Lords2 PIT calibration
Task: NEW-GAME-LORDS2-DEMO-20261015. Evidence: scratch/runs/20261107-lords2-pit-calibration/. Follows the repaired CMOS installation attempt.
The prior original is evidence, not another run authorization. Its retained blocks show eleven helper calls: one warmup followed by five pairs of timing samples. Each helper lowers channel 2's gate through port 61h, writes B4h to port 43h and FFFFh low/high to port 42h, raises the gate, executes a divide loop, restores the gate-low control, and reads the counter low/high. The sample pair compares 16-bit and operand-size-prefixed 32-bit division loops. The captured returns are both AX=1; five subtractions accumulate zero. Final block 1:9E3A averages that zero and subsequently divides a nonzero product by the zero average. The prior run did not separately capture the fault-time instruction or exception code. prior-calibration-blocks.json retains the selected evidence and identifies its exact source hash.
Intel's 8254 datasheet, order 231164-005, describes B4h as channel 2, low/high access, binary mode 2. A low gate stops decrementing; its next rising edge reloads on a clock pulse. A count latch persists until consumed and ignores repeated latch commands; status read-back takes priority over count bytes. Count-register writes and counting-element transfers are distinct. Intel's PIIX4 datasheet, order 290562-001, sections 4.2.3, 4.2.4.1 and 8.8, documents the port mapping, channel-2 gate on 61h bit 0 and output status on bit 5.
Before the repair, both retained module variants reproduced the complete causal chain using generated ordinary Win16 instructions: ten samples equal 1, average zero, exception C0000094 at the generated DIV instruction, then unhandled CPU termination and yield 2. This confirms the timer mismatch without changing exception handling. The new production regression rejects those old modules because they lack the PIT import.
The repair routes immediate and DX IN/OUT instructions for ports 40h–43h and 61h through a process-owned polled binary mode-2 8254. It implements low/high and single-byte access, independent read/write phases, count/status latches and read-back priority, deferred count transfer, periodic reload, gate freeze/restart, channel-2 OUT status, and zero-as-65536. Counter 0 starts in the firmware rate-generator configuration used by the existing seed consumer. Clock input is elapsed monotonic time at 1,193,182 Hz, independent of guest batch time and calendar changes; reads do not advance it. Cooperative contexts share the device and OS Workers serialize returning imports through existing synchronous RPC. Unsupported modes, BCD, illegal mode-2 count 1 and wide accesses fail explicitly. IRQ/audio generation is outside this polled device.
The locked full build passed, as did generated instruction calibration and sequencing controls on standard and compatibility modules, host-state/isolation and real OS-Worker controls, CMOS/calendar/worker integration, existing x86 operation and upper-half regressions, and the channel-0 seed consumer. Armed/disabled/saturated observer controls preserve architecture and the complete 512 MiB memory on both modules, including the strict floating-port fault control. The first fixture-control failure was a test scratch-store collision with its inherited fixture's saved SP; relocating the new scratch bytes fixed the fixture without production changes.
A separate natural-clock generated consumer measured unequal samples, but still raised C0000094 at the final 16-bit division: its measured average (294 standard, 292 compatibility in this observation) produces a quotient above FFFFh for the generated numerator. This is retained as a calibration-range limitation, not a success or reason to alter clock rate or substitute a guest result. The ordinary original result below establishes this guest's outcome; the generated natural-clock overflow remains disclosed.
Evidence: baseline-causal-*.json, instructions-*.json, focused-results.json, observer-control-results.json, natural-*.json, and build.log in the run directory. Runtime/source/build were acquired only after Aranna's explicit 01:18:35Z release and independent process, port, profile and four-lock cleanup verification.
The repair was published as 1de8ccb957ce2ef81301c927b45eb3e855e03e99 on main, followed immediately by successful strict-SSH pull/push. The single original attempt ran 01:25:44Z–01:27:51Z. Before its launch, a missing preparation lease PID was detected; no browser, server, runtime-start or original-attempt guard had been created. The supervisor then acquired the three runtime locks directly under the unchanged board ownership. prelaunch-lease-loss.json records that prelaunch recovery; this was not a second original.
Personally reviewed ordinary menu screenshots show Install, System Test Yes, Wave Yes, CD-ROM Detect OK, Video Display Test OK, and completed System Test Results. The test reports 65,152 KB physical memory and a Pentium Pro at 13 MHz. The CPU routine executed all eleven helpers and 110 port operations with no CPU exception, termination or PIT error. The negated samples were 119 (warmup), then (1073,596), (836,357), (716,358), (596,358), (716,477). The five 16-bit differences wrap to 65059,65057,65178,65298,65297; the average is 65177. The recorded return to selector 95:4BA7 carries AX=117. The installer subsequently converts/classifies that heuristic into its displayed CPU result; this is not a claim of accurate host MHz or instruction-cycle timing. The wall-clock device does not manufacture sample values or calibration outcomes.
Using the installer's ordinary Yes option after its minimum-requirements warning reached a _setup dialog: “Insufficient memory to perform operation.” The retained USER ordinal-1 entry records that exact text and caption, with return address selector 103:6265. This is the first remaining blocker. Its underlying allocation or API mismatch has not been established, and the message alone is not proof of memory exhaustion. Installation, installed-state launch, and gameplay were not reached. No second original was started and no guest/media/configuration bytes were edited.
original-causality-review.json identifies the full trace, all samples, final arithmetic and return blocks, and the memory-dialog API frame. The full original state snapshots and output are retained losslessly compressed where large; compression.json records original and compressed hashes and recovery instructions. HTTP receipts contain 213 responses: 203 exact file matches, two index responses whose normal dev-server injection was reconstructed and hash-verified, and eight retained 404 build-info/CARDS probes. The served module matches the retained tested standard module and both observer module identities; the host-import bytes also match. All 181 original media files and the registered manifest were verified in place before and after.
The driver and Chrome exited normally with code 0. All recorded runtime PIDs, server and debugging ports, browser profile and profile-matching processes were absent; all four runtime/build locks were independently acquired and released. Explicit runtime/source/build release was posted after cleanup. No boxes were created or assigned. Remaining work is a focused investigation of the new memory dialog using the retained API/block evidence, followed by separately authorized ordinary installation progress.