TIM: real INT33 mouse callbacks and original Puzzle 1

Run: scratch/runs/20261114-tim-int33-callback/result.json. The original installed, packed TIM accepted the actual printed-page-1 manual answers and reached PUZZLE 1: TUTORIAL: PUT THE BALL IN THE HOOP. This proves entry into Puzzle 1, not a solved puzzle or complete gameplay compatibility.

The predecessor's final note and result were read before editing: tim-puzzle1-gameplay-20261114.md and its named run. Its generated guest registered a callback but exited with failure because no callback executed. The restored relative-input experiment had not solved the original game.

Machine.int33 now stores function 0Ch's CX mask and ES:DX FAR handler. Input producers queue snapshots of actual movement and button transitions; they never execute guest code synchronously. At a real-mode DosSession handback with interrupts enabled, outside the host interrupt stub and any active mouse invocation, the driver saves the interrupted register context, pushes a real 16-bit FAR return, and enters the handler. AX contains actual conditions, BX button state, CX/DX virtual coordinates, SI horizontal and DI vertical accumulated motion counters, cleared by function 0Bh. DS belongs to the driver. Entry clears IF/TF/DF; guest RETF reaches a reserved continuation, verifies balanced SS:SP, and restores the interrupted context, including flags and nonvolatile registers. Multiple queued transitions retain their own data. STI in a callback cannot recurse into another mouse callback. Protected-mode delivery is deferred; this change does not implement a protected-mode mouse callback ABI.

Mask zero, hardware reset and software reset discard pending callbacks. Changing registration during a callback leaves that invocation's saved return context alive. Function 14h exchanges the previous CX/ES:DX tuple. Function 0Ch supplies no invented success result. Mouse hardware reset now takes the host interrupt path rather than the Wasm fast return, so it clears the real registration too.

The ABI was checked against Microsoft's original Mouse Programmer's Reference, printed 6-18/19 and 6-27/28 and RBIL's function 0Ch/14h entries and tables 03171/03172. RBIL resolves the older manual's transposed SI/DI descriptions. The reference hash and reviewed pages are recorded in abi-reference.json. The FreeDOS CuteMouse driver source also confirms that callback SI/DI read the accumulated mickeys fields which function 0Bh exchanges with zero; a native guest test verifies this lifetime.

Actual TIM registers mask 001Fh at 1D35:5D7F. Its handler switches to a private guest stack, calls its own routine, restores the original stack and RETFs; original-handler.txt records the bounded disassembly. Session 1 proved this handler ran for balanced press/release, but the pointer landed at roughly one quarter of the requested position and no answer appeared. That session ended with an explicit review-response timeout, not success.

The startup trace explains the second primitive: TIM requests horizontal 0–2556 and vertical 0–1916 with functions 07h/08h, then programs ratio 3:3. These ranges were previously ignored. The driver now stores/clamps the ranges and position and stores function 0Fh ratios. Browser and ordinary controlled input map their screen grid into the guest's configured virtual grid; motion counters follow the configured ratio. Existing input without explicit ranges retains its conventional grid. This is absolute frontend mapping, not a full hardware acceleration or sensitivity model.

In session 2, each action followed a newly viewed frame and each balanced release preceded a fresh consequence/review:

Reviewed screen Ordinary action Observed consequence
Original title Enter Manual prompt for printed page 1
Empty answer boxes Fishbowl at (543,103) Fishbowl in first box
First box filled Cat at (351,103) Cat in second box
First two boxes filled Lightbulb at (350,200) Accepted answers; black transition
Black transition Wait without input Actual Puzzle 1 tutorial screen
Puzzle 1 Bound stop STOPPED, no held input

The actual manual's printed page 1 (PDF index 2), viewed from the predecessor's retained page image, lists fishbowl, cat, lightbulb. Session 2's actual guest callback records contain CX/DX (2172,412), (1404,412), (1400,800), with movement-plus-press AX=3/BX=1 and release AX=4/BX=0 for each selection. After the final accumulated-counter ABI correction, session 4 repeated the same six individually viewed reviews and reached the same original Puzzle 1 screen. Its callback motion counters accumulated across releases and later presses, as required. Session 3 had safely timed out at the first review with no input while review notification was missed after a lock wait; that failed attempt is retained explicitly. Only selected, viewed key PNGs remain, all named by result.json.

The original EXE is scratch/game-state/tim-20261114/C/sierra/tim/tim.exe, SHA-256 34c844ba4586f27c817442ac5ed2f5b76e35d0c6521a9a9d4e7db66404e7c51c. The manual SHA-256 is 16115f62a074f1dfb5fb80a6c34e4162d717fa74c12c078d35655b11296cb196. All 13 installed files, floppy and manual are pinned before/after. Both original sessions used a fresh private writable overlay; successful session 2 made no overlay writes. No guest, solver, configuration, clock or score patches were used. The rendered duplicate lower strip is an existing video presentation issue; Puzzle 1's title, control panel and instruction are visible, but this task does not claim to fix that presentation.

test/toyvm-mouse-callback.test.js generates native 16-bit COM instructions. The actual CPU registers, executes and RETFs from the handler, which writes its received data into guest memory and deliberately clobbers registers and flags. It verifies selection masks, simultaneous conditions, signed motion, five queued transitions, full interrupted context and caller stack, nonrecursive lifetime, active unregister/both resets, no later callbacks, mainline reset with fast interrupts enabled, function 14h exchange, and native guest range/ratio programming. Removing delivery produces the expected real-guest timeout. The predecessor's exact COM hash fbdea67ec39e9d6d8fbe66f76bbe7ad8656f4539ac4bde4d929760934a9eced1 now exits 0 after 10,035 dispatches through its executed callback. The actual original and generated guest runs used module SHA-256 44dbe86c807add8ad8c1f8010ed455ab8b011d282f9387dfaf42481109710fa1.

Released predecessor WIP sharing these source files was reviewed and adopted coherently, rather than publishing unexplained whole-file hunks. It includes short-DMA completion timing, MZ/EXEC bounds and parent flags, FAT12 startup, FCB search, read-only CD/MSCDEX mounting, VGA palette-page rendering and controlled-session input/release support. Its six helper source hashes are unchanged in incoming-helper-pins.json. Focused native short-DMA/audio/floppy tests, 118 loader/FCB/CD/disk/palette tests, controller lifetime tests, existing polling/mouse closeout/chord tests and browser CD-wrapper loading passed. DOM tests are adjunct regressions, not callback execution evidence. Both browser bundles were regenerated from stable source. Full flock /tmp/wine-build.lock bash tools/build.sh passed while also holding wine-heavy-runtime.lock; test commands and results are in the run JSON/logs.

Every launched native run/build held the heavy lock, with an independent supervisor deadline armed before child launch, actual child wait and both stdout/stderr EOF. Session 2 had an immutable 115-second controller deadline and stopped normally at 73.4 seconds; final session 4 also stopped normally inside its immutable 115-second deadline. Their independent controller watchdogs were closed. Queued lock retries launched no guest; they are recorded separately. No box was created, resumed or deleted. Git plus retained binary diff/new-file patches is the source recovery record; no source archive or media copy was made. Next work would be ordinary Puzzle 1 gameplay from this verified entry point.