Files
Neutron/.planning/phases/01-reproducible-fidelity-baseline/01-05-SUMMARY.md
T

16 KiB

phase, plan, subsystem, tags, requires, provides, affects, tech-stack, key-files, key-decisions, requirements-completed, requirements-acceptance, coverage, actuals, plan_head_before, plan_head_after, duration, started, completed, status
phase plan subsystem tags requires provides affects tech-stack key-files key-decisions requirements-completed requirements-acceptance coverage actuals plan_head_before plan_head_after duration started completed status
01-reproducible-fidelity-baseline 05 testing
offline
powershell
windows
fidelity
sha256
phase provides
01-02 Offline paired review viewer
phase provides
01-03 Capability ledger and identified source links
phase provides
01-04 Passive port observation and package flags
Immutable offline checkpoint with identified original and Linux/Windows packages
Complete eighty-obligation Windows agenda with separate original-DxWnd and native-port collection
Package-free PowerShell selftests and truthful terminal/media results
01-06
01-07
added patterns
exclusive allocation
bounded hashing
atomic publication
explicit open acceptance
inert argument provenance
created modified
tools/recon/fidelity_checkpoint.py
scripts/prepare-fidelity-checkpoint.sh
scripts/collect-fidelity-windows.ps1
scripts/collect-fidelity-windows.cmd
tests/test_fidelity_checkpoint.py
tests/fidelity-windows.Tests.ps1
docs/evidence/windows-session.md
READY validates immutable prepared materials only; actual original-DxWnd/native Linux/native Windows acceptance remains open 0/120.
Runtime package source/dirty/diff identity stays bound to its actual completed publish, separately from later collector and document commits.
Selected external tools are identified without redistributing executables; raw MKV, terminal failures and actual derivation provenance remain local.
EVID-01
EVID-02
EVID-04
OPEN: assigned preparation contributions are complete; shared requirement IDs remain Pending until unfinished sibling capture/review plans and actual acceptance gates close.
id description requirement verification human_judgment
offline-checkpoint Identified complete offline bundle, preservation and tamper checks EVID-01
kind ref status
integration tests/test_fidelity_checkpoint.py pass
kind ref status
other artifacts/fidelity/verification/01-05-complete-bundle.json pass
false
id description requirement verification human_judgment
collection-tooling Package-free collector identity/path/argument/process/result functions on synthetic fixtures EVID-02
kind ref status
integration tests/fidelity-windows.Tests.ps1 pass
false
id description requirement verification human_judgment
combined-session-materials Eighty Windows obligations, session instructions and settings/annotation templates are bundled EVID-04
kind ref status
integration tests/test_fidelity_checkpoint.py#test_complete_session_materials_and_eighty_windows_obligations pass
false
id description verification human_judgment rationale
actual-original-native-acceptance Actual original-DxWnd/native Windows execution, audiovisual preflight and full scenario review
true Linux-hosted synthetic PowerShell tests, prepared bytes and apphost help do not establish Windows gameplay, actual browser/audio playback or whole-original fidelity; plan 01-07 owns the combined session.
tokens tasks commits basis
21862 2 7 ceil(chars/4) over the realized seven-file production/test/document diff: 87448 characters. Measured ledger range includes five plan task commits plus the intervening primary handoff and quota-guard commits; no invented task-commit count.
de68796df1 aa4b862a12 12h 4min 2026-11-13T04:40:08Z 2026-11-13 complete

Phase 01 Plan 05: Offline Windows Checkpoint Summary

An identified 10,983-file offline checkpoint carries preserved originals, both actual port packages, review/protocol materials and a complete Windows original-DxWnd/native-port collection agenda with open acceptance.

Performance

  • Two tasks, seven production/test/document files. The recorded wall interval starts at the earliest task RED commit, 04:40:08 UTC, and includes the long quota interruption before task 2 resumed. It is not twelve hours of active execution.
  • Final targeted verification: 24 Python checkpoint tests, 17 actual Linux-hosted PowerShell checks, four PowerShell-produced synthetic receipts validated by the shared Python validator, shell syntax and 539 unchanged original files.
  • The broader fidelity regression passed 133 tests in 123.784 seconds with authorized local loopback/process access before the final mode-only correction; the affected checkpoint suite then passed all 24 tests in 4.742 seconds. No skipped tests are represented as passes.
  • Actual full-bundle Python and Linux-hosted PowerShell preflights both pass, checking 10,983 immutable entries. The PowerShell traversal took several CPU-active minutes and completed before its explicit 600-second supervisory bound. It did not wait for interactive input.

Accomplishments and Interfaces

  • fidelity_checkpoint.py seals actual completed publishes, prepares a fresh offline tree and preflights immutable file identities. Existing outputs, overlaps, aliases, tampering, stale DLLs, interrupted/concurrent assembly and missing dependencies are rejected with local diagnostics. The generated original seed is verified before copying and again against its authoritative inventory. Hashes stream; assembly uses an exclusive reservation and retains incomplete diagnostics on failure.
  • The bundle contains manifest.json, READY.json, original-pristine/, original-runs/, port/linux-x64/, port/win-x64/, review/, protocol/, results/, both collector launchers, the session guide and standalone PowerShell selftests. Protocol materials include scenarios, capabilities, original inventory, an 80-obligation Windows agenda and explicit pending settings/annotation templates. Empty/false fields require actual observations during the later session; they are not fabricated identities or implementation stubs.
  • collect-fidelity-windows.ps1 supports Preflight, Run, Retained, Finish, Skip and Abort. Native port and original-DxWnd receipts keep distinct categories. It identifies actual selected external tools without copying third-party executables/installers, preserves the user's profile plus generated original pre/post state, uses quoted direct process invocation and records honest exit/log information. No original is executed in place, no command from a receipt is evaluated, and no dependency is downloaded.
  • Raw MKV and actual audio evidence are required to finalize captured results. Derived artifacts retain source/output references, actual tool/version/argv/time/crop metadata and explicit browser-playback declaration. Synthetic promotion, terminal-outcome promotion, reused results and invalid references are rejected. Failed/skipped/aborted results remain available for collection without inventing successful media. All collector outputs retain open acceptance and require the shared validator and later human review.
  • windows-session.md gives one ordered agenda: local selections/provenance, package-free checks, separate original/port visual/audio preflights, all eight matrix cells and five obligations, paired browser/exclusive-audio review, collection/postflight, then one local return transfer with hash checks. It recommends offline OBS Display Capture initially while requiring actual selected settings. It does not request capture now.

Task Commits

  1. Task 1 RED: 446930d — specify immutable offline checkpoint round trip.
  2. Task 1 GREEN: 6112339 — assemble identified offline Windows checkpoint.
  3. Task 2 RED: 818191c — require complete Windows session materials and open results.
  4. Task 2 GREEN: 8574665 — complete offline Windows collection agenda and selftests.
  5. Task 2 correctness follow-up: aa4b862 — retain executable permissions in offline copies.

Task 1 was already complete and was not repeated during recovery. The partially edited collector and checkpoint module were preserved and completed. Normal Git hooks ran with explicit scoped staging, protected-branch configuration and the supplied-root guard; no tracked files were deleted. Unrelated tools.yaml, the primary's quota tooling commit c9b9f2e and its ownership were preserved. The primary owns review, integration reports and push.

Actual Offline Bundle and Package Identities

Fresh final bundle: artifacts/fidelity/01-05-session-checkpoint. Durable identity/audit evidence: artifacts/fidelity/verification/01-05-complete-bundle.json.

Material SHA-256
manifest.json, 2,689,788 bytes 1098cf07030d5acf71b37d2132ec8bf39360855e9f7b2203cd130c9a7ac011be
READY.json, 358 bytes 1bb81a44da9f5ce92d5549f1608fcc6f46a47e2010d5160b39dcb471d1e8cccd
Original 539-file canonical tree 419af98a75a9f1d4ae3eceb5c7807bc9634912c2c3918c308470ad798dcef181
Linux sealed package file-list hash 124cf271e4952f3822eec5c24358447fcffc9e9e1803aded8d96d4d4472dc5a3
Windows sealed package file-list hash 335e5b75ec15c0fa904113d8b69d4f50e2cfa403f1f2f225353bd9e586322ef4
Linux GamePort.dll, 734208 bytes 145de80f2628f836cd77f0e8f66cb1b9d5abb8445088ff8ba8eb82527b6cd150
Windows GamePort.dll, 734208 bytes d21c917c1d260f3cd2c5b3d19397ab78ae23108c90b4f393eba4904d80e512a1

Both packages already existed under artifacts/fidelity/01-05-packages; recovery did not rebuild them. Their actual BUILD-INFO identifies source 446930d2a30d6d616c60ac5e59e4169a840bea09, dirty yes, diff SHA-256 90cef95669caece53fe2d2c4736cf2ac05ed27f5d25ca03ab572589686bcb73d. Publish timestamps are 2026-11-13T04:45:41.527524+00:00 and 04:47:05.457630+00:00 respectively. Neither the resumed collector revision nor the later clean task commit is stamped over that runtime provenance. Historical dist/ packages remain unchanged.

The final copied Linux apphost was permission-repaired from its byte-identical package source to mode 0755, then actually ran --help successfully. The generator now preserves ordinary source permissions for future copies while dropping privileged mode bits. That permission-only repair changes no manifest/file byte identity. Apphost help establishes startup/CLI availability only, not gameplay, sound, GPU or platform acceptance.

Verification Commands and Scope

python3 -c 'import sys,unittest; s=unittest.defaultTestLoader.discover("tests",pattern="test_fidelity_checkpoint.py"); assert s.countTestCases() >= 16; sys.exit(not unittest.TextTestRunner(verbosity=2).run(s).wasSuccessful())'
bash -n scripts/prepare-fidelity-checkpoint.sh
python3 tools/recon/inventory.py ../Original --verify docs/original-manifest.json
python3 tools/recon/fidelity_checkpoint.py preflight --checkpoint artifacts/fidelity/01-05-session-checkpoint

The cached .tools/powershell-check/root/opt/microsoft/powershell/7/pwsh is usable with XDG_CACHE_HOME, XDG_CONFIG_HOME and XDG_DATA_HOME set to task-local /tmp/gameport-pwsh-* directories. Its default home cache is read-only in the sandbox. No install/download was performed. Actual commands used that executable with -NoProfile -File tests/fidelity-windows.Tests.ps1 -FixtureOutput /tmp/gameport-01-05-ps-results and separately -NoProfile -File scripts/collect-fidelity-windows.ps1 -Action Preflight -Bundle artifacts/fidelity/01-05-session-checkpoint.

The 17 selftests exercise IDs/paths, hashes/copy identity, arrays/write-once JSON, argument quoting, an actual child with spaced arguments and exit 7, stdout/stderr, timeout/start failure, categories/terminal outcomes, manifest tampering and finish links. Their four captured/failed/skipped/aborted receipts all validate structurally through fidelity_receipts.validate_receipt; all four remain synthetic and unaccepted. Linux PowerShell is not Windows PowerShell 5.1 execution evidence. The exact native Windows selftest command is included in the future combined agenda.

TDD Gate Compliance

Both retained RED reports, artifacts/fidelity/tdd/01-05-01-red.json and 01-05-02-red.json, return RED_EVIDENCE_OK through the current classifier. Task 1's named target executes its intended missing preparation-API assertion. Task 2's named target completes assembly then fails the planned missing-session-material assertion; its separate missing status-API error remains honestly recorded. Both RED commits precede their GREEN commits by ancestry. No separate refactor was needed. Final targeted tests pass.

Deviations from Plan

  1. [Rule 1 — collector correctness] Real PowerShell execution exposed top-level empty arrays serializing incorrectly and asynchronous .NET completion values leaking into the process-result pipeline. Write-FidelityJson now uses ConvertTo-Json -InputObject; asynchronous completions are suppressed, output closure is bounded and final log size is checked. Standalone serialization/process/terminal tests pass; 8574665.
  2. [Rule 2 — runnable offline copy] Final bundle inspection found that byte-only copying dropped Linux apphost execution permission. The generator now preserves ordinary source mode bits, and a regression checks both mode and bytes. The generated apphost was repaired from its verified source and ran help; aa4b862. No source/package bytes or gameplay algorithms changed.

No new trust surface beyond the plan's selected-process/local-filesystem boundaries was introduced. The supplied threat mitigations remain applicable: no eval/elevation/downloads, explicit local selections, contained paths, hashes, immutable result IDs and categories that cannot imply acceptance. No implementation stub or skipped mandatory automated verification remains; actual Windows/human checks are intentionally assigned to the later combined session.

Issues Encountered

Quota interruption required a specific task-2 recovery from the existing RED commit and partial edits. An accidentally duplicated slow preparation was stopped; its incomplete directory and reservation remain under .01-05-complete-checkpoint* as diagnostics, and the final assembly uses a new root. No lock was stolen or failed partial tree promoted. The sandboxed broader regression denied loopback sockets and timed out .NET setup; its authorized local rerun passed all 133 tests. The optional full PowerShell preflight was CPU-active and bounded, then passed. No unresolved test failure is hidden by these environment adaptations.

Next Phase Readiness

The full bundle and session tooling are prepared. Plan 01-06 still owns actual native Linux capture, and plan 01-07 owns the combined original-DxWnd/native Windows session and audiovisual review. The real scenario audit is structurally ready with accepted 0 / required 120, all eight cases missing and no invalid receipts. Original launch/profile validity, native Windows execution, real browser playback/audio/crop and full-route fidelity remain open. This plan makes no new original-binary, REA, Wine, native Windows gameplay or whole-original compatibility claim.

Shared EVID-01/EVID-02/EVID-04 IDs remain Pending because requirements.ready-ids returns no ready IDs while sibling plans are unfinished. Root progress/tasks/decisions/compatibility integration and consumed-handoff cleanup belong to the primary.

Self-Check: PASSED

All seven created files exist. All five plan task commits are reachable on this branch, and evaluation-scope resolves the recorded plan commits. Final counted checkpoint/PowerShell checks, shared receipt validation, both full-bundle preflights, shell syntax, Linux apphost help and 539 original hashes pass. The persisted ledger measures seven commits before summary metadata, including the two intervening primary commits explained above. No generated package/media/original is staged, no tracked deletion occurred, and unrelated tools.yaml is preserved.