Archived 2026-11-08. This is the former implementation plan. It is kept for traceability, stable IDs, evidence, and completed history. Do not add new tasks here. The active task list is ImplementationPlan.md.

5DO — Implementation Plan¶

Snapshot: 2026-11-08 (feasibility revision) Current release: 1.6.0 Repository: https://github.com/purelogiccode/5DO
Branch: master
Status: 2,051 tests pass; build, format, and InspectCode are clean. The local collection now inventories 1,371 disc images, which collapse to 414 unique titles and a 378-title game test set. The new 3-way matrix matches 232 titles to 4DO and 246 to Phoenix (194 to both), with 0 loadability regressions, 5 titles where 5DO is ahead of 4DO, and 0 untested; the four no-video results are secondary media/homebrew, not retail game regressions. Tier B batches 1-4 reviewed 107 titles (81 Good, 26 Playable); 307 titles still have automated loadability only. The remaining Behind rows are 22 vs 4DO / 25 vs Phoenix, all explicit and documented. The 50-title accuracy panel keeps 50/50 frame hashes and 150/150 audio hashes at frames 60/120/300. The 594-disc legacy sweep remains 593 rendered, 1 known failure (Crash 'n Burn Japan). Full evidence: docs/CompatibilityComparison.md, docs/DifferentialHarness.md §8, docs/GameCompatibilityResults.md. Goal: beat 4DO and Phoenix on measured, player-visible compatibility, and remain the best standalone desktop 3DO emulator on Windows, Linux, and macOS.
Plan shape: three phases, one corpus, one comparison, four gates. Research, product, and speculative timing work are out of scope until the compatibility gate is green.

2026-11-08 feasibility revision. The previous plan carried 49 open IDs, eight gates, and an 800-850 developer-day backlog. That is not a plan; it is a wish list. This revision keeps every stable ID for traceability, but only nine backlog entries are allowed to block a release. Everything else is explicitly deferred or closed in §3.4, with a revisit condition. The full per-image differential is replaced by the 414-title unique corpus; the open-ended timing audit is replaced by a time-boxed residual that must show a game-visible effect to stay open.

This was the single source of truth until 2026-11-08. It replaced the former dashboard, milestone log, feature spec, parity audit, and session handoff. It is now historical; the active task list is ImplementationPlan.md.

1. How to use this plan¶

1. Work the phases in order. Phase 1 is the only compatibility phase; do not start Phase 2 before every unique title has a status, and do not start Phase 3 before the peer comparison has no unexplained regression.
2. The unique corpus is the test set. Compatibility claims are made on the 378-title game test set, not on 1,371 regional copies. The image inventory is for loadability and duplicate accounting only.
3. One feature per change, tests first. Every behavior change needs a unit, golden, differential, or scripted test. State the reference: Opera, 4DO, hardware documentation, or a 5DO decision.
4. Evidence before checkbox. Mark an item done only when its "Done when" line is satisfied and the evidence is committed or linked.
5. No work without a player-visible outcome. If a task cannot change a game's status, a gate, or a release artifact, it belongs in §3.4, not in the backlog.
6. Timing work is conditional. Open CORE-02/CORE-03/CORE-06/CORE-07 only when a matrix regression or a panel divergence points at that subsystem. A trace difference with no game-visible effect is documentation, not a release blocker.
7. No default timing change without panel evidence. A 40-title, 300-frame sample is a smoke test, not a gate.

Legend¶

• [ ] — remaining work. [x] is used only in the completed log.
• Pending — not started, or research/design only.
• Partial — a usable subset ships; the remaining scope is listed in the item.
• Conditional — work only if the measurement panel proves it changes a game-visible outcome.
• Deferred — explicitly out of scope for this plan; see §3.4.
• Closed — a recorded decision not to pursue the work; it is not "done".
• Priority — P0 correctness/compatibility, P1 release evidence, P2/P3 post-compatibility.
• Effort — S under 3 developer-days, M 3-10, L 10-30, XL 30-90+.

2. Current state¶

2.1 Verified snapshot (2026-11-08)¶

The corpus and comparison numbers come from artifacts/corpus/inventory.json and docs/data/compatibility-comparison.json; the J: corpus is not mounted in every session. Re-run §7 before quoting them.

2.2 What "better than 4DO and Phoenix" means¶

"Better" is not a feature list. It is measured on the same BIOS, same unique titles, and same input scripts:

A Good title boots, reaches gameplay, survives 10 minutes of scripted input without a hang or crash, and has no visible or audio defect a player would notice. Playable means it runs but has a documented defect. Unplayable means it does not reach or sustain gameplay.

2.3 Shipped baseline¶

These are done and must not regress. They are listed here once; the backlog contains only missing work.

Scope decisions (unchanged): WebAssembly/browser, Android/iOS, Raspberry Pi, and Steam Deck are out of scope. Desktop Windows, Linux, and macOS (x64/arm64) remain the supported targets. 4K internal CEL rendering is research, not a release blocker; 4K output upscaling is supported. Distribution remains free, open-source donationware under the FreeDO modified GNU LGPL (non-commercial; voluntary donations only).

2.4 Evidence gaps that block the claim¶

1. Loadability is not playability. The matrix has 107 manual gameplay grades and 307 automated loadability results, with 0 untested; the four No video rows are secondary media/homebrew. Tier B's ~100-title target is met; the grade gap is now the work.
2. The peer comparison is now explicit, not unknown. After Tier B batches 1-4, 107 titles have a manual grade and 22 rows remain Behind vs 4DO / 25 vs Phoenix. Every one has a recorded reason (lightgun menu, media disc, instructional disc, or a missing title-specific script); none is an unexplained loading failure.
3. The CD cadence residual is still open. Eleven 50-title panel titles diverge in command cadence at frame 300; video and audio still match. It is a diagnostic difference until a title shows a player-visible effect.
4. Platform evidence is asymmetric. Windows and Linux are validated; macOS is documented best-effort. That is acceptable for the compatibility claim but must be stated in the release notes.

Honest current claim: 5DO is a better product than 4DO, a credible challenger to Phoenix, and — on the tested panel — video/audio-accurate. It is not yet a proven playability win. Phase 1 exists to settle that.

2.5 Completed IDs (traceability only)¶

Full completed-ID index (including the nine above): FND-01, FND-02, FND-03, CORE-09, CORE-12, CORE-14, CORE-15, SYS-04, SYS-06, SYS-07, SYS-08, SYS-09, SYS-10, SYS-11, SYS-12, SYS-13, SYS-14, VID-04, VID-05, VID-06, VID-14, VID-15, AUD-05, TEST-07, TEST-08, UX-01, UX-02, UX-04, UX-11, UX-12, UX-15, UX-19, UX-23, UX-26, ENG-03, PERF-02, VID-12, BAK-12, BAK-17.

Milestone aliases: M9R/M9P → P3/P6/TEST-04; M10 → PERF-02; M12 → PERF-08; M13.5/M13.6 → TEST-06; M14 → VID-01/02/03; M15 → CORE-11/16 + FND-05.

3. Missing features — canonical backlog¶

Each entry is a single feature or a tightly-scoped bundle. No feature appears in more than one category. The backlog is deliberately short: everything that does not block Gates A-D lives in §3.4.

3.1 P0 — Playability proof¶

• TEST-06 — Playability audit on the unique corpus Status: Partial · Priority: P0 · Effort: L Remaining:
• Tier A (all 378 unique games): done. Every unique title has a 5DO status; the initial pass had 371 automated loadability results and 43 manual grades, with 0 untested. The 44 previously untested images now have a status; the four No video results are secondary media/homebrew, not retail games.
• Tier B (target ~100 titles): met. 107 titles have a 10-minute scripted pass and a checkpoint capture; 307 titles remain automated loadability only.

• Tier C (top 50): full gameplay review to Good/Playable/Unplayable with a reason and evidence. Done when: all 378 games have a reviewed status; Tier B has zero unexplained regressions; Tier C is ≥95% Good/Playable; the matrix and per-game database are updated. Evidence (2026-11-08 batch 1): the ten comparison diffs were run through a 10-minute 5DO-only scripted pass (artifacts/tierb-behind-clean, benchmarks/diff/tierb-behind.json). Five reached and sustained gameplay (Battle Chess, Crash 'n Burn, Myst, Olympic Summer Games, Orbatak) and were promoted to Good; the other five stay Playable pending title-specific scripts (Crime Patrol, Mad Dog II, Guardian War, Shootout at Old Tucson, Slayer). Evidence (2026-11-08 batch 2): 32 titles where both 4DO and Phoenix claim Playable/Good were reviewed with --bios-auto; 25 were promoted to Good and 7 kept Playable (20th Century Almanac, BladeForce, Carrier, Club 3DO, DeathKeep, Dennis Miller, Digital Dreamware). The ledger is docs/data/playability-grades.csv. Evidence (2026-11-08 batch 3): 32 more peer-flagged titles were run with --bios-auto; 24 were promoted to Good and 8 kept Playable (ESPN Golf, Family Feud, Hell, Insect War, John Madden Football, Jurassic Park Interactive, Kingdom, Life Stage). An 18,000-frame audio-hash pass produced non-zero hashes and dspRunning on all 32. The comparison moved to 75 reviewed titles (54 Good, 21 Playable), 339 automated loadability only, and 17/20 explicit Behind rows. See TestCorpus.md §1.1bk. Evidence (2026-11-08 batch 4): the last 32 titles in the peer-flagged set were run with --bios-auto; 27 were promoted to Good and 5 kept Playable (N.O.B., Off-world Interceptor, Olympic Soccer, PGA Tour 96, Return Fire - Maps O' Death). Title-specific tierb-panzer-start and tierb-quarterback-start scripts confirmed Panzer General's scenario map and Quarterback Attack's kickoff/football gameplay. An extra capture pass confirmed Real Pinball's dark late frame is an attract/transition view, not a render failure. The comparison now has 107 reviewed titles (81 Good, 26 Playable), 307 automated loadability only, and 22/25 explicit Behind rows — more evidence, not a regression. See TestCorpus.md §1.1bl. Why first: this is the only evidence that answers "better than 4DO and Phoenix" for a player. The old 594-title "renders" count does not. Ref: docs/CompatibilityComparison.md; docs/Compatibility.md §10/§13.

• CORE-13 — Matrix-driven compatibility fixes Status: Partial · Priority: P0 · Effort: L Remaining: review every title that CompatibilityComparison.md marks Behind; fix or re-grade it with evidence; fix Crash 'n Burn (Japan); retry any Tier A No video that is a retail title; add a regression test or matrix note for each change. Done when: the matrix has no unexplained regression; every remaining exception has a root cause and a documented workaround; Tier B is clean. Rule: do not add speculative per-title fixes. Every fix starts from a matrix row. Ref: References/4DO-master/FourDO/Emulation/GameSource/GameDatabase.xml; docs/GameCompatibilityResults.md.

• CORE-11/16 + FND-05 — Regression harness on the unique corpus Status: Partial · Priority: P0 · Effort: M Remaining: keep the 50-title panel green; add the 378-title loadability set to the weekly run; keep the image inventory as the duplicate/region record; publish a nightly dashboard with frame/audio/CD fields; document every allowance with a date. Done when: the nightly run has no unexplained divergence, and every allowance is documented and dated. Rule: a frame-hash match without PC/CD/audio evidence is a smoke result, not a pass. Evidence (2026-11-08): the audio-gated 50-title panel (test93-panel50-audio/report.json) captures 50/50 with 0 errors, 50/50 frame hashes, and 150/150 audio hashes at 60/120/300; it leaves 39/50 phase-only and 11 CD-only first behavioral divergences at f=300. The same-frame retired-instruction delta is 11-20 at f=60, -46..2,164 at f=120, and -36,104..3,591 at f=300; 50/50 first PC/CPSR divergences are same-frame phase. See DifferentialHarness.md §8.

3.2 P0 — Accuracy, time-boxed¶

• CORE-04 — CD-ROM command and timing residual Status: Partial · Priority: P0 · Effort: M (time-boxed) Remaining: close the 20th Century guest read-schedule residual or document it as a guest-scheduling difference with no video/audio impact; keep the command/status/metadata transcripts green; verify CD-DA and multi-disc only on matrix titles that need them. Done when: the 50-title panel has no video/audio divergence; every remaining CD cadence difference is documented with its first mismatch and its player-visible impact; no 594-sweep regression. Why time-boxed: the command table, status bytes, READ_ID payload, sector-time formula, FIQ delivery, and timer period already match. The residual is guest software pacing, not a CD register-contract bug. Evidence (2026-11-08 increment 4): the 5DO timer trace now emits the reference's FIQ/XBUS event kinds (fiqgen, fiqwrite, xbussel, xbuspoll, xbusfiq with source 1-5). On 20th Century both engines read the active-FIQ register exactly 99 times with identical masks. 5DO leaves the OS idle loop on the VBlank FIQ at 59,021,271 and issues READ_DATA at 59,063,402; Opera stays idle until READ_ID at 70,080,576. The periodic READ_ID cadence matches within +0.21M cycles. The residual is the guest file-system read-request schedule. 2,051 tests pass. Ref: opera_cdrom.c; docs/ReferenceStrategy.md §6.

• CORE-02 / CORE-03 / CORE-06 / CORE-07 — Conditional timing evidence Status: Conditional · Priority: P1 · Effort: S each Remaining: open only when a matrix regression or a panel divergence points at CLIO timers, CEL busy, SPORT/XBUS/DMA, or FIQ/IRQ ordering. Record the trace, the game-visible effect, and the fix. Otherwise leave the subsystem at its measured-parity state and cite the evidence in DifferentialHarness.md §8. Done when: each item is either closed with reference evidence or promoted to P0 with a named title that changes status. Evidence: timer 14 is 12,384 cycles in Opera and 12,396 in 5DO; the Lemmings READ_ID-to-READ_ERROR residual is guest copy/decrypt work; the XBUS CD FIQ is level-held like the reference. CORE-01 is closed (see §3.5).

• AUD-01 — Audio output quality, reduced scope Status: Partial · Priority: P1 · Effort: S Remaining: one 10-minute real-time soak on Windows and Linux with the null device and one real device; confirm 0 dropped samples/underruns; keep the existing sync-offset and device-recreation tests. Done when: the soak is clean on both platforms and the result is recorded in Audio.md. Evidence: AudioSyncMilliseconds applies a fixed host offset; OpenAlAudioOutput recreates output safely and exposes drop/underrun/latency counters; a 240-frame AOT null-device run reported 0 drops and 0 underruns.

3.3 P1 — Release blockers (not accuracy blockers)¶

• ECO-04 — Windows and Linux release validation Status: Partial · Priority: P0 for release · Effort: M Remaining: Windows x64/arm64 bundle smoke; Linux x64 (Hyper-V LinuxMint) full suite, CLI smoke, frontend --smoke-test, and a 3,000-frame disc boot; macOS x64/arm64 build validation with GPU presentation documented as best-effort. Done when: Windows and Linux pass the release gate; macOS builds and is documented; the six bundles are published with SHA256SUMS.txt. Note: this blocks the release gate, not the compatibility claim.

• SYS-03 / SYS-15 / SYS-01 / SYS-02 — Conditional disc and input coverage Status: Conditional · Priority: P1 · Effort: S each Remaining: run only the matrix titles that need PAL, lightgun/arcade/flightstick/trackball input, CD-DA, or multi-disc. Record the title, BIOS, input script, and result. Do not build a speculative full-library program. Done when: every matrix title with that requirement is verified or documented as a known limitation.

• ENG-01/09 — Minimal release CI Status: Partial · Priority: P1 · Effort: S Remaining: keep dotnet build, dotnet test, dotnet format --verify-no-changes, and jb inspectcode --severity=WARNING green on Windows; run the docs JSON --check; publish the six bundles and checksums. SBOM and reproducible-build work are release-note items, not blockers. Done when: every release candidate passes the four commands and publishes the bundles.

3.4 Deferred and cut (not release blockers)¶

These items are explicitly not part of this plan. They stay in the repository's history and stable-ID index; they do not block Gates A-D. Reopen only under the stated condition.

3.5 Closed decisions (not backlog)¶

These are recorded decisions, not completed work. Do not reopen them without new measurement.

4. Suggested execution order¶

This is the phased order. The IDs are the canonical entries in §3.

If you only do three things: TEST-06, CORE-13, and CORE-11/16. They produce the playability proof, the fixes that proof drives, and the regression signal that keeps both honest.

Rule: do not start Phase 3 while any Phase 1 P0 item is open. Do not start §3.4 work at all until Gate B is green.

4.1 Effort reality¶

By the plan's own S/M/L/XL scale, the blocking path through Gate D is roughly 60-80 developer-days, down from 240-850 in the previous plan. The savings come from the unique corpus, the folded evidence items, and the explicit cut list — not from lowering the bar.

• A 378-title loadability pass is ~1 hour at the current CLI speed.
• A Tier B 10-minute scripted pass is ~2-3 hours of emulated runtime for ~100 titles, plus visual review.
• Tier C is 50 titles × review; budget 1-2 weeks even with automation.
• The capture, audit, and hardware half does not compress with AI assistance.

A realistic wall-clock target for Gate D is 8-12 weeks at the project's recent AI-assisted velocity, with Tier C review as the bottleneck. The deferred list is intentionally large; it is the post-compatibility product roadmap, not this plan.

5. "Best 3DO emulator" gates¶

These are the release/quality gates for the goal, not a second backlog. They reference canonical IDs only. A gate is green only when the evidence is committed, reproducible, and dated.

Release rule: Gates A-D must be green. There is no Gate 5-8. Performance, platforms beyond Windows/Linux, product, and research are not release blockers for the compatibility claim; they are post-compatibility roadmap items.

6. Reference strategy¶

6.1 Primary and secondary references¶

Reference hierarchy: Opera wins on timing where it is measurable; 4DO wins on known game-specific fixes; hardware documentation wins when the two disagree. RMG and rpcs3 are product inspiration, not accuracy references.

6.2 Rules¶

1. Port behavior, not code. 5DO is a clean C# implementation; do not copy C/C++ source.
2. One reference per change. Record the source in the commit/PR and in docs/PortingNotes.md if the behavior diverges.
3. Tests first. Add the failing test or differential fixture before the fix.
4. Preserve determinism. No wall clock, unordered iteration, or environment-dependent seeds in emulated state.
5. Keep the core single-threaded by default. Threaded work must be opt-in and bit-identical; CORE-05 is closed.
6. Do not anchor to Opera's bugs. Opera is a reference, not ground truth. If the hardware documentation is clear, follow the hardware and record why.
7. Do not ship copyrighted BIOS/game content. Embedded boot ROM work is closed (SYS-05); retail titles still need a user-supplied BIOS.
8. Document intentional divergence in docs/PortingNotes.md and docs/Compatibility.md.

7. Verification commands¶

Run these before marking work done.

# Build and test
dotnet build 5DO.sln -c Debug
dotnet test 5DO.sln -c Debug --no-build --no-restore
dotnet format 5DO.sln --verify-no-changes --no-restore

# Static analysis
jb inspectcode 5DO.sln --output=inspectcode.sarif --format=Sarif --sEverity=WARNING --no-build

# Unique corpus inventory and 3-way comparison (TEST-08)
python tools/corpus-inventory.py --jobs 8 --out artifacts/corpus/inventory.json
python tools/generate-compat-comparison.py --inventory artifacts/corpus/inventory.json --sweep artifacts/compat-full/summary-merged.csv --sweep artifacts/compat-unique/summary.csv --reference docs/GameCompatibilityList.md

# Compatibility sweep (full images for loadability; unique test set for the matrix)
pwsh -NoProfile -File tools/compat-sweep.ps1 -Roots "J:\Panasonic 3DO","G:\MAME\MAME Software List CHDs\3do" -Frames 3000 -Resume
pwsh -NoProfile -File tools/compat-sweep.ps1 -FileList artifacts/corpus/unique-testset.txt -Frames 3000 -Resume -OutDir artifacts/compat-unique

# Differential smoke and showcase (current oracle)
dotnet run --project tools/FiveDo.DiffHarness -c Release -- --manifest benchmarks/diff/smoke.json
dotnet run --project tools/FiveDo.DiffHarness -c Release -- --manifest benchmarks/diff/showcase.json
dotnet run --project tools/FiveDo.DiffHarness -c Release -- --manifest benchmarks/diff/showcase.json --compare-pc --core artifacts/diff/opera-build/out/OperaCore.dll

# After TEST-07: the 50-title accuracy panel
dotnet run --project tools/FiveDo.DiffHarness -c Release -- --manifest benchmarks/diff/panel50.json --strict --compare-pc --core artifacts/diff/opera-build/out/OperaCore.dll

# Benchmark gate
dotnet run --project benchmarks/FiveDo.Benchmarks -c Release -- --frames 120 --label ci --json --gate

# GPU probes (Windows)
dotnet run --project src/FiveDo.Frontend.Avalonia -c Release -- --gl-probe --gl-probe-backend wgl --gl-probe-capture
dotnet run --project src/FiveDo.Frontend.Avalonia -c Release -- --d3d-probe artifacts/d3d-probe/parity.json --d3d-probe-size 320x240 --d3d-probe-frames 20 --d3d-probe-warmup 5 --d3d-probe-filter accurate --d3d-probe-capture artifacts/d3d-probe/parity.png
dotnet run --project src/FiveDo.Frontend.Avalonia -c Release -- --vk-probe

# Frontend smoke
dotnet run --project src/FiveDo.Frontend.Avalonia -c Release -- --smoke-test

# Frontend fullscreen launch evidence (marker records windowState=FullScreen)
dotnet run --project src/FiveDo.Frontend.Avalonia -c Release -- --smoke-test --fullscreen --smoke-output artifacts/frontend-fullscreen-smoke.txt

Differential rule: attach the report (manifest hash, BIOS hash, disc hash, frame count, build) to any accuracy claim. A frame-hash match alone is not evidence of accuracy.

Linux validation: use the local Hyper-V VM LinuxMint (D:\Hyper-V\VMs\LinuxMint). Read AGENTS.local.md for credentials. Publish linux-x64 on the host and copy the output into the guest, or install the SDK in the guest. Record evidence in docs/Compatibility.md §10.3 and docs/KnownIssues.md.

Release: run the commands in ReleaseChecklist.md; build bundles with packaging/package-bundle.ps1; keep a local copy in src/FiveDo.Frontend.Avalonia/bin/Release; remove PDB/XML files; publish SHA256SUMS.txt.

8. Maintenance rules¶

1. This file is the single source of truth. Do not recreate separate plan, backlog, handoff, or roadmap files.
2. One canonical entry per feature. Do not add a second checklist for the same work. Reference IDs instead.
3. Retire duplicates. BAK-* and MOON-* are aliases only; see Appendix B. Do not reintroduce them.
4. Check the box only when "Done when" is satisfied. Include behavior, tests, docs, and no regressions.
5. Move completed items to §2.5 and the completed log. Do not leave them in the missing-feature category.
6. One feature per pull request. No mixed accuracy/rendering/UX changes.
7. No default timing change without panel evidence. A 40-title, 300-frame sample is a smoke test, not a gate.
8. Keep evidence current. Update TestCorpus.md, Compatibility.md, PortingNotes.md, and KnownIssues.md when behavior changes.
9. Keep the plan readable. Keep §§1-8 lean; the completed log is append-only history and the appendices are reference material. Put long evidence in the linked docs; keep the decision and the number here.
10. Record closures. A "won't do" decision needs a measurement and a revisit condition, like §3.5.
11. Refresh the snapshot monthly or after every release: version, test count, compatibility counts, performance numbers, and platform validation status.
12. Keep the cross-document anchors stable. #2-current-state, #3-missing-features-canonical-backlog, #4-suggested-execution-order, #7-verification-commands, and #9-completed-log are linked from NextSession.md.

9. Completed log¶

Append new entries here; do not delete history. Each entry is a completed canonical item or a release milestone.

• 2026-11-08 - TEST-06 Tier B batch 4 (last 32 peer-flagged titles). Generated benchmarks/diff/tierb-panel-4.json and ran the 36,000-frame --bios-auto pass. All 32 completed with exit code 0 and dspRunning: true; 27 promoted to Good and 5 kept Playable (N.O.B., Off-world Interceptor, Olympic Soccer, PGA Tour 96, Return Fire - Maps O' Death). Title-specific tierb-panzer-start and tierb-quarterback-start scripts confirmed Panzer General's scenario map and Quarterback Attack's kickoff/football gameplay; a 5-frame Real Pinball recheck showed the full table at frame 30,000 and confirmed the dark frame is the attract/transition view. The comparison now has 107 reviewed titles (81 Good, 26 Playable), 307 automated loadability only, 22 Behind vs 4DO / 25 vs Phoenix, and 5 titles where 5DO is ahead of 4DO. The higher Behind count is the explicit grade gap, not a loadability regression. See TestCorpus.md §1.1bl.

• 2026-11-08 - TEST-06 Tier B batch 3 (32 peer-flagged titles). Generated benchmarks/diff/tierb-panel-3.json and ran the 36,000-frame --bios-auto pass. All 32 completed; 24 promoted to Good and 8 kept Playable (ESPN Golf, Family Feud, Hell, Insect War, John Madden Football, Jurassic Park Interactive, Kingdom, Life Stage). An 18,000-frame audio-hash pass produced non-zero hashes and dspRunning on all 32. The comparison now has 75 reviewed titles (54 Good, 21 Playable), 339 automated loadability only, and 17/20 explicit Behind rows. See TestCorpus.md §1.1bk.

• 2026-11-08 - TEST-06 Tier B batch 2 (32 peer-flagged titles). Generated the next batch with tools/build-tierb-manifest.py (both 4DO and Phoenix claim Playable/Good, 5DO had only loadability). All 32 completed a 36,000-frame --bios-auto pass; 25 were promoted to Good and 7 kept Playable with recorded reasons (20th Century Almanac, BladeForce, Carrier, Club 3DO, DeathKeep, Dennis Miller, Digital Dreamware). An 18,000-frame audio-hash pass produced non-zero hashes and dspRunning on all 32; 13,243,230 samples each (Defcon 5: 15,876,000). Added the Tier B grade ledger docs/data/playability-grades.csv and the contact-sheet tool. The comparison now has 43 reviewed titles (30 Good, 13 Playable) and 10/12 explicit Behind rows. See TestCorpus.md §1.1bj.

• 2026-11-08 - TEST-06 Tier B batch 1 (10 comparison diffs). Ran the ten titles where 5DO differed from 4DO or Phoenix through a 5DO-only 36,000-frame scripted pass (tools/playability-sweep.py, benchmarks/diff/tierb-behind.json, benchmarks/diff/scripts/tierb-10min.txt) with captures at 600/3,600/18,000/36,000. Five titles reached and sustained gameplay and were promoted to Good: Battle Chess, Crash 'n Burn, Myst, Olympic Summer Games, and Orbatak. The five promoted titles were re-run on stock BIOS behavior with 13,243,230 samples and non-zero audio hashes; Crash 'n Burn's 18,000-frame WAV is non-silent (peak 7,631, RMS 1,190, 0 clipped). Crime Patrol, Mad Dog II, Guardian War, Shootout at Old Tucson, and Slayer remain Playable pending title-specific scripts. The comparison moved from 8 Behind vs 4DO / 10 vs Phoenix to 4 vs 4DO / 5 vs Phoenix, with 0 loadability regressions. See TestCorpus.md §1.1bi.

• 2026-11-08 - TEST-08 unique-corpus inventory and 5DO/4DO/Phoenix comparison. Added tools/corpus-inventory.py, which fingerprints every local image with the 5DO gameId (MD5 of the first two 2048-byte sectors) and reads the database metadata, and tools/generate-compat-comparison.py, which collapses identical images, groups regional revisions, chooses a canonical disc, parses both saved compatibility lists, and writes the comparison. On the current J:/G: collection the inventory finds 1,371 images, 653 distinct game IDs, and 414 unique titles, including a 378-title game test set; 232 titles match 4DO and 246 match Phoenix (194 both). A 44-disc gap-fill pass with -AllowUnsigned classified the remaining titles: 40 rendered, 4 no video (all secondary media/homebrew), so 0 titles are untested. The matrix finds 0 loadability regressions, 2 titles where 5DO is ahead of 4DO (Orbatak and Shootout at Old Tucson), 8 Behind rows vs 4DO, and 10 vs Phoenix (all manual Playable vs peer Good, not failures). docs/CompatibilityComparison.md, docs/data/unique-corpus.csv, and docs/data/compatibility-comparison.json are committed. tools/compat-sweep.ps1 gained -FileList and -AllowUnsigned for the unique test set and homebrew images. 2,051 tests pass; build and format are clean.

• 2026-11-08 - CORE-04/CORE-07 guest wake-source parity and 20th Century read schedule. The 5DO --clio-timer-trace now emits the reference debug core's FIQ/XBUS events: event=fiqgen (CD bit-4 edge only), event=fiqwrite for 0x40-0x6c, event=xbussel/event=xbuspoll, and event=xbusfiq with the reference source codes (1 tick, 2 poll tick, 3 built-in command, 4 device command/poll, 5 built-in poll); two new ClioTests and an XbusTests tick-source test cover the names and the edge rule. The probe (artifacts/diff/probe-20th-wake, 20th Century, 290 frames) shows both engines read the active-FIQ register exactly 99 times during the 57.8M-59.0M wait with identical masks (38x 00000001, 33x 80000001, 23x 80000801, 5x 80000003). The guest is in the Portfolio OS idle loop at 0x205f4..0x20600; 5DO leaves it on the VBlank FIQ at 59,021,271 and issues READ_DATA at 59,063,402, while Opera stays idle until its READ_ID at 70,080,576. On 900-frame captures (artifacts/diff/probe-20th-wake-900) the periodic READ_ID cadence then matches (Opera 70.08M/82.59M/95.11M/107.62M/120.13M/132.64M; 5DO 70.29M/82.80M/95.31M/107.83M/120.34M/132.85M, +0.21M phase) and the divergence is 5DO's READ_DATA bursts between polls (59.06M/59.12M/61.56M/64.07M/66.57M/69.07M). The READ_ID payload is identical (83 00 10 00 01 00 00 00 00 00 00 e3). FIQ delivery, masks, timer scheduling, SPORT, the command table, status bytes, and the ID payload are therefore not the residual; the guest file-system read-request schedule is. The next measurement is the FS read request and the CD task wake queue. 2,051 tests pass; build and format are clean.

• 2026-11-08 - CORE-04 CD register trace and status/metadata parity. Added reference-side recording of every CD command byte, poll write/read, status read, and data read with cycle and PC (sg_debug_cd_register_*), harness --cd-register-trace output (fivedo.cdreg.txt/opera.cdreg.txt), and 5DO datr tracing for non-empty data reads. On 20th Century Video Almanac (290 frames, unbounded speed) the command packets, status bytes, and 2048-byte sector data are identical through command 105, the last command before the 5DO READ_DATA/Opera READ_ID sequence divergence. The trace exposed four reference divergences now fixed: the XBUS 2x status bit (MEI_CDROM_DOUBLE_SPEED is 0x80, and cdrom_mode_reset starts at double speed; the guest mode byte is stored separately and re-applied on inject); EffectiveSpeed follows cdrom_effective_speed (unbounded stays unbounded, configured >1x drops to 1x while the guest selects single speed); the readable-disc limit is the 3DO volume header's volume_block_count (sector 0 offset 0x50), not the physical image size (20th Century: 327,168 declared vs 327,320 physical); and the TOC control includes the Q-position bit plus the 8-byte READ_SESSION_INFO payload. The CD-ROM save-state section is version 5 for the guest speed mode and readable limit. The remaining 20th Century wait is guest sleep time (6 frames vs 60) with matching CD values, so the residual is not the CD register contract; the next measurement is the wake source (FIQ/timer/SPORT). The 11-title cd-cadence panel keeps the same first mismatch indices. The confirmation panel (test110-panel50-metadata, 50 titles, 300 frames) captures 50/50 with 0 errors, 0 frame-hash divergences, 150/150 aligned audio fields, 39 phase-only titles, and the same 11 CD-only first behavioral divergences at f=300. 2,048 tests pass; build and format are clean. The local InspectCode host crashed before emitting a result in this session; the prior complete run was clean. Evidence: artifacts/diff/test104-20th-cdreg3 through test110-panel50-metadata; DifferentialHarness.md §8; PortingNotes.md 12.19; TestCorpus.md 1.1bf.

• 2026-11-08 - TEST-07 audio gating, scripted input, and the nightly baseline. The differential harness now gates the cumulative audio hash at aligned sample counts, reports a count difference as cumulative hash not comparable ... (reference 16.16 field-rate quantization, 1.214 ppm), and adds audio aligned/lag/diverged summary rows plus an HTML card. The alignment probe (audio-probe.json, Lemmings, 3,000 frames) measured 5DO's exact 60000/1001 rate against OperaCore's 3,928,227/65,536 (59.939987 Hz, 1.214 ppm lower): the sample counts match through 1,000 frames, then the reference is +1 at 2,000 and +2 at 3,000 (2.68 predicted). The 50-title panel (test93-panel50-audio/report.json) captures 50/50 with 0 errors, 50/50 frame hashes, and 150/150 audio hashes at 60/120/300; the 11 CD-only first behavioral divergences at f=300 remain. Scripted input is wired through both engines: test94-samples5-input matches 5/5 frame hashes at 60/300/600, the generic menu script changes guest output at 600/900, the no-input control matches the same checkpoints, and Road Rash's f=600 divergence reproduces without input (test95-roadrash-noinput). The nightly workflow now captures fresh (--force), uploads report.html, and fails only on a 5DO field change against the checked-in benchmarks/diff/baseline/showcase-report.json; baseline --write generates the trimmed baseline. 2,040 tests pass; build, format, and InspectCode are clean. Moves TEST-07 to completed.

• 2026-11-08 — UX-27 Tooltip coverage, host logging hardening, XML-doc enforcement, and UI test expansion. Every window in FiveDo.Frontend.Avalonia now exposes ToolTip.Tip on all interactive controls and on the information labels that previously had none: DiscInfoWindow (all 24 fields plus Close), DiscBrowserWindow field labels/values, GameMetadataWindow (developer, publisher, genre), MovieEditorWindow (11 pad buttons), CheatsWindow syntax help, DiscPlaylistWindow status, ScreenshotCompareWindow captions and panes, MainWindow library/icon/log headers, and the BIOS/NVRAM tab strip. A new FrontendToolTipTests parses every Views/*.axaml as XML and fails if an interactive element (Button, TextBox, ComboBox, Slider, CheckBox, ListBox, TabControl, MenuItem, …) has no ToolTip.Tip on itself or an ancestor, so the requirement is enforced by CI. Logging: added the missing Debug events to fallback/cleanup catch blocks (Direct3D/Vulkan surface initialization, cover resolution/download, DAT reads, void-color and refresh-rate fallbacks, random-seed parsing), guarded App.Initialize and the donation/update dialog ShowAsync entry points with try/catch + log, and kept every Warning+ event in the bug-report sink with the required environment/error/exception blocks. Naming/visibility: CoverCache.LoadAsyncCore→LoadCoreAsync and OnCompareClick→OnCompareClickAsync (the only production async methods without the Async suffix; Main is the framework exception), Program properties and AsyncEvent.Run are now internal to match their internal containing types. Documentation: GenerateDocumentationFile is enabled for FiveDo.Cli and FiveDo.Frontend.Avalonia, closing the last five CS1591 gaps. Dead code: InspectCode can be run at WARNING with 0 results; a reference scan of the two host projects found no unreferenced members. Added 94 tests (LoggingInfrastructure, LibraryPresentation, UpdateInfo, FrontendToolTip); the suite is 2,033 tests, with build, dotnet format, and InspectCode clean.

• 2026-11-08 — CORE-02/CORE-04/CORE-07 CLIO timer/read trace, CD poll enables, and level-held XBUS FIQ. Added reference-side CLIO timer/read/XBUS/FIQ event instrumentation (opera_clio_debug_*, sg_debug_timer_*) and harness --timer-trace; 5DO gained --clio-read-trace and cycle stamps on --clio-trace. Measured timer 14 underflow periods on Lemmings under the accuracy profile: 12,384 cycles in Opera, 12,396 in 5DO, a constant ~285-cycle phase offset, proving the old "~10,400" figure was the FreeDO mode and the residual is not the timer period. The READ_ID-to-READ_ERROR wait remains +22,834 cycles but is +8,694 retired instructions in a data-dependent guest copy/decrypt loop; the first 0x118 read is 29,491,640/29,469,632 and the last pre-READ_ERROR reads are identical. The trace exposed two reference-parity bugs: Opera resets the CD poll to CDROM_POLL_INTEN_MASK (0x07) plus the boot media latch (0x87), while 5DO reset to 0x80 and lost all enables on the guest's read-modify-write; and Opera asks every XBUS device for XBP_FIQ each batch (level-held) while 5DO only raised FIQ 4 on sector/status publication. CdromDevice.ResetPoll, Xbus.FiqPending, and the FiveDoMachine level check now match the reference. The A/B then isolated the old READ_ID media-access heuristic: poll fix only and level FIQ only both kept Space Ace's 3,000-frame hash at 0x46916C61, while both together diverged; retiring the heuristic restored the hash and left Road Rash's 0xD5895AD5 unchanged. The final 50-title panel (test91-panel50-heuristic-retired) keeps 50/50 frame hashes, 0 errors, improves Daedalus Encounter to a full match, and leaves 38 phase-only plus the same 11 CD-only behavioral divergences; the 5-title sample keeps 5/5 frame hashes. The 3,000-frame sample (test92-samples5-long-final) keeps 4/5 video matches (Alone 0x175B8FE4, Lemmings 0xEBDB01CA, Space Ace 0x46916C61, Wolfenstein 0x175B8FE4); Road Rash remains the only video divergence (0xD5895AD5 vs 0x0528251F). Lemmings is still 109/105 commands at f=300. 1,925 tests pass; dotnet format and InspectCode are clean. Evidence: artifacts/diff/test69-timer-traces through test92-samples5-long-final; TestCorpus.md 1.1az; PortingNotes.md 12.38.

• 2026-11-07 — TEST-07/CORE-07 frame-boundary parity and checkpoint capture order. Two independent defects produced the universal frame-60 phase. OperaCoreHost finalized each checkpoint at the next video callback, so ARM PC/CPSR/instructions and CD state belonged to the following field while the audio hash was already correct; the harness now captures ARM/CD state in the checkpoint's own video callback and fills the audio hash when that field's audio callback completes. FiveDoMachine.RunFrame also ended on the rounded 208,541-cycle budget while Opera ends on the 263-line scanline boundary (opera_3do_process_frame), and the discarded last-batch overshoot accumulated about 3.3 instructions per field; RunFrame now ends on Quartz.VdlFieldIndex, matching the reference's line < opera_region_scanlines() loop. On the five-title sample (test63-line-boundary/report.json) the same-frame retired-instruction delta at f=60 fell from ~89,000 (misaligned) / ~219 (after the capture fix) to 8-17, and at f=120 to 24-84; PCs match within one instruction. The 50-title confirmation panel (test64-panel50-line-boundary/report.json) captures 50/50 with 0 errors and 50/50 frame hashes, with a same-frame delta of 11-20 at f=60 and 23-2,223 at f=120. The CD count matches on 50/50 titles at f=60 and f=120 (down from 36 first divergences at f=120), and the first behavioral divergence moved to 11 titles at f=300 (8 count differences, 3 one-command sequence differences); the other 39 titles are phase-only. The 3,000-frame showcase keeps 4/5 video matches (Road Rash is the only frame-hash divergence) with CD counts Lemmings 365/352, Gex 548/547, Road Rash 851/836, Wolfenstein 862/861, and Alone 932/932 (Alone's sequence hash differs but the video matches). This supersedes the earlier "5DO frame 60 aligns with Opera frame 59" interpretation. 1,923 tests pass; build, format, and the focused machine/Quartz suites are clean.

• 2026-11-07 — TEST-07/CORE-04 CD register trace and cadence root-cause split. Added --cd-register-trace (cdreg f=<frame> cycle=<cycle> pc=<pc> op value) covering command writes, poll writes/reads, status reads, data writes, and empty data reads. On Space Ace the READ_DATA status is drained by cycle 58,934,229, then there is no CD register access for 237,794 cycles before command 164 at 59,174,910. On Lemmings all 12 READ_ID status bytes are drained by 29,431,952, then there is no CD access until 31,599,968 (2.17M cycles). Both large cadence steps are therefore guest waits on a non-CD source (timer/VBlank), not CD readiness. This confirms the transcript-classifier result (39 identical / 8 prefix / 3 READ_ID-shift streams) and moves the residual lead from CORE-04 to CORE-07. 1,922 tests pass; build, format, and the focused CD/DiffComparer suites are clean.

• 2026-11-07 — TEST-07/CORE-04 CD transcript classification and cadence-step evidence. Added CdCadenceAnalyzer to the differential harness. Every report now carries cdCadence (Equal / PrefixAhead / SequenceDiverged, first mismatch, last matched cycle/instruction delta) and the largest cadence step (cycle jump, opcodes, command-issue PCs, retired-instruction jump); the Markdown/HTML reports add a CD cadence column and detail line. On the existing 50-title panel the command streams are 39 identical, 8 prefix-only, and 3 sequence-shifted; the three shifts are a one-command READ_ID poll placement (Chiki Chiki 111/112, Dragon's Lair 111/108, Virtual Horror 110/107). 24/50 titles have a significant step above 1,000 cycles and all 24 issue the step command from the same PC on both engines (17 at 60-100k, six at 20-60k, one >=100k). The fresh five-title test60-cadence-samples5/report.json confirms Alone 361/361 (+379), Road Rash 395/395 (+468), Wolfenstein 341/341 (-345), Lemmings 109/105 with one +22,834 step at command 51 (PC 0x00022140, +8,242 instructions), and Space Ace 168/168 with one +80,567 step at command 164 (PC 0x0002aa18, +32,330 instructions). The residual cdCommandCount difference is guest wait-loop cadence, not a divergent CD command table; the next measurement is a CD register-access trace around the step. benchmarks/diff/samples5.json now points at the local J:/Panasonic 3DO corpus used by the panel. 1,921 tests pass; build, format, and the focused DiffComparer suite are clean.

• 2026-11-07 — TEST-07 per-command PC/instruction stamps and READ_DATA first-sector parity. The debug oracle now exports sg_debug_cd_pc_at and sg_debug_cd_instructions_at; 5DO records the same PC/retired-instruction stamps with every issued opcode. --cd-commands writes opcode cycle pc instructions, and captures carry cdCommandPcs/cdCommandInstructions. The new evidence showed the 20th Century boot handoff (last READ_DATA to READ_ID) cost +207,707 cycles / +86,511 instructions; matching the oracle's first-sector POLST publication reduced it to -818 / -376, and the f=300 command-104 cycle delta fell from +324,027 to +111,990. READ_DATA deadlines now accumulate like cdrom_data_schedule_next instead of re-basing to now + sector_cycles. The confirmation panel (test55-panel50-first-sector) is 50/50 hashes, 0 errors, with 48/50 first PC divergences at frame 59 and 2/50 at frame 119, and the same 36/13/1 CD count distribution as test43. The 5-title sample keeps 5/5 frame hashes at 60/120/300 at unbounded and stock 2x speed; the 3,000-frame sample keeps the established 4/5 video result (Road Rash still diverges) and the same CD counts. 1,919 tests pass; build, format, and InspectCode are clean.

• 2026-11-07 — Frontend -fullscreen launch alias. The desktop shell now accepts -fullscreen in addition to -f and --fullscreen, with a unit test and an end-to-end smoke check. --smoke-test --fullscreen enters fullscreen before the smoke marker is written; the marker records windowState=FullScreen (normal smoke records windowState=Normal). docs/UserGuide.md documents all three spellings. 1,919 tests pass.
• 2026-11-07 — CORE-04 READ_DATA validation and negative oracle evidence. CdromDevice.ReadDataCommand now rejects an invalid MSF with MEI address error, rejects a transfer that starts at or runs past the last disc sector with MEI end address, and aborts any previous transfer so a stale sector cannot leak into the FIFO. Three CD tests cover the two error paths and the abort. The 5-title and 20th Century CD counts are unchanged. A separate oracle build with the current Opera transfer-completion read-status rule also produced identical counts, ruling out both the local oracle's first-sector status publication and past-end reads as the cause of the residual FMV cadence difference. 1,916 tests pass. Superseded: the first-sector entry above shows the publication rule did change the handoff cycle/instruction timing; the earlier count-only test was too narrow.
• 2026-11-07 — 50-title panel video-clean (CORE-13/VID-13 lead). The VDL background word now keeps all 24 bits and the CPU/OpenGL/Direct3D background paths output them directly; the previous 5-bit narrowing rendered 0x010101 as pure black. A zero VDL display-list word is now a stop that retains the previous control word and bitmap, matching opera_vdlp.c, instead of blanking the last visible scanline. The VDL frame save-state section is version 4 and converts version 1-3 packed backgrounds on load. Result: the 50-title panel matches 50/50 frame hashes at frames 60/120/300 (up from 32/50), and the 5-title panel matches 5/5 (Space Ace and Wolfenstein 3D were previously divergent). Road Rash still diverges at 1,000/3,000. 1,916 tests pass; build, format, and InspectCode are clean. Evidence: artifacts/diff/test43-panel50-bg/report.json, artifacts/diff/test39-vdlstop-samples5/report.json, artifacts/diff/test44-long-bg/report.json.

• 2026-11-07 — TEST-07 instruction-phase triage and harness robustness. The harness now searches the checkpoint frame window for the neighboring reference frame whose retired-instruction count aligns with 5DO, records pcPhaseFrame/pcPhaseInstructionDelta per title, and labels every PC/CPSR divergence as an instruction phase in the report. The 50-title panel proves all 50 first PC divergences align to Opera frame 59 with a residual of only -55..-256 instructions (same-frame delta ~89,000), fixing the earlier "fixed startup offset" interpretation. ISO staging is now atomic (.tmp + move, no partial file on failure), retries once on a transient CHD read error, and a per-title capture exception is reported as an Error instead of aborting the panel. 1,916 tests pass.

• 2026-11-08 - v1.5.0. GPU HQx 2x/3x/4x rule-table passes in OpenGL/GLES, Direct3D 11, and Vulkan; 24-bit VDL background and zero-entry stop; reference field-boundary RunFrame; CD poll/FIQ/READ_DATA parity and first-sector status; instruction-phase and CD-cadence harness reporting; atomic ISO staging with a transient-CHD retry; -fullscreen desktop alias. The 50-title panel matches 50/50 frame hashes at 60/120/300 with 0 errors; the command-50 boot handoff fell from +207,707 cycles / +86,511 instructions to -818 / -376. VDLP state is version 4 and Machine state version 3 for the 24-bit background. 1,939 tests pass; build, format, and InspectCode are clean. See WhatsNew.md and ReleaseNotes-1.5.0.md.

• 2026-11-07 - v1.4.0. Six framework-dependent single-file bundles plus SHA256SUMS.txt; online cover art (UX-26), VID-12 latency controls, SYS-06 unsigned-disc support, the exact 60000/1001 NTSC field clock and rational VDL accumulator, VDLP blanking/background fixes, live OpenAL device-switching and test-tone fixes, and differential-harness CD-sequence/cycle evidence. The machine frame section now preserves dimensions and pixel aspect across load, run-ahead invalidates its snapshot after external machine actions, and the cover CLI/UI fail clearly when the catalog is unavailable. 1,905 tests pass; build, format, and InspectCode are clean. Release run https://github.com/purelogiccode/5DO/actions/runs/42399055155; CI run https://github.com/purelogiccode/5DO/actions/runs/42398646569. See WhatsNew.md and ReleaseNotes-1.4.0.md.

• 2026-11-07 - UX-26 online cover catalog. Added FiveDo.Host/Covers (catalog model and validation, content-addressed cache, conditional HTTP downloader, local/online resolver, annotation stripping, Jaro-Winkler, token/no-space/subtitle aliases, sequel-number guard, region/year/retail tie-breaks), the covers CLI, Settings > Covers, and Configuration > Cover Art. The packager deduplicates the 1,356-file source collection to 473 unique images, encodes 600 px WebP q85 (25.4 MB, 26.7 MB ZIP), maps 645/653 4DO game IDs (9 manual overrides, 5 weak, 8 homebrew/emulator entries without covers), and reloads the generated catalog to verify every image hash and unique game-ID assignment. The catalog is hosted at https://assets.purelogiccode.com/5do/v1; see CoverArt.md. 14 host cover tests pass.

• 2026-11-06 — InspectCode cleanup. Removed the last seven jb inspectcode --sEverity=WARNING warnings without changing behavior. StateWriter lost two redundant (ushort)0 casts. JitInstruction no longer stores the unused Terminator property (the classifier's terminator local still stops the block) and declares LayoutKind.Auto. The unused EmitStoreUserConstant helper was deleted. EmitDataProcessing and EmitBranch no longer take an unused state local. EnterFiqFromJit is invoked only from generated JIT IL, so it is now looked up with nameof and carries a scoped SuppressMessage("ReSharper", "UnusedMember.Local") with a justification rather than being deleted. InspectCode reports 0 results; 1,885 tests pass and format is clean.

• 2026-11-06 — Exact NTSC field clock and rational VDL (TEST-07/CORE-13 lead). NTSC was modeled at 60.00 Hz (208,333 cycles/field) instead of the hardware and reference 60000/1001 (208,541.67). The 1 ms per 60 fields difference produced a 44-sample audio drift at frame 60 (44,101 vs 44,144) and 2,514 samples at 3,000 frames. The VDL accumulator also added the per-batch rounded 8.24 ARM fraction, running the scanline clock ~0.09% fast and slipping the VDLP phase. VideoRegionProfile now exposes exact numerator/denominator and GetFrameCycles; Quartz uses an exact rational VDL accumulator (cycles * fieldLines * fieldRateNumerator against armClock * fieldRateDenominator) and the Quartz state section is version 4 (v1-3 states convert their 8.24 fraction on load). On the 50-title panel every sample delta is now ≤9 at frame 300, and 32/50 titles match every frame hash at 60/120/300. At 3,000 frames Space Ace, Wolfenstein, Alone, and Lemmings match Opera's hash exactly; Road Rash still diverges. 1,885 tests pass.

• 2026-11-06 — TEST-07 command-cycle and instruction-parity instrumentation. CdromCommandTranscript records the ARM cycle of every issued opcode; the debug reference core adds sg_debug_cd_cycle_at (exact command boundary via cmd_idx, fixing the old byte-modulo-seven transcript for one-byte ABORT packets) and sg_debug_arm_instructions. The harness now reports the first behavioral divergence separately from PC/CPSR frame-boundary phase, prints the first CD command cycle delta, and records retired-instruction counts. 50-title result: 50/50 captured, 0 errors; CD count matches 49/50 at frame 60; The absolute instruction count carries a fixed ~89,000-instruction startup offset, but per-frame retirement matches within 0.1% on 4/5 titles; PC/CPSR is a phase sample, not a cycle-model gap. See DifferentialHarness.md §8 and TestCorpus.md 1.1ax.

• 2026-11-06 - TEST-07 oracle increment. Added sg_debug_set_seed, sg_debug_arm_cpsr, sg_debug_cd_command_count, sg_debug_cd_command_hash, and sg_debug_cd_command_at to the debug reference core; added fixed-seed plumbing (default 1488), CPSR and CD command/sequence comparison, sampleTolerance, and firstDivergence reporting to FiveDo.DiffHarness; added issue-time command counters and the opcode list to CdromCommandTranscript and CdromDevice. benchmarks/diff/samples5.json runs five local titles at 300 frames. Two consecutive runs produced identical checkpoint arrays for both engines. Result: 0 match / 5 diverged / 0 error; first divergence is PC at frame 60 on all five; Lemmings CD commands match 9/9 at f=60 and 50/50 at f=120, then 108/105 at f=300; CPSR differs at f=60 by the FIQ-disable bit. Added the panel generator and produced benchmarks/diff/panel50.json (50 evenly spaced titles from the 594-disc manifest). The 3,000-frame follow-up shows Lemmings matching the frame hash at every checkpoint, Wolfenstein/Alone recovering after transient mismatches, and Road Rash/Space Ace visibly diverging; a ±60-frame window found no neighboring hash match. 1,881 tests pass. See DifferentialHarness.md §8 and TestCorpus.md 1.1aw.

• 2026-11-06 — Plan rewrite (focused execution plan). Reorganized the backlog around the compatibility claim. Added TEST-07 (reference oracle instrumentation and first-divergence metric), promoted TEST-06 to P0 as a staged top-100 playability audit, moved performance and product work to §3.5, moved research to §3.6, and closed PERF-08/CORE-05 with their measured negative evidence. SYS-05 is blocked on GPL-2.0 licensing, not engineering. No IDs were retired; SYS-06, SYS-08, SYS-12, SYS-17, SYS-18, and VID-04 moved to §2.5. The plan retains the five anchors other docs link to.
• 2026-11-06 - VID-12 latency features. Added 0-2 frame run-ahead and 0-8 frame presentation delay in CoreRunner, VRR/adaptive-sync presentation, and a −250…250 ms audio sync offset. Run-ahead uses a reusable uncompressed MachineStateSnapshot; hidden frames suppress audio/host submission but still render through Vdlp.DoLine so visible hashes match the normal timeline. State frame section v2 writes only visible lines and the RAM loader clears only the previously addressable tail. Each depth step removes one field of emulator-side input delay (16.7 ms at 60 Hz, 20 ms at 50 Hz PAL). bench --latency measured a 9,348,438-byte snapshot, ~3.6-3.9 ms capture, and ~1.6-2.0 ms restore; one-frame run-ahead stays inside the 60 Hz budget on a typical title, and the runner auto-disables after 45 overruns or a rejected snapshot. LatencyTests covers snapshot round-trip, 1-/2-frame timeline parity, visible-only audio, frame-delay holding, range validation, and telemetry. Settings, per-game profiles, the FPS overlay, and --audio-sync expose the options. Real AOT frontend run: Lemmings booted with run-ahead 1, frame delay 1, VRR, and audio sync 40 ms; OpenAL initialized at 44.1 kHz through the null device with no fault. Full optical input-to-photon measurement and real adaptive-sync hardware confirmation remain validation caveats documented in Performance.md §8.7 and KnownIssues.md. See Audio.md §3 and Video.md §7.
• 2026-11-06 - PERF-02/BAK-08 NativeAOT, trimming, and audio validation. CLI and frontend AOT win-x64 single-file publishes build and start. The CLI AOT executable passes --version, selftest --json, bench --frames 2 --json, and bench --latency; a 3,000-frame Lemmings boot and a 240-frame realtime OpenAL run with --audio openal --audio-sync 40 under ALSOFT_DRIVERS=null completed with 0 dropped samples and 0 underruns. The AOT frontend --smoke-test returns FRONTEND SMOKE OK 1.3.0, opens the real OpenGL 4.0 surface, boots Lemmings, and logs OpenAL audio ready (44100 Hz stereo, 100 ms, sync 40 ms) with run-ahead 1, frame delay 1, and VRR enabled. Sizes: CLI AOT 16.6 MB exe / 18.7 MB runtime folder; frontend AOT 34.5 MB exe / 54.6 MB runtime folder; trimmed frontend 27.9 MB exe / 48.0 MB folder; trimmed CLI 19.4 MB. Startup: CLI AOT --version 69 ms, bench --frames 2 65 ms; AOT frontend --version 70 ms and smoke 1.2 s; trimmed frontend smoke 3.2 s. Added four opt-in .pubxml profiles. Artifacts remain framework-dependent by default; AOT/trimming stay opt-in. Silk.NET IL3000/IL3002 and CHDSharp/SharpGen IL2104 warnings remain visible but non-fatal. See Trimming.md.
• 2026-11-06 - SYS-06 RSA/unsigned-disc option. Added BiosRsaPatch, the documented two-site media-RSA bypass, applied only in memory. Applying it to retail panafz10.bin and panafz10e-anvil.bin reproduces the known No-RSA MD5s 1477bda8... and cf11bbb5... exactly. The BIOS catalog now identifies the No-RSA FZ-1J/FZ-10/FZ-10E images, automatic selection prefers them when --allow-unsigned-discs/Settings > Paths is enabled, and a compatible retail image is patched otherwise. run/trace report biosRsaPatched/biosRsaDisabled, bios info reports rsaDisabled/rsaPatchSites, and bios select accepts --allow-unsigned-discs. Signed Wolfenstein 3D and Road Rash 600-frame hashes are unchanged; the option never writes the BIOS file and no BIOS/game content is bundled. See TestCorpus.md 1.1au.
• 2026-11-06 - v1.3.0. Six framework-dependent single-file bundles plus SHA256SUMS.txt; 480i/240p field handling (auto detection, bob, weave, blend), interlace-aware viewport geometry, screenshot/GIF support, per-game and CLI mode selection, and 1,833 tests. The release is presentation-only over 1.2.2: default bob hashes and save states are unchanged. Release workflow https://github.com/purelogiccode/5DO/actions/runs/42235476257; docs/wiki workflow https://github.com/purelogiccode/5DO/actions/runs/42235456021.
• 2026-11-06 - VID-04 480i/240p field handling. Added InterlaceProcessor with bob, weave, blend, and content-based auto detection; wired it through FiveDoMachine, AppSettings/GameProfile, Settings > Video, screenshots/GIF, and --deinterlace; added VdlFrame.PixelAspect and interlace-aware viewport scaling so a woven 320x480 frame keeps its display geometry and integer scaling. Sixteen processor tests plus bob-vs-auto hashes on five local titles (3,000 frames for Road Rash, Space Ace, and Alone in the Dark) are unchanged. See Video.md §1.
• 2026-11-05 - PERF-06 ANGLE revalidation. Re-ran the 4K offscreen probe through ANGLE/D3D11 on the RTX 4070: OpenGL ES 3.0, integer textures, sharp-bilinear present mean 0.67 ms / p99 1.29 ms over 120 measured frames, with 0 B/frame in the renderer. Real GLES 3.1 hardware and Tier B percentile runs remain open, so PERF-06 stays Partial. See Performance.md §8.
• 2026-11-05 - ENG-02 interpreter hot-path increment. Added a direct DRAM/ROM instruction fetch (IAddressBus.ReadInstruction), inline LSL#0 and rotated-immediate operand handling, a flag-preserving data-processing dispatch that avoids the generic ArmAlu call, BitOperations.PopCount/TrailingZeroCount LDM/STM iteration, and a cached PC-adjust source. Interleaved same-host A/B: synthetic 1.58→1.33 ms/frame and Lemmings 1.43→1.13 ms/frame, with unchanged hashes. ENG-02 stays Partial pending a decoded block cache; see Performance.md §8.6.
• 2026-11-05 - PERF-08 managed IL JIT second pass. Replaced per-instruction verification with exact RAM-write invalidation, kept PC/CPSR/operands/flags in block locals, compiled through RAM stores and conditional-branch fallthroughs, added direct IRam load/store fast paths, and fixed R15 operands to use the block-local PC. Correctness: 1,000-frame CLI hashes match the interpreter on Road Rash 0x41D2A1DA, Powers Kingdom 0xDCE31DC5, and Lemmings 0x110E7DC5, with identical instructions/cycles. On the synthetic ALU/store loop the old JIT measured 1.32-1.36 ms/frame and the new JIT 0.78 ms/frame (1.7x faster); Lemmings improved from 1.58-2.60 to 1.40-1.78 ms/frame but remains slower than the 1.08-1.25 ms/frame interpreter, so the interpreter remains the default and PERF-08 stays Partial. See Performance.md §8.4.
• 2026-11-06 - PERF-08 JIT context ABI and register shifts. Compiled blocks now share PC/CPSR/cycle/instruction state through ref parameters, so a chain of blocks inside one 16-cycle batch does not reload and flush architectural state per block. The FIQ exit stores the post-vector context, fixing a frame-loop rewind to the interrupted PC. Register-specified shifts are compiled with the interpreter's extra internal cycle and have a dedicated parity test. The compile-hit counter is an 8,192-entry direct-mapped table; 4,096 hits remains the best measured threshold. Lemmings improved from 1.60-1.66 to 1.23-1.37 ms/frame (~20% faster than the previous JIT) but is still ~1.25x slower than the 1.00-1.02 ms/frame interpreter; the synthetic loop is 0.89-0.99 versus 1.20-1.30 ms/frame. Hashes and instruction counts match on the 600-frame Lemmings checkpoint and 500-frame CLI runs for Road Rash, Powers Kingdom, and Lemmings. PERF-08 stays Partial. See Performance.md §8.4.
• 2026-11-06 - CORE-05 threaded-DSP analysis. Measured the DSP share on Lemmings (USA): 0.141 ms/frame of ~1.0 ms/frame (~14%), with 3.2M DSP instructions, 441k loop calls, and ~16.8 instructions per running loop. A per-sample worker handoff would cost more than the ~0.19 µs of DSP work, and a one-sample-ahead speculative pipeline must validate every ARM write to FIFO buffers, FIFO descriptors, semaphore, DSP memory, and CPU-supplied coefficients, replaying on any change. No threaded path was shipped because it would be slower on the measured workload and risks FIQ/FIFO timing; CORE-05 remains Pending. See the CORE-05 item for the full analysis.
• 2026-11-05 - PERF-05 high-resolution VDLP copy. High-resolution scanline DMA now reads the raw word backing span directly instead of four IRam.Read16 calls per pixel, with the same lane and plane handling. Interleaved A/B on Lemmings measured high-res 2.42-2.47→1.30-1.43 ms/frame (about 1.8x) with an identical high-res frame hash; PERF-05 stays Partial pending rasterizer and Tier B work. See Performance.md §8.6.
• 2026-11-05 - Rewind capture stall and state-size increment. Fixed the sub-second gameplay freezes caused by the default rewind history capturing a keyframe on the emulation thread every 120 frames. Rewind keyframes now use CompressionLevel.Fastest and unchecked transient CRCs, the memory section (version 4) stores only the live high-resolution RAM range, StateCrc32 is slicing-by-8, and RewindBuffer/StateWriter pool their capture stream, state arrays, payload buffer, and compression scratch. The BIOS MD5 in the header is cached, and the frontend opens discs through the prefetch cache. On Lemmings the worst keyframe capture fell from 139.3 ms to 7.4-9.2 ms (one 14.1 ms outlier), allocations fell from ~1.5 GB to 16 MB per 1,200 frames, and the uncompressed stock-layout state fell from ~26 MB to ~10 MB. See Performance.md §8.5 and SaveStates.md.
• 2026-11-04 - SYS-07 memory capacity options. Added the eight Opera DRAM/VRAM layouts (21, 22, 41, 42, 81, 82, E2, F1) through MemoryConfiguration/MemoryLayout, dynamic RAM/VRAM decode, VRAM base and masks for VDLP/SPORT/CEL, capacity-derived MADAM RED SYSBITS, --memory/Settings/profile/state plumbing, memory state section v3, Options section v8, a homebrew probe ROM that uses 4 MB DRAM + 2 MB VRAM, and 34 memory tests. Stock hashes are unchanged; expanded memory is documented as homebrew-only.
• 2026-11-04 - SYS-08 clock matrix increment. bench --cpu-clock/--memory measure the supported clock range; the 0.5x-4x matrix and unchanged stock default are recorded in Performance.md §8.1.
• 2026-11-05 - Library duplicate visibility. The library identifies discs by the MD5 of their first two data sectors (the 4DO game ID), so renamed copies and GoodTools/Redump/No-Intro variants collapse to one entry. The header now reports 653 games · 1,366 images, and Settings > Paths > Show duplicate disc images lists every file. GameLibraryScanner.BeginScan/InspectPending/Scan accept deduplicate, and library verify checks every image. Added scanner tests for duplicate listing and ImageCount.
• 2026-11-05 - Settings game folders and fast library listing. Added a Game folders editor to Settings > Paths: type or paste one or more folder paths, press Enter or click Add, Browse, or Remove. The list persists in LibraryFolders and refreshes the library when it changes. GameLibraryScanner now builds a per-directory cover index instead of calling File.Exists for four extensions per disc, sibling folder, and cover folder, and LibraryItem decodes covers on bounded background tasks at 256 px so the non-virtualizing grid never blocks the list. On the local 2,732-image collection the cached listing dropped from 2.8 s to 0.33 s (86% of the old time was cover probes), and a cold listing now returns 2,732 placeholders in 17 ms before background metadata hashing and cover decoding. Added cover-index tests for cached entries and PNG-over-JPG precedence.
• 2026-11-05 - PERF-08 managed IL JIT (partial). Added Arm60.Jit.cs, a DynamicMethod JIT behind --cpu-backend jit that compiles hot data-processing, load/store, and branch blocks. Blocks verify every instruction against current memory, stop at the machine's 16-cycle batch boundary, return after stores/branches, and fall back for unsupported instructions, SWIs, mode changes, register shifts, self-modifying code, and runtimes without dynamic code. Correctness: 100M-instruction lockstep parity on Lemmings and identical 3,000-frame Lemmings plus 1,000-frame Road Rash/Powers Kingdom hashes. Per-class microbenchmarks are 1.9-3.3x faster, but end-to-end boot is 1.17-1.85x slower, so PERF-08 remains Partial and the interpreter remains the default. See Performance.md §8.4.
• 2026-11-05 - PERF-04 prefetch buffer increment. PrefetchingDiscSource now reuses evicted cache arrays through a bounded pool, keeps one worker buffer for 2048-byte data reads and one for 2352-byte audio reads, and clears the pool on dispose. Cache hits are allocation-free on the caller thread and covered by a 1,000-read test. Continuous miss/eviction streaming still needs an end-to-end allocation gate.
• 2026-11-05 - SYS-18 auto per-game profile completion. Database inputDevices now drives AppSettings.ResolveInputDevices and CLI runs without --input-device; auto-input seeding no longer persists a shadowing override. The settings dialog applies one effective profile snapshot, updates input/filter/high-resolution live, and restarts only for BIOS/NVRAM/fix/timing changes. GameProfile gained a CheatBundle field and Clone. Evidence: WorldCup Super Stadium (JP) auto-selected FZ-10J + Kanji and rendered "Story of the World Cup"; Powers Kingdom Europe auto-selected Anvil/PAL1; Slayer USA auto-selected FZ-10; Crash 'n Burn USA applied fix bit 1; Crime Patrol auto-selected the lightgun and rendered its intro.
• 2026-11-05 - SYS-17 game database expansion. Grew GameDatabase.xml from 518 to 653 records, added revision, discNumber/discTotal, bios/bios2, fixMode, inputDevices, nvram, and cheats, and wired fixMode into GameCompatibility, BIOS fields into BiosLibrary, and input fields into GameControlDefaults. Added gamedb stats|scan|export|merge, the filename parser, serial retry for transient CHD read failures, and docs/GameDatabase.md. gamedb stats --folder "J:\Panasonic 3DO" reports 1,366 readable images, 653 unique IDs, and 100% metadata coverage with no errors; the three former exceptions were audio-only [CD-Audio] CHDs, which now open and get TOC-derived IDs.
• 2026-11-05 - SYS-01 audio-only CHD increment. ChdDiscImage no longer requires a data track, audio-only discs get a stable TOC-derived GameId, and info reports audioCd/trackCount. This let the game database catalogue the three local [CD-Audio] bonus discs. In-emulator playback still needs a disc with a data track or a guest CD player, so SYS-01 remains partial.
• 2026-11-05 - SYS-08 CPU clock validation. Recorded the five-title 0.5x/1x/2x/4x A/B in Performance.md §8.3. Every non-stock clock changed the frame hash, and AutoBahn Tokio's wall time rose from 9.1 s at 1x to 84.4 s at 2x and 271.7 s at 4x. The stock 12.5 MHz default and the no-per-title-clock decision are unchanged.
• 2026-11-04 - SYS-12 Portfolio OS SWI HLE. Ported Opera's 14 fixed-point math SWIs (0x50000-0x50012, with the object/Dot4 cases left to the guest), added --swi-hle/Settings/profile/state v9 plumbing, --swi-comments histograms, 33 CPU/machine tests that match opera_fixedpoint_math.c, and a real-BIOS synthetic 4x4 benchmark. Default-off hashes are unchanged; HLE is an opt-in performance mode and intentionally skips the guest handler's cycles and register clobbers, so some titles (AutoBahn Tokio, Ballz) change frame hashes.
• 2026-11-04 - FND-01 machine option registry. Added MachineOptionRegistry (13 validated options with CLI/Settings/profile/state mappings), typed MachineOptionValidationException/UnknownMachineOptionException, Opera DIPIR and soft-reset controls, --opera-dipir/--opera-soft-reset/--cpu-backend, settings options generated documentation, docs/MachineOptions.md, save-state Options v7 coverage for every option, mismatch warnings, and CLI/Settings/profile parity tests.
• 2026-11-04 - Plan rewrite. Consolidated the former dashboard, parts I-IV, feature spec, parity audit, and handoff into this single backlog; retired duplicate BAK-*/MOON-* trackers.
• 2026-11-06 — v1.2.2. Six framework-dependent single-file bundles plus SHA256SUMS.txt; interpreter, high-resolution VDLP, and rewind performance; opt-in managed JIT; expanded game database; library health/DAT verification and duplicate image counts; eight Opera memory layouts; eight PBUS ports; opt-in Portfolio OS SWI HLE; audio-only CHD cataloging; 1,815 tests.
• 2026-11-03 — v1.2.1. Six x64/arm64 bundles plus SHA256SUMS.txt; CI verified every apphost machine type and every checksum.
• 2026-11-03 — v1.2.0. Direct3D 11 presenter, startup release checks, deterministic fuzz harness, per-edge overscan/void border, screenshot gallery filters/compare/export, display-refresh control, flat CHD/raw ISO detection, FIQ/DSP telemetry, CLIO/DIPIR/FIFO/MCTL/XBUS audit fixes.
• 2026-11-02 — Hardening pass. Fuzz harness, overscan profiles, gallery features, FIQ request-to-vector latency, DSP per-tick statistics, 1,622 tests.
• 2026-11-01 — Disc/UX/accuracy pass. CD-DA and multi-track, M3U/live swap, PAL/NTSC, BIOS manager, per-game NVRAM, rewind, fast-forward, cheats, TAS, disc browser, input devices, OSD/quick menu, library metadata, portable mode, instance lock, save-state format v3.
• Earlier: v1.0.0 documented-compatibility release; 594-disc sweep; legal/attribution sign-off; six-RID packaging; OpenGL/GLES, Vulkan, and Direct3D 11 presenters; 4K output; differential harness.

Appendix A — Opera option coverage¶

This is a crosswalk, not a second backlog. "Canonical item" is the only place the remaining work is tracked. Rows updated in the 2026-11-06 rewrite reflect the plan's own completed evidence.

Additional 4DO/Opera UX gaps already covered elsewhere: physical CD drive (SYS-21), disc conversion (SYS-22), void patterns (VID-17), auto-load/auto-open (shipped), localization (UX-09), installers/file associations (UX-16), session focus behavior (UX-25).

Appendix B — Retired and duplicate IDs¶

Priority changes in the 2026-11-06 rewrite (historical traceability):

This table records the 2026-11-06 rewrite. The 2026-11-08 feasibility revision supersedes it: SYS-05 is closed, CORE-01 is closed, and TEST-06 now starts with the unique corpus (Tier A complete) rather than a 594-disc audit.

Dropped scope (do not add checkboxes): WebAssembly/browser (ECO-02), Android/iOS (ECO-03), Raspberry Pi/arm64 GLES optimization (ECO-05), Steam Deck. The libretro core wrapper is not planned; 5DO stays standalone.

Appendix C — Source-of-truth links¶

• Status — short current answer.
• Compatibility — supported set and platform evidence.
• Game compatibility results - per-disc sweep table.
• 5DO vs 4DO vs Phoenix - unique-title comparison.
• Known issues — current limitations.
• Performance — budgets and measurements.
• Test strategy and Test corpus — test evidence.
• Differential harness — Opera comparison.
• Reference strategy — Opera/4DO mapping.
• High-resolution CEL research — VID-01/02/03.
• Release checklist — release gate.
• Porting notes — intentional divergences.