Original TIM Puzzle 1: scanout repair and incomplete native play
The preserved packed original reaches the actual editable Puzzle 1 board, “PUT THE BALL IN THE HOOP.” A reviewed pickup/drop attempt leaves the ramp inventory at three and no new part on the settled board. A subsequent short held drag has delivered DOWN, held movement, independent UP and real execution after release, but its next image is a transient empty board. Placement, simulation, CLEAR and next puzzle are not established. The final session reaches the matching control panel and times out without another input; the later stale command is rejected locally. This is an incomplete gameplay result, not proof of a native physics or placement defect.
Evidence is scratch/runs/20261114-tim-puzzle1-solve/result.json. Named key PNGs
include fixed-editable-no-placement.png, short-drag-transient.png,
final-puzzle1-panel.png, and baseline-editable.png. Raw pixels are retained;
no cropping, compositing or synthesized gameplay is used. The panel's miniature
machine preview, editable board, and running simulation are separate states.
Only the first two were observed. Lower-screen numbers are not score evidence.
Generic VGA repair
Actual read-only video diagnostics show mode 12h, 640×480, four planes, 80 bytes per plane row, CRTC start 6400 (other transitions use 35840), and line compare 367: CRTC 18h=6fh, 07h=1fh, 09h=00h. The predecessor renderer ignores the line-compare address reset. Its bottom 112 rows exactly repeat the top 112 rows; addressed plane-byte hashes corroborate the linear address path.
The IBM VGA hardware reference,
September 1992, Video Subsystem printed page 2-75, describes the lower-screen
address reset after line compare. dos.js now decodes all ten compare bits and
expresses the boundary in rendered rows. framebuffer.js resets the planar row
address below that boundary. BIOS mode initialization disables the split beyond
the display. Both browser bundles are regenerated. There is no game-specific
pixel rule or guest register write in this repair.
Native reruns now show stored SCORE/BONUS graphics below the boundary instead of the previous exact repetition. Residual bottom repetition and broader scanout correctness remain unqualified. An empty upper board immediately after an action is transient: ordinary further execution recovers the board. Do not interpret that transient as lost game state or as a persistent blocker.
Five new generated-plane tests cover all compare bits, reset boundary,
nonzero start/stride, disabled defaults and planar mode X. They fail against
predecessor source and pass with the repair. Together with existing palette
coverage, 14 focused tests pass. Browser bundle loading, persistent input/session
controls, and run-local drag sequencing tests pass. These are implementation
and controller tests, not native gameplay fixtures. The full build passes under
both heavy-runtime and build locks; see pipeline-results.json and build logs.
Original media and ordinary input
The 13 preserved installed files under scratch/game-state/tim-20261114/C,
original floppy /dev/shm/wine-owner187-tim-disk.img, and scanned manual
scratch/media/tim-original-manual.pdf are referenced by path and SHA-256.
All 15 hashes and the exact installed-file closure match after the runs.
The installed executable hash is
34c844ba4586f27c817442ac5ed2f5b76e35d0c6521a9a9d4e7db66404e7c51c.
No unpacked executable or solver state was loaded. Each startup uses a private
in-memory file overlay. No controller action writes guest configuration, scores,
physics or clock state. No native gravity or air-pressure choice was made.
The actual manual images were viewed: printed page 1 supplies fishbowl, cat, and lightbulb; page 10 describes the panel green arrow; page 13 describes part pickup/drop, the runner start button and left click to stop. Ordinary fresh startup passed the manual prompts. No running simulation was entered.
Session 1 is nonqualifying: two commands incorrectly described expected screens
while the actual screen was still the title or a black transition. Its explicit
correction is session1/review-corrections.json. Sessions 2 and 5 time out at
the first review with zero input. Session 3 reaches the editable board before
the independent deadline kills its owner. Session 4 performs reviewed pickup
and drop, then waits to the unchanged settled board. Session 6 performs the
short held drag, releases and advances, then loses the next review. Session 7
accepts the manual answers and reaches the panel; review response timeout
prevents the planned longer held drag. It was never delivered. These timeouts
are controller interaction limits, not demonstrated guest faults.
Every native lease is immutable at 115 seconds, with an independent 119-second outer deadline armed before launch. The response window is 28 seconds. No active lease is extended. Heavy-lock queue time precedes launch. Every supervisor has actual child wait and both stream EOFs; session 3's killed owner and reaped watchdog are recorded. Postflight checks find no owned runtime/watchdog left. No box was created, resumed or deleted: ID none, state NOT_CREATED, expiry none, stop receipt not applicable. Disk remains above 1.5 GB.
The next useful work is a fresh, promptly reviewed native pickup/drop and settled part-count check, followed by the runner and visible motion. The current result cannot support a solve claim. Preserve the residual scanout issue as a separate question until actual register/address evidence explains it.