance bounded by the gate deadline. Menu, Normal, objectives, loadout (batch 2400), and world/HUD (3500) were inspected. Browser milliseconds were not replayed as native batches. Captures sometimes preceded consumption/presentation: the map became visible at 4376 after its key at 4085. The initial five-batch turn release was still queued at 3535; another turn was issued before it drained, so early captures do not isolate those turns. Later controls used longer settling. All 20 keyboard edges / ten balanced pairs ultimately dequeued, routed and dispatched to HWND 0x10001. Final known queues were empty; guest post queue unknown. No interaction-button consumption claim.
Before/held/released/settled captures document approach to the window. Neutral batches 5431–5631 retained identical sampled window, sill, floor and HUD pixels; torch/snow animation is not interaction progress. All 61 PNGs passed CRC/deflate checks, full compressed runtime log decoded, and focused identity/input/evidence checks passed. PID 3671463 exited 0 after 366.488 seconds at batch 5631 without external timeout. Runtime disk minimum was 2,038,202,368 bytes; no RAM gate. Native local box ID/state/expiry N/A; local shutdown explicitly delegated. No audio, FPS, GPU, browser or full-suite qualification is added.
Automated recovery, 2026-10-14
Fresh acquisition from https://archive.org/download/ThiefIiTheMetalAgeDemo/thief2.exe verified SHA1 e812b6d8a41c7755f87e1e17bf8e7ef11a57de5a (metadata size136880529). Extract with7-Zip23.01 and copy archive/thf2demo/ to game/, then write documented DARKINST.CFG: install_path,resname_base,load_path,script_module_path each C:; movie_path C:\MOVIES; language english. This file is reconstructed installer state, not original installer output.
Run node tools/fetch-candidate-corpus.js --id=thief2-demo, then node test/test-thief2-recovery.js on the pristine recovery. The41-file inventory test/candidate-corpus/thief2-demo-files.json pins paths,sizes,SHA256 and labels DARKINST.CFG reconstructed. All other files matched extracted source bytes; thf2demo.exe matches its prior recorded hash. Comparison with restored fixture differed at: none. Runtime-written settings may differ after playing; no new gameplay or audio/FPS qualification is claimed. Existing private-fixture restrictions unchanged; no public deployment.
Rightward rooftop route and disk stop, 2026-11-01
Run scratch/runs/20261101-thief2-reachable-door-route/ established no door
opening or loot acquisition. One original reached Normal / The Unwelcome
Guest. The predecessor's map-drained and map-closed PNGs both show the world,
not a usable map. This run actually displayed both map pages at batches 3812
and 4312, then closed the map. Neither captured page supplied a usable current
position marker. From the opening arch, rightward turns inspected a different
approach from the predecessor's torch-side barred window. A short S walk
approached the stone wall/parapet beside the arch; no highlighted reachable
door or identified loot appeared. The first visible obstacle on this attempted
route was this stonework. Collision or impossibility of passing it was not
established: the disk guard ended exploration. The leftward rooftop route,
other roof edges, jumps and further streets remain explicitly untested. No
window was treated as a door and no emulator defect is inferred.
The concrete stopping blocker was disk space. The original controller detected 1,999,806,464 available bytes, below the required 2,000,000,000-byte floor, and gracefully quit: PID 3695951 exited 0 after 272.932 seconds, final batch 5873. Therefore the continuous disk-floor requirement failed; this is not an accepted interaction result. Later closeout samples were lower still. No foreign cleanup or second original was attempted. The auxiliary disk guard ended before runtime completion (its last retained sample was 13:21:40Z); the runtime controller continued sampling while reading output and caught the floor. Do not describe that coverage as uninterrupted independent monitoring.
Gate passed at 13:18:25Z only after Quake II worker 01a19730's fresh launcher
final.txt/result.json and explicit runtime/source/build RELEASE. Actual main
was 8f2f8a802db44745a1ac00fe3c65e25965eb587f, versus dispatch e6814d7b9;
the 25-minute deadline was 13:43:25Z. All 4,392 accepted source pins and 141
accepted read pins matched, allowing reuse of module SHA256
dd43e4a6624d2449d7133b731d687c32530f7fcf0041a69ba71a09f78c09aaf4.
Actual runtime compilation matched it. Before launch the observer passed actual
sync/async execution, whole shared 512 MB read-only comparison, host/native
clock identity and emitted late-10001 retention. All 4,444 source/module pins,
43 original/manifest/inventory pins, full fixture tree and 142 observed runtime
reads remained unchanged. Original EXE SHA256 remained
5e5f831ba549081c35e592afaa540f336f76c828a996b4fcb74edef46beba942.
Read coverage is string-path sync/callback/promises readFile, not exhaustive
OS/descriptor I/O or compiler-source provenance.
Original readme and bindings were read again: S walks, Up runs, right mouse uses items, left mouse uses weapons, M maps, F4/F5 select lockpicks. No use, weapon or lockpick input was sent. Ten keyboard edges/five balanced pairs (M, three Right turns, S) dequeued/routed/dispatched. Before/held/released/ settled captures were preserved. Each subsequent action waited for the prior release dispatch and empty known queues; presentation received additional neutral batches. Even so, map presentation lagged consumed input. After the S walk, 200 neutral batches retained identical sampled wall, parapet and floor pixels; HUD pixels differed, so no whole-frame stationary claim. This is a neutral comparison, not a blocked-forward collision test or interaction.
Final scheduled/pending/cross-thread/renderer queues were empty, with S key-up last dispatched. Used-key edge balances ended released, but final physical/ async states for M/Right/S and mouse-button mask were not captured before the disk-triggered exit. Modifier/Space observations were false. Guest post queue, button queue and SDL consumption remain unknown; trace retention is 1,000 observations and counters saturate at 2^32−1. Do not infer missing states from balanced edges alone.
All 34 PNGs passed CRC/deflate checks and the full runtime gzip decoded.
To preserve evidence while reducing only owned storage, screenshots were
losslessly recompressed: decompressed PNG scanline bytes were compared equal,
with original and resulting hashes in capture-compression.json. Exact-content
hardlinks are recorded separately; no foreign file was modified or removed.
Identity/input/pixel checks passed; the disk criterion failed. No tracked
source/helper/test/config/backend/registration edits, build, browser, GPU,
audio, FPS or full-suite claim. Own runtime/helper PIDs were absent after
shutdown, no new child temp directories remained. Native local box ID/state/
expiry N/A; local own-process shutdown was explicitly delegated. Further work
requires disk headroom and a fresh ordinary route; this run stops here.
Left-side roof edge, 2026-11-01
Run scratch/runs/20261101-thief2-left-route-interaction/ established no door
opening, loot acquisition or objective interaction. One original reached
Normal / The Unwelcome Guest. From the opening view, successive left turns
looked past the torch/window wall to the other opening of the starting stone
structure. The selected route followed the wooden platform toward the open
edge facing a timber-framed building, beside a sloping stone support. This
differs from the earlier torch-window approach and rightward stonework route.
S walks, Up running, a Z strafe away from the support, and one intentional
Up+Space forward-jump attempt changed the view but did not establish a crossing
of that edge. A downward inspection showed the platform edge and building
across it. The first blocker is navigation beyond this platform/open edge:
no reachable highlighted door or identifiable loot was found. No right-mouse
use, weapon or lockpick input was sent against unidentified geometry. Collision,
an impassable gap and an emulator defect are not established. Further roof
edges, streets and other jump trajectories remain untested; no second original.
Original readme and both bindings were inspected again: S walks, Up runs, Z strafes left, Space jumps, V centers view, right mouse uses items, M maps, and F4/F5 select lockpicks. The readme requires approaching a highlighted door and holding right mouse, possibly alternating picks. Highlight alone is not an interaction result. Existing map captures were reviewed; neither supplies a usable position marker. This run navigated from its actual opening view.
The gate passed at 13:47:18Z after fresh Quake II eight-save-onward launcher
final.txt/result.json, its 13:47:11Z explicit release, and the coordinator's
13:47:10Z genuine runtime/source/build release. The predecessor's earlier false
gate was not consumed. Actual main was ba0ba17ebf6d1baa1cab8a4c483e4cd5647857a3,
versus dispatch 2ec46b5d7b3d423bd8191684c66dbf24e2f7caca; deadline 14:12:18Z.
All 4,392 accepted source pins and 141 accepted runtime-read pins matched,
allowing reuse of module SHA256
dd43e4a6624d2449d7133b731d687c32530f7fcf0041a69ba71a09f78c09aaf4.
Actual compilation used those bytes. Before original launch the observer passed
actual sync/async execution, whole shared 512 MB read-only comparison, native/
host clock identity and emitted late-10001 retention. Its added frozen host-key/
button observations passed a read-only parked-state check. The 4,444 source/
module pins, 43 original/manifest/inventory pins, complete fixture tree and 142
observed runtime reads remained unchanged. EXE SHA256 remained
5e5f831ba549081c35e592afaa540f336f76c828a996b4fcb74edef46beba942.
String-path sync/callback/promises readFile coverage is not exhaustive descriptor/
OS I/O or compiler-source provenance.
The CLI used software rendering, 100,000-block batches and the unchanged 200 ms/batch guest clock. The first loadout capture at 2400 was black; actual loadout was reviewed at 3400 and world at 4500. Each keyboard action waited for its press dispatch, then its release dispatch and empty known queues before 300 neutral batches; the deliberate forward-jump chord balanced both keys. All 28 keyboard edges / 14 pairs dequeued, routed and dispatched in order. Two relative mouse movements inspected height. Final known queues were empty, all seven used keys had physical/async held states false, and the host mouse button mask was zero. Their asyncPressed latches remained true: these record earlier presses, not held keys. Guest post/button queues and SDL consumption remain unknown. Trace retention is 1,000 observations; counters saturate at 2^32−1. Before/held/released/settled captures are preserved. A 300-batch matched neutral retained identical sampled timber, edge, floor and HUD pixels; no whole-frame stationary or blocked-forward collision claim is made.
PID 3725735 exited 0 after 450.302 seconds, final batch 11799. The enclosing execution-tool shell reported 143 despite the saved original exit 0 and normal driver completion; its cause is unknown. All 64 PNGs passed CRC/deflate checks, the full runtime gzip decoded, and focused identity/input/pixel checks passed. Owned captures were losslessly recompressed with equal decoded scanline hashes and exact-content hardlinks. Runtime controller minimum was 2,145,902,592 bytes; auxiliary minimum 2,145,906,688 bytes, both above 2,000,000,000. The auxiliary guard remained active through publication checks. It polled every 0.25 seconds, with controller checks on the output loop; no complete auxiliary sample history was retained, so sub-sample dips cannot be excluded. No RAM gate or foreign cleanup. Native local box ID/state/expiry N/A; own-process shutdown explicitly delegated. No tracked source/helper/test/config/backend/registration changes, build, browser, GPU, audio, FPS or full-suite claim.
2026-11-11: authentic save restored; native water-arrow ammo consumed
Run 20261111-thief2-save-mouse-weapon-response-1040 continued the 0940
browser save with the original demo EXE, matching media/source pins and the
successful serial WebGL renderer configuration. The unchanged predecessor
quick.sav.gz decoded to SHA256
009d761b8b94b0511f2e9b2fc7ff6b97f4f332c4b6af0f2c20e69b5fc3cb4a57.
Normal filesystem create/write/close restored it, and the actual Load Game /
QuickSave / Load UI restored the rooftop torch scene. No save editing occurred.
The shipped user.bnd binds mouse1 to +use_weapon. Key 4 selected Water
Arrow; frame 20 shows ammo 10. A native left-button press held for three
seconds and released produced ammo 9, visible in frames 22 and 23. Ordinary
mouse input also changed the view; torch extinction and projectile impact were
not established. The original command table points use_weapon at 0x542a20;
the existing passive execution counter changed from zero to two across that
press/release. This establishes a native weapon/ammo consequence.
DirectInput consumption was not established: renderer and guest memory objects
and backing buffers were identical (512 MB), the generated-map mouse queue
head stayed zero while its tail advanced 26 to 28, traced DirectInput calls
were absent, and the final DX object census had no DirectInput root/device.
WM messages alone are not evidence of DirectInput reads. This run demonstrated
working game-handler dispatch without reproducing an emulator/control defect,
so no production repair was made. Matching module SHA256 was
75ecda5701c15f6fe2c269455ead4e15457b90653733df15d35674b08e8849d5;
the inherited locked-build proof matched, and fresh test-thief2-recovery.js
(41 recovered files), test-directinput-device.js, and same-box control/guard
checks passed.
Native F11 generated an updated authentic save, retained under this run's
updated-authentic-saves/quick.sav.gz; raw SHA256 is
10d98308157e3f3805eed90235fb812ba09477d4de6139b810ce57ec3e8ef178.
The before/after full captures, handler observation, save and full terminal
are locally hash-verified. Auxiliary recovery was incomplete: the worker
stopped the box after a compact-download failure before checking recovery
success. Of 204 remote manifest paths, 85 exact raw hashes were retained or
reconstructed locally; 119 (mostly passive captures) remain unavailable.
recovery-reconciliation.json and local-encoding-map.json explicitly map
retained bytes and missing paths; this is not a complete auxiliary archive.
Browser/controller exits were normally waited with exit zero and both EOFs;
independent checks confirmed PID/profile/port absence. Box bx_4xb5wgpa was
normally archived at 06:59:15Z, stop stop_3c0f88fb5858cf68, completed without
force/error and independently confirmed. Its initial expiry was 07:25:34.595Z;
the prearmed fixed watchdog exited unfired. No resume, deletion or deployment.
2026-11-11: saved rooftop water arrow extinguishes the torch
Run 20261111-thief2-saved-arrow-world-interaction-0730 restored the unchanged
1040 updated quicksave (raw SHA256 10d98308157e3f3805eed90235fb812ba09477d4de6139b810ce57ec3e8ef178)
through normal filesystem create/write/close and the actual Load Game /
Quick Save / Load UI. Original media and all 1,335 provisioning pins matched;
the successful serial-thread browser configuration and module SHA256
75ecda5701c15f6fe2c269455ead4e15457b90653733df15d35674b08e8849d5
were reused with matching locked-build evidence.
Ordinary shipped lookcenter/turn/look controls aligned the torch. An initial off-center pointer press changed the camera without consuming ammo. Moving the pointer to the gameplay center, realigning with keyboard controls, and holding native mouse1 for three seconds then releasing succeeded. Full frame 70 shows the lit torch and Water Arrow 9; frame 72 shows the same view with the flame gone, nearby wall/floor illumination reduced, and ammo 8. Frame 74 confirms the dark torch after native F11. All three full images were personally reviewed. No guest-state changes, save editing, or production repair occurred.
The new game-written updated-authentic-saves/quick.sav.gz is retained locally:
raw 2,328,726 bytes, SHA256
5326f3a67f08c2677effb45527812ddcdd595fc04154139639eda4e1075bde6e.
Reload of this newest save remains untested. Fresh recovery (41 files),
DirectInput, and same-box balanced-input/negative guard checks passed. This
establishes torch extinction through authentic gameplay, not mission completion,
audio/FPS validation, or DirectInput consumption.
All 424 auxiliary paths and the saves were recovered and raw-hash verified
before normal box stop; recovery-local-map.json resolves exact retained
bytes through identity/gzip encodings. No source/media archive was made.
Chrome/controller actual waits were zero with both EOFs, independently checked
alongside absent processes/profile/port. The enclosing local execution session
reported 143 and did not retain its parent stream terminal; its cause is unknown,
and the original command was never replayed. This does not replace the retained
remote child wait evidence. Shared disk growth caused transient dips below the
1.5 GB floor; coordinator alerts and owned lossless compaction are recorded.
Box bx_ef3fprsg was normally archived at 07:41:00.506Z, stop
stop_050bfe23cb1e0a64 completed without force/error and independently confirmed.
Initial expiry was 08:12:07.684Z; current expiry is null. The fixed independent
watchdog was prearmed before creation and exited unfired with actual wait zero
and both EOFs. No workers, restart, resume, deletion, or deployment.
2026-11-14: "white sparkles" on walls were sky stars drawn without depth
The scattered grey/white dots over the opening arch's ceiling and walls are
the mission's 19 sky stars, not rasterization cracks. The game draws them as
one TL POINTLIST (19 vertices, sz=0.999949, rhw 0.0041, colours
0x7b7b7b..0xdddddd) after ~1,600 world draws per frame, with ZENABLE=1,
ZFUNC=LESSEQUAL, ZWRITE=1, and relies on the walls' depth to hide them. Both
D3DIM point paths (WebGL _drawPoints and WAT software) ignored depth and
plotted 2x2 dots, so all 19 showed through geometry; they are static across
frames (snow would move), which is the quick tell. Fixed by depth-testing TL
points ($d3dim_plot_point, _drawPoints); the 2x2 footprint is unchanged.
Before: scratch/runs/20261114-thief2-sparkle-gpu3/ (20 dots); state probe:
20261114-thief2-sparkle-points-probe3/; after: 20261114-thief2-sparkle-fix/
(2 stars left, both in the visible sky). Route: clicks 310,240@400,
330,238@800, 392,444@1400, 530,456@3400 and @3700 with --batch-size=100000;
gameplay is on screen by ~3800. The software rasterizer additionally leaves a
few single coloured pixels near the far window (not present on WebGL).