Files
Neutron/docs/evidence/windows-session.md
T

15 KiB
Raw Blame History

One combined local Windows fidelity session

This is the prepared agenda for plan 01-07, not a request to capture now. READY.json means immutable materials are prepared. It never means original, native Windows, browser, audio or scenario acceptance passed. Linux PowerShell selftests, cross-publish, Wine and synthetic media cannot fill those slots. Keep every package, original copy, raw recording and result on local storage.

The checkpoint contains manifest.json, READY.json, both identified port/{linux-x64,win-x64} packages, original-pristine/, writable original-runs/, protocol/, offline review/, collectors, these instructions, package-free selftests and results/. No SDK or Python is needed on Windows. The manifest records actual package BUILD-INFO and sealed source identities; the later bundle revision does not replace the runtime's source identity. Do not alter immutable materials to make a preflight pass. Mutable results and generated run copies stay outside that immutable hash set.

Ordered session agenda

  1. Select local dependencies and capture provenance. Transfer the entire prepared checkpoint to a fresh local directory, preferably a short path with sufficient recording space. Select the actual Windows PC, DxWnd executable, original exported DxWnd profile, recorder, browser and any media-probe/ derivation executables already installed locally. No download, installer redistribution, elevation or guessed DxWnd switches are part of this bundle. Copy protocol/settings-template.json into results/session-settings.json and fill observed values. Record each tool's actual local source, version, SHA-256, settings/export and limitations, Windows build, GPU/driver, Vulkan and SDL versions as actually available; missing identities remain pending. The collector hashes selected external tool executables without copying them into returned results. Retain settings/profile exports, not installers. Initial recorder recommendation: offline OBS Display Capture, with the actual monitor, capture source, crop/aspect, resolution/frame rate, encoder, audio device/mix and MKV container recorded. Record notification/desktop exposure and wrapper timing/rendering limitations in the local notes.

  2. Run immutable preflight and package-free selftests. From the checkpoint directory run the following exact commands. Require exit zero, at least eight executed checks and zero failures; preserve the output in local session notes. A missing PowerShell or dependency is a failure to record, not a silent pass.

    powershell.exe -NoProfile -File .\scripts\collect-fidelity-windows.ps1 -Action Preflight
    powershell.exe -NoProfile -File .\tests\fidelity-windows.Tests.ps1
    

    PowerShell 5.1 is the Windows target; a pwsh run on Linux exercises tooling only. Executables are launched directly with quoted arguments, never from receipt-supplied commands. Keep script execution subject to the PC's existing policy; record a policy failure rather than changing security policy here.

  3. Perform separate short original and native-port visual/audio preflights. Use new diagnostic IDs for each. Inspect the selected DxWnd distribution's own local help/config. Import the user's retained exported profile and change only executable/working-directory paths to the generated copy printed by the collector; export that relocated profile to the result's artifacts directory. Do not force timer, renderer or compatibility settings or guess profile syntax. If automation is unsupported, use the transparent DxWnd GUI launch. Retain original and relocated exports and describe exact changes. The collector opens DxWnd with no assumed switches and does not infer the original child's exit. Never select the authoritative installation or original-pristine/ as target. Play each raw preflight: require visible moving game content, intelligible game audio, correct crop and aspect, rather than merely detected ffprobe streams. Record failures before proceeding. Keep diagnostic runs separate from required cases. Native port launch uses the existing NewGame.tsk story entry; absence of a reconstructed menu remains an explicit comparison gap.

  4. Record all matrix cells and obligations. Use protocol/agenda.json as the checklist: 8 cells × 5 obligations × 2 Windows sides = 80 Windows obligations. The separate Linux side contributes 40; all 120 remain subject to actual receipt validation and human review. The table below and protocol/scenarios.json specify the matrix, actual route, required events and end conditions. Each fresh run has a new launch ID and original copy. For retained runs, record prerequisite lab departure to VR04 and return in the same process/clock before the tested interval; do not relaunch or reset state. Capture the retained generated-copy identity during that lifetime with -Action Retained, preserving the original pre-copy identity. Start the recorder before interaction, retain the entire raw MKV and audible tail, then finalize the specific scope after the process ends. Follow each cell's natural/skip and ordinary/held-input procedure through its focused setup.

  5. Review paired browser video and exclusive audio. Open review/index.html using the selected local browser. Return results are imported into a fresh local review export for pairing; the initially empty viewer does not invent captures. Preserve actual speed and separate raw clocks, annotate offsets, playback drift, crop and uncertainty. Listen to only one source at a time. Require actual browser playback/seeking/audio checks for every preview codec. A codec name, probe success or file extension is insufficient. Retain raw MKV even if the browser cannot decode it. If a preview or extracted WAV is needed, use a selected existing local tool and retain exact executable/version/hash, inert argv, UTC derivation time, raw/output identities, timestamp mapping and crop/aspect changes. Never silently retime or discard the source. Record unsupported codecs or inaudible/missing content as open failures.

  6. Collect, postflight and audit. Finish each run with actual annotations, recording files, observer logs, settings, DxWnd exports and limitations. Include missed, skipped, aborted, crashed and incomplete runs. A process exit or captured means collection only. Original-DxWnd provenance is separate from native port receipts. Preserve generated originals and their post-run changes, copy-pre.json, retained-copy.json where applicable and copy-post.json; never restore them over the authoritative originals. Re-run immutable preflight after collection and preserve output. Actual review verdicts and shared schema/matrix validation remain separate gates.

  7. Return once using local transfer. Copy the complete checkpoint (including raw recordings, results and original-runs) by USB/local filesystem or existing local share to a new return directory. Do not upload assets/packages/media. Recheck manifest/READY SHA-256 and all immutable entries, then rehash all receipt-linked artifacts on return using the shared Python receipt validator. Preserve a source-side file/hash inventory for the mutable result and original-runs trees and compare it with the returned copy before importing. A mismatch or missing file is an explicit transfer failure. Do not overwrite the source copy, canonical originals, historical packages or an earlier return.

Matrix and required procedures

Case ID Lifetime Carl dialogue Physical input policy
fresh-natural-edge Independent fresh launch Natural completion Held through dialogue/RETURN
fresh-natural-ordinary Independent fresh launch Natural completion Ordinary play/release
fresh-skipped-edge Independent fresh launch Actual supported skip Held through dialogue/RETURN
fresh-skipped-ordinary Independent fresh launch Actual supported skip Ordinary play/release
retained-natural-edge Observed departure/return, same lifetime Natural completion Held through dialogue/RETURN
retained-natural-ordinary Observed departure/return, same lifetime Natural completion Ordinary play/release
retained-skipped-edge Observed departure/return, same lifetime Actual supported skip Held through dialogue/RETURN
retained-skipped-ordinary Observed departure/return, same lifetime Actual supported skip Ordinary play/release

Every cell requires separate continuous, dialogue-return, bank-left, bank-right and landing-idle obligations on both original-DxWnd and native Windows port. Continuous means ordinary lab → VR04 → lab → outside in one uninterrupted recording and process, including Carl dialogue, first neutral input, both banked jumps/landings, gadget use, RETURN cleanup and outside arrival. Focused clips retain their actual setup/prerequisite recording. A direct scene, debug warp or injected fixture cannot replace the route or retained prerequisite. Record keys, skip action and held/released boundaries as actually performed. Preserve legitimate landing bank and the implemented TSK lifetime ordering; do not insert a leveling reset or tune movement for the recording.

Collector commands and failure results

Replace paths and IDs with actual selections. Commands are examples for the later combined session; use a fresh unique ID for each independent launch. Start external recording when prompted and keep the collector alive for the entire retained setup and tested interval. The recorder/browser are identified, not automatically driven. The .cmd wrapper is an optional equivalent entry.

powershell.exe -NoProfile -File .\scripts\collect-fidelity-windows.ps1 -Action Run -Mode port -SessionId port-fresh-natural-ordinary-continuous-01 -CaseId fresh-natural-ordinary -Scope continuous -RecorderPath 'C:\Tools\OBS\bin\64bit\obs64.exe' -BrowserPath 'C:\Tools\Browser\browser.exe' -SettingsPath '.\results\session-settings.json'
powershell.exe -NoProfile -File .\scripts\collect-fidelity-windows.ps1 -Action Run -Mode original -SessionId original-retained-natural-edge-bank-left-01 -CaseId retained-natural-edge -Scope bank-left -DxWndPath 'C:\Tools\DxWnd\dxwnd.exe' -ProfilePath 'C:\Local\exported.dxw' -RecorderPath 'C:\Tools\OBS\bin\64bit\obs64.exe' -BrowserPath 'C:\Tools\Browser\browser.exe' -SettingsPath '.\results\session-settings.json'
powershell.exe -NoProfile -File .\scripts\collect-fidelity-windows.ps1 -Action Retained -SessionId original-retained-natural-edge-bank-left-01
powershell.exe -NoProfile -File .\scripts\collect-fidelity-windows.ps1 -Action Finish -SessionId port-fresh-natural-ordinary-continuous-01 -AnnotationPath '.\results\port-notes.json' -ArtifactsDirectory 'D:\LocalCaptures\port-run-01'
powershell.exe -NoProfile -File .\scripts\collect-fidelity-windows.ps1 -Action Skip -Mode original -SessionId original-missed-01 -CaseId fresh-skipped-edge -Scope bank-left -Reason 'Actual reason this obligation could not run'

Skip/Abort create a new terminal result without requiring recorder/browser selection. An interrupted active run retains its initial incomplete.json; when safe, collect its artifacts with Finish -Outcome aborted -Reason '...' instead of pretending a separate aborted launch describes its lifetime. When only incomplete.json survived, this command preserves its exact bytes as a diagnostic artifact and permits only aborted. End clocks remain absent; initial placeholder end values are discarded. Keep recovered media and source notes, with an empty events array because no complete launch interval was observed. The receipt remains open and importable; recovery does not establish missing executable identity, process completion, or a captured/reviewed run. Failed starts preserve run.json, failure.json and available logs. To collect failed/skipped/aborted media use Finish -Outcome failed|skipped|aborted matching the prior status, a reason and annotation/artifact directories (empty arrays are allowed for these outcomes). Never promote a failed or terminal run to captured; use a new independent session for a retry. Existing evidence is write-once.

Annotation field contract

Copy protocol/annotation-template.json to a mutable result note and fill it with actual values. Empty templates cannot finalize capture. All paths are relative, case-unique and contained below ArtifactsDirectory; no commands are read from annotations. Every captured run retains a raw .mkv video and actual audio evidence. If raw audio is embedded, retain a separately named byte-identical MKV copy with role audio, or a proven extracted WAV with derivation provenance. This preserves unique artifact paths while binding the actual audio bytes.

  • artifacts: entries {path, role, duration_seconds}; roles include video, audio, input_events, prerequisite, diagnostic. Preserve raw and preview as separate entries. Original prerequisite recordings/copy-state records and port observer/events.jsonl, summary, stdout/stderr must be included as appropriate; do not rename synthetic fixtures as native evidence.
  • events: {name, artifact, media_seconds, elapsed_seconds, method, uncertainty_seconds} using the exact required names/order in scenarios.json. Original methods are manual or procedural; markers describe visible input and event intervals, never invented original simulation ticks. Port events map observer data to actual media clocks; simulation_tick remains separate. Preserve observer launch_id, clock_id, requested/accepted input distinctions, sequence and possible truncation/overhead. Never reset elapsed clocks at scenes. All times are finite, nonnegative and within actual recording/run duration.
  • derivatives: {path, source, tool, version, argv, created_utc, timestamp_mapping, crop_aspect, browser_playback_verified}. The collector stores argv as inert provenance; it does not execute it. Select true browser playback only after observed playback. Hash tool binaries and retain source identity/settings in local session notes, including whether crop/time changed.
  • continuity: lifetime_id equals session ID, prerequisite_artifact names a hashed artifact, copy_pre_sha256 binds the actual copy-state record. Retained additionally needs departure_session_id and return_session_id equal to that same session ID and retained_copy_sha256 identifying the retained snapshot. prerequisite_departure and prerequisite_return markers precede the tested interval on the prerequisite artifact. Include these state records as hashed identities/artifacts, never made-up placeholders. The port's actual package/ startup-state record must be identified separately; missing continuity remains open rather than treating a second fresh launch as retained.
  • diagnostic: true for observation overhead, preflight, fixtures or other diagnostic limitations; limitations: actual missing tools, crop/audio/render/ timer/wrapper differences, incomplete transitions, uncertain markers and review failures. A successful collector does not clear them or acceptance automatically.

Finish prefixes local media/event/derivative references with media/ and hashes copied bytes. It never accepts an annotation-supplied acceptance verdict. The shared receipt validator and later full matrix/human review decide applicability. No universal timing, pixel, angle or audio tolerance is asserted by this agenda.