Session Handoff [process-findings]: Promote and Release 3.23

Next steps, in priority order

Promote develop to main when the maintainer asks, with drive-pr to a mergeable promotion pull request. Its body carries one Fixes line per issue fixed on develop and not on main: Fail One File Instead of the Whole Run on a Per-File Exception #1069, Count an Orphan Sidecar Deletion as a Success, Not an Error #1070, Stop Setting Verified When the Verify Option Is Off #1071, Replace the Original Only After ReEncode and DeInterlace Fully Succeed #1072, Stop Remux and Lowercase Renames From Overwriting an Existing File #1073, Fail or Report RepairMetadataErrors When Errors Remain After Repair #1074, Re-Run the Full Cleanup After a Re-Encode Repair #1075, Guard ReEncode and DeInterlace Against Repeating on Every Run #1076, Remove Cover Art That Only MediaInfo Reports, and Re-Check the FfProbe Path #1077, Clear BitrateExceeded and Handle Failed Bitrate and Packet Analysis #1078, Decide Whether IsMediaAndToolsCurrent Has the Tool-Change Sense Inverted #1079, Skip Registering Deleted Files in FileIgnoreList #1080, Handle Write Failures for the Settings and Results Files After a Run #1081, Keep the State Set Before a Rename When the New Name Has a Sidecar #1082, Decide How a Signal-Killed Repair Run Affects DeleteInvalidFiles #1095, Enforce a Minimum Version for Every Media Tool Family #1104, Trim the HISTORY Release Entries to the Changes a User Needs #1105, Add a Codecov Coverage Badge to the README Build Status #1106, Split the Remux and Container Steps Diagram into Narrower Charts #1107, Keep a File Found When a Multi-Step Rename Fails Partway #1109, Set FileDeleted When Only the Sidecar Delete Fails in DeleteFailedFile #1116, Harden FileIgnoreList Pruning and Post-Run Write Retry #1120, Consolidate Fake Tool Fixture Setup in FakeTools #1124, Report Every Repair a Track Needs When Its State Holds One #1127, Share One Log Capture Helper Across Test Classes #1129, Filter find_run on Corpus So a Reduced Run Is Not Read as Full-Corpus Ground Truth #1134. Re-derive the set from origin/main..origin/develop and each pull request's Closes on promotion: line rather than trusting this list. Done is a promotion pull request meeting every Merge Gate item except the merge permission.
Release 3.23 after the promotion merges, when the maintainer says so, via merge-and-release. HISTORY.md 3.23 is in its release shape. Share One Log Capture Helper Across Test Classes #1129 and Filter find_run on Corpus So a Reduced Run Is Not Read as Full-Corpus Ground Truth #1134 are test and regression tooling only and need no line.
Close this track with close once the release ships, unless new issues have been filed by then. No other work is queued on it.

External blockers

The promotion merge and the release each need the maintainer's explicit go-ahead, per merge-and-release.
local_review.py record accepts a typed findings count with nothing tying it to the reviewer's real report, Bind a Local Review Record to the Reviewer's Actual Report ProjectTemplate#3699. Record a pass only from the reviewer's harness-delivered completion.
The README coverage shield has no fleet spec class yet, Add a Coverage Shield Class for Repositories That Upload Coverage ProjectTemplate#3692.

Internal dependencies

develop takes no merge commits. A promotion is a merge of develop into a branch off main, since Dependabot bumps hub workflow pins on main.
The release follows the promotion merge, since it ships from main.

State

develop is at de6c2a4 (Filter find_run on Corpus So a Reduced Run Is Not Read as Full-Corpus Ground Truth #1135), 21 commits ahead of main. main holds the earlier 3.23 promotion, unreleased.
No feature worktree is registered and no pull request is open. The primary checkout is on develop, clean, at origin/develop. Re-read with git worktree list and gh pr list.
Open issues besides this handoff are the promotion's Fixes set above and the long-standing external issues HandBrake converts all subtitles to ASS vs. passthrough #95, FFmpeg removes IETF language tags #148, HandBrake removes IETF language tags #149. Re-read with gh issue list --limit 200.
Hub issues still open from earlier rounds: Add a Coverage Shield Class for Repositories That Upload Coverage ProjectTemplate#3692 and Bind a Local Review Record to the Reviewer's Actual Report ProjectTemplate#3699.

The parked decision queue

It holds 0 issues. No open issue in this repository carries the decision label.

What the last round did

Fixed Filter find_run on Corpus So a Reduced Run Is Not Read as Full-Corpus Ground Truth #1134 in Filter find_run on Corpus So a Reduced Run Is Not Read as Full-Corpus Ground Truth #1135, squash-merged into develop. find_run takes a corpus argument, default full, and skips a run whose buildinfo.json Corpus differs. A run without the field reads as full.
The issue proposed falling back to the -reduced directory suffix for runs without Corpus. The local strict pass showed the harness has written Corpus since its first commit, the same one that introduced -reduced directories. So the fallback is "full", and the suffix check was dropped before the first push.
Copilot recommended approval with no findings, and Qodo found nothing. CodeRabbit raised two test nitpicks with no threads, a second full run to cover newest-run selection and an assertion on the no-match message. Both were fixed, and its review of the fix found nothing actionable.
No new issues were filed this round.

What not to repeat

Check the history of a field before writing a back-compat path for it. The suffix fallback the issue proposed covered a run shape the harness never wrote, and cost a fix commit and a second strict pass before the push.

New learnings

None beyond the entry above.

Parked decision queue, presented at the session close: 1 issue. #1079 was put to the maintainer, who chose to leave it open for now, so it stays in the queue.

Round 1, closed 2026-10-30: Session Handoff [process-findings]: Fix the Process Command Suspect Branches

Next steps, in priority order

Fix the process-command suspect branches that Add Process Command Logic Diagrams to ARCHITECTURE.md #1083 documented, one issue per pull request into develop. Highest impact first:
- Stop DeleteFailedFile From Deleting Files on Cancel and Read-Only Skips #1068 (a cancel or skip deletes the file)
- Fail One File Instead of the Whole Run on a Per-File Exception #1069 (one exception ends the run)
- Treat an FfMpeg Run That Did Not Complete as a Tool Error, Not a Decode Error #1084 (an FfMpeg run that did not complete reads as a decode error and can delete the file)
- Replace the Original Only After ReEncode and DeInterlace Fully Succeed #1072 (the original is replaced before the step finishes)
- Stop Setting Verified When the Verify Option Is Off #1071 (Verified with Verify off)
Done: the fix merges, the matching ARCHITECTURE.md "Suspect Branches" entry is removed, and the diagram is redrawn to the new behavior in the same pull request.
Then work the rest: Count an Orphan Sidecar Deletion as a Success, Not an Error #1070, Stop Remux and Lowercase Renames From Overwriting an Existing File #1073, Fail or Report RepairMetadataErrors When Errors Remain After Repair #1074, Re-Run the Full Cleanup After a Re-Encode Repair #1075, Guard ReEncode and DeInterlace Against Repeating on Every Run #1076, Remove Cover Art That Only MediaInfo Reports, and Re-Check the FfProbe Path #1077, Clear BitrateExceeded and Handle Failed Bitrate and Packet Analysis #1078, Skip Registering Deleted Files in FileIgnoreList #1080, Handle Write Failures for the Settings and Results Files After a Run #1081 and Keep the State Set Before a Rename When the New Name Has a Sidecar #1082.
Answer Decide Whether IsMediaAndToolsCurrent Has the Tool-Change Sense Inverted #1079 (a decision) before touching SidecarFile.IsMediaAndToolsCurrent().
Open a pull request for the pushed docker-layout-user branch, a one-line GOVERNANCE.md Repository Layout fix left over from Fail Closed When the Nonroot Docker User Does Not Resolve #1067. Done: it merges into develop.

External blockers

Decide Whether IsMediaAndToolsCurrent Has the Tool-Change Sense Inverted #1079 needs the maintainer to say whether the tool-change sense is intended. Meanwhile, the other issues can proceed.

Internal dependencies

Change a branch and its diagram together. ArchitectureDiagramTests fails when the Process.ProcessFile() step order or a named method moves without the diagram following, so a fix that reorders or renames a step updates ARCHITECTURE.md in the same commit.
Stop DeleteFailedFile From Deleting Files on Cancel and Read-Only Skips #1068 and Treat an FfMpeg Run That Did Not Complete as a Tool Error, Not a Decode Error #1084 both touch DeleteFailedFile() and the abort path, so do them in one lane or in sequence, not in parallel.
Replace the Original Only After ReEncode and DeInterlace Fully Succeed #1072 and Re-Run the Full Cleanup After a Re-Encode Repair #1075 both touch the re-encode and repair paths in ProcessFile.cs, so sequence them as well.

State

develop holds Add Process Command Logic Diagrams to ARCHITECTURE.md #1083, merged as a39dbf0. Its worktree and branch are cleaned up.
The branch docker-layout-user (e78aa9c) is pushed with no PR yet, and its worktree is still registered. Re-read with git worktree list.
The primary checkout is clean, on develop.
No release was dispatched. Unreleased 3.23 work on develop is still waiting for its promotion.

The parked decision queue

It holds 1 issue: #1079, which asks whether IsMediaAndToolsCurrent() inverts the SidecarUpdateOnToolChange sense.

What the last round did

Merged Add Process Command Logic Diagrams to ARCHITECTURE.md #1083, which adds the ARCHITECTURE.md "Process Command Logic" Mermaid diagrams and ArchitectureDiagramTests. The test holds the pipeline step order and the named methods to the code.
Filed Stop DeleteFailedFile From Deleting Files on Cancel and Read-Only Skips #1068 through Keep the State Set Before a Rename When the New Name Has a Sidecar #1082 and Treat an FfMpeg Run That Did Not Complete as a Tool Error, Not a Decode Error #1084 for the suspect branches the diagrams expose.

What not to repeat

Do not add ARCHITECTURE.md, or any docs path, to the release paths filter in test-pull-request.yml. That breaks WORKFLOW.md D1.4, and the validate unit-test job already runs dotnet test on every pull request. It was tried in 38fb3e8 and reverted in e304e3c.
Keep the Docker builder's dotnet test and its ARCHITECTURE.md copy. The maintainer chose to keep testing inside the image being built.
A full-diff local strict pass over this doc cost about 200k to 230k tokens each, and the later passes mostly found wording precision. For a fix PR, keep each change small, so its diff and its passes stay small.
local_review.py does not run from a git archive export of the hub's scripts/, because it imports a sibling module the export lacks. Use a shallow clone of the hub's main.

New learnings

Pass-cost and hub-clone notes are recorded in the maintainer's PR-loop memory. No governance change was made.

Parked decision queue, presented at the session close: 1 issue, now 0.

Decide Whether IsMediaAndToolsCurrent Has the Tool-Change Sense Inverted #1079: the maintainer chose to invert IsMediaAndToolsCurrent() to IsToolsCurrent() || !SidecarUpdateOnToolChange, matching Read(). The decision label is off. It is now an ordinary fix in next step 2, and next step 3 no longer blocks.

Round 2, closed 2026-10-30: Session Handoff [process-findings]: Fix the Process Command Suspect Branches

Next steps, in priority order

Fix the remaining process-command suspect branches, one issue per pull request into develop. Highest impact first:
- Treat an FfMpeg Run That Did Not Complete as a Tool Error, Not a Decode Error #1084 (an FfMpeg run that did not complete reads as a decode error and can delete the file)
- Fail One File Instead of the Whole Run on a Per-File Exception #1069 (one exception ends the run)
- Replace the Original Only After ReEncode and DeInterlace Fully Succeed #1072 (the original is replaced before the step finishes)
- Stop Setting Verified When the Verify Option Is Off #1071 (Verified with Verify off)
Done: the fix merges, the matching ARCHITECTURE.md "Suspect Branches" entry is removed, and the diagram is redrawn to the new behavior in the same pull request.
Then work the rest: Count an Orphan Sidecar Deletion as a Success, Not an Error #1070, Stop Remux and Lowercase Renames From Overwriting an Existing File #1073, Fail or Report RepairMetadataErrors When Errors Remain After Repair #1074, Re-Run the Full Cleanup After a Re-Encode Repair #1075, Guard ReEncode and DeInterlace Against Repeating on Every Run #1076, Remove Cover Art That Only MediaInfo Reports, and Re-Check the FfProbe Path #1077, Clear BitrateExceeded and Handle Failed Bitrate and Packet Analysis #1078, Skip Registering Deleted Files in FileIgnoreList #1080, Handle Write Failures for the Settings and Results Files After a Run #1081 and Keep the State Set Before a Rename When the New Name Has a Sidecar #1082.
Answer Decide Whether IsMediaAndToolsCurrent Has the Tool-Change Sense Inverted #1079 (a decision) before touching SidecarFile.IsMediaAndToolsCurrent().

External blockers

Decide Whether IsMediaAndToolsCurrent Has the Tool-Change Sense Inverted #1079 needs the maintainer to say whether the tool-change sense is intended. Meanwhile, the other issues can proceed.
A squash merge into develop needs the maintainer's explicit go-ahead in the session. The auto mode classifier denies gh pr merge until then.

Internal dependencies

Change a branch and its diagram together. ArchitectureDiagramTests fails when the Process.ProcessFile() step order or a named method moves without the diagram following.
Treat an FfMpeg Run That Did Not Complete as a Tool Error, Not a Decode Error #1084 changes what reaches DeleteFailedFile() from the verify path. Keep Skipped and Cancelled Files Out of DeleteInvalidFiles #1086 changed that method, so read its current guard first.
Replace the Original Only After ReEncode and DeInterlace Fully Succeed #1072 and Re-Run the Full Cleanup After a Re-Encode Repair #1075 both touch the re-encode and repair paths in ProcessFile.cs, so sequence them.
Stop DeleteFailedFile From Deleting Files on Cancel and Read-Only Skips #1068 stays open until the develop -> main promotion pull request closes it with a closing keyword, with the other issues the 3.23 promotion owes.

State

develop holds Keep Skipped and Cancelled Files Out of DeleteInvalidFiles #1086 (64ee6c0) and Correct the Docker Layout Entry on the Image User #1087 (16f8ccc). Both worktrees and branches are cleaned up.
The primary checkout is clean, on develop. No other worktree is registered. Re-read with git worktree list.
No release was dispatched. Unreleased 3.23 work on develop is still waiting for its promotion.

The parked decision queue

It holds 1 issue: #1079, which asks whether IsMediaAndToolsCurrent() inverts the SidecarUpdateOnToolChange sense.

What the last round did

Merged Keep Skipped and Cancelled Files Out of DeleteInvalidFiles #1086 for Stop DeleteFailedFile From Deleting Files on Cancel and Read-Only Skips #1068. A cancel and the read-only media and sidecar skips no longer reach the delete. DeleteFailedFile() waits for a pending cancel through Program.IsCancelledError(), after its DeleteInvalidFiles check, and deletes the sidecar only for an MKV file. It adds DeleteFailedFileTests, and removes the S1 suspect entry.
Merged Correct the Docker Layout Entry on the Image User #1087, the GOVERNANCE.md Repository Layout Docker/ line left over from Fail Closed When the Nonroot Docker User Does Not Resolve #1067.

What not to repeat

Do not use Program.IsCancelled() after a tool step to decide something destructive. A Ctrl-C can fail the tool before the handler sets the token. Use Program.IsCancelledError(), and put it behind the option check so the 100 ms wait runs only when it matters. The strict pass and then Qodo each caught one half of this on Keep Skipped and Cancelled Files Out of DeleteInvalidFiles #1086.
A unit test cannot cover a cancel. Program's cancel token source is static and cannot be reset, so Program.Cancel() in a test cancels every later test.
Each strict pass over a small fix cost about 65k to 110k subagent tokens. A narrow pass over just the fix commit was enough for each later round.

New learnings

PR loop notes from this round are in the maintainer's PR-loop memory. No governance change was made.

Parked decision queue, presented at the session close: 1 issue, now 0.

Decide How a Signal-Terminated FfMpeg Verify Run Is Classified #1090: the maintainer chose to keep a signal-terminated FfMpeg verify run as DecodeError but never delete on it, choice 3. The decision label is off. It is now an ordinary fix that should share the tool-did-not-run signal Keep a File Out of DeleteInvalidFiles When a Tool Did Not Run #1088 adds, so next step 3 no longer waits, and it sequences after Keep a File Out of DeleteInvalidFiles When a Tool Did Not Run #1088.

Round 3, closed 2026-10-30: Session Handoff [process-findings]: Fix the Process Command Suspect Branches

Next steps, in priority order

Fix the remaining process-command suspect branches, one issue per pull request into develop. Highest impact first:
- Keep a File Out of DeleteInvalidFiles When a Tool Did Not Run #1088 (S16): a tool that fails to launch, on a repair ladder re-verify or in any other step, can still delete the file. The fix is one "tool did not run" signal that DeleteFailedFile() reads, rather than one guard per step.
- Fail One File Instead of the Whole Run on a Per-File Exception #1069 (one exception ends the run)
- Replace the Original Only After ReEncode and DeInterlace Fully Succeed #1072 (the original is replaced before the step finishes)
- Stop Setting Verified When the Verify Option Is Off #1071 (Verified with Verify off)
Done means the fix merges with its ARCHITECTURE.md "Suspect Branches" entry removed. The same pull request redraws the diagram to the new behavior. A newly found diagram-to-behavior mismatch is filed as its own issue and marked as a new S entry.
Then work the rest: Count an Orphan Sidecar Deletion as a Success, Not an Error #1070, Stop Remux and Lowercase Renames From Overwriting an Existing File #1073, Fail or Report RepairMetadataErrors When Errors Remain After Repair #1074, Re-Run the Full Cleanup After a Re-Encode Repair #1075, Guard ReEncode and DeInterlace Against Repeating on Every Run #1076, Remove Cover Art That Only MediaInfo Reports, and Re-Check the FfProbe Path #1077, Clear BitrateExceeded and Handle Failed Bitrate and Packet Analysis #1078, Decide Whether IsMediaAndToolsCurrent Has the Tool-Change Sense Inverted #1079, Skip Registering Deleted Files in FileIgnoreList #1080, Handle Write Failures for the Settings and Results Files After a Run #1081 and Keep the State Set Before a Rename When the New Name Has a Sidecar #1082.
Decide How a Signal-Terminated FfMpeg Verify Run Is Classified #1090 (S17) waits on its decision before any code change.

External blockers

Decide How a Signal-Terminated FfMpeg Verify Run Is Classified #1090 needs the maintainer to choose how a signal-terminated FfMpeg verify run is classified. The issue states three choices. Meanwhile the other issues can proceed.
A squash merge into develop needs the maintainer's explicit go-ahead in the session.

Internal dependencies

Change a branch and its diagram together. ArchitectureDiagramTests fails when the Process.ProcessFile() step order or a named method moves without the diagram following.
Keep a File Out of DeleteInvalidFiles When a Tool Did Not Run #1088 and Fail One File Instead of the Whole Run on a Per-File Exception #1069 both change what reaches DeleteFailedFile() from the abort path in Process.ProcessFile(). Land Keep a File Out of DeleteInvalidFiles When a Tool Did Not Run #1088 first, or design the two together, so the tool-error signal and the per-file exception handling do not each add their own delete guard.
Decide How a Signal-Terminated FfMpeg Verify Run Is Classified #1090's chosen answer may reuse the ToolError path, so read VerifyResult and DeleteFailedFile() as they stand after Keep a File Out of DeleteInvalidFiles When a Tool Did Not Run #1088.
Replace the Original Only After ReEncode and DeInterlace Fully Succeed #1072 and Re-Run the Full Cleanup After a Re-Encode Repair #1075 both touch the re-encode and repair paths in ProcessFile.cs, so sequence them.
Stop DeleteFailedFile From Deleting Files on Cancel and Read-Only Skips #1068 and Treat an FfMpeg Run That Did Not Complete as a Tool Error, Not a Decode Error #1084 stay open until the develop -> main promotion closes them. It carries closing keywords for every issue the 3.23 promotion owes.

State

develop holds Treat an FfMpeg Verify That Did Not Run as a Tool Error #1089 (ea7be41) for Treat an FfMpeg Run That Did Not Complete as a Tool Error, Not a Decode Error #1084. Its branch and worktree are cleaned up.
The primary checkout is clean, on develop, and no other worktree is registered. Re-read with git worktree list.
No release was dispatched. Unreleased 3.23 work on develop is still waiting for its promotion.

The parked decision queue

It holds 1 issue: #1090, which asks how a signal-terminated FfMpeg verify run is classified.

What the last round did

Merged Treat an FfMpeg Verify That Did Not Run as a Tool Error #1089 for Treat an FfMpeg Run That Did Not Complete as a Tool Error, Not a Decode Error #1084. An FfMpeg verify run that fails to launch, is cancelled, or whose execution throws returns VerifyResult.ToolError. VerifyAndRepair() then fails the file with its state unchanged, RepairTimestamps() reports the failure, and DeleteFailedFile() keeps the file. Added VerifyToolErrorTests, and removed S15.
Filed Keep a File Out of DeleteInvalidFiles When a Tool Did Not Run #1088 (S16) and Decide How a Signal-Terminated FfMpeg Verify Run Is Classified #1090 (S17, a decision) from the strict review and CodeRabbit.

What not to repeat

Do not describe ToolError as "did not run to completion" or with a "before it exits" limit. ExecuteStreamStdErr() returns false on a cancel or on any exception up to the end of the stderr drain. A non-zero exit returns true, and it reads as DecodeError. Four review rounds each tightened this wording.
Do not claim a ToolError keeps the file everywhere. Only the file's own stream verify writes _lastVerifyResult, in Verify() or RepairTimestamps(), and the repair ladder's re-verifies do not.
The full-diff strict pass cost about 130k subagent tokens. Each narrow pass over one fix commit cost about 20k to 60k more. The two-round edit budget was spent once, on wording churn.

New learnings

ProcessFile steps can be tested with no media tools installed. Load the media props from a sample sidecar, with each tool's Info version set to match it and the stored state cleared. PlexCleanerTests/VerifyToolErrorTests.cs is the worked example. No governance change was made.

Parked decision queue, presented at the session close: 0 issues. Nothing was asked.

Round 4, closed 2026-10-30: Session Handoff [process-findings]: Fix the Process Command Suspect Branches

Next steps, in priority order

Decide How a Signal-Terminated FfMpeg Verify Run Is Classified #1090 (S17), decided: a signal-terminated FfMpeg verify run stays DecodeError for the repair ladder, but DeleteFailedFile() never deletes on it (choice 3). Done means the fix merges into develop with its ARCHITECTURE.md "Suspect Branches" entry removed and the Verify and Repair diagram redrawn. Read the design note under "Internal dependencies" first.
Fix the remaining process-command suspect branches, one issue per pull request into develop, highest impact first:
- Fail One File Instead of the Whole Run on a Per-File Exception #1069 (one exception ends the run)
- Replace the Original Only After ReEncode and DeInterlace Fully Succeed #1072 (the original is replaced before the step finishes)
- Stop Setting Verified When the Verify Option Is Off #1071 (Verified with Verify off)
  Same done criteria: the entry is removed and the diagram is redrawn in the same pull request. A newly found diagram-to-behavior mismatch is filed as its own issue with a new S number.
Then work the rest: Count an Orphan Sidecar Deletion as a Success, Not an Error #1070, Stop Remux and Lowercase Renames From Overwriting an Existing File #1073, Fail or Report RepairMetadataErrors When Errors Remain After Repair #1074, Re-Run the Full Cleanup After a Re-Encode Repair #1075, Guard ReEncode and DeInterlace Against Repeating on Every Run #1076, Remove Cover Art That Only MediaInfo Reports, and Re-Check the FfProbe Path #1077, Clear BitrateExceeded and Handle Failed Bitrate and Packet Analysis #1078, Decide Whether IsMediaAndToolsCurrent Has the Tool-Change Sense Inverted #1079, Skip Registering Deleted Files in FileIgnoreList #1080, Handle Write Failures for the Settings and Results Files After a Run #1081 and Keep the State Set Before a Rename When the New Name Has a Sidecar #1082.

External blockers

A squash merge into develop needs the maintainer's explicit go-ahead in the session. Ask for it as a prompt as soon as the Merge Gate is met.
CodeRabbit allows 2 reviews per hour on this account, then posts "Review limit reached" and skips a head. Its follow-up findings land in its review body with no thread, so read every CodeRabbit review across every head before calling it silent.

Internal dependencies

Decide How a Signal-Terminated FfMpeg Verify Run Is Classified #1090 and ToolErrorScope: the per-file count that ProcessDriver.ProcessFiles() begins does two things. DeleteFailedFile() keeps a file with any recorded tool error. And a repair that fails while its count moved leaves VerifyFailed and RepairFailed unset. Choice 3 wants only the first. So recording a signal-terminated verify into the same count would also suppress the failure state on a failed repair. Choice 3 does not ask for that. Either add a separate keep-the-file signal or confirm that suppressing the state is acceptable before building on the count.
Fail One File Instead of the Whole Run on a Per-File Exception #1069 also changes what reaches DeleteFailedFile() from the abort path in Process.ProcessFile(). Its per-file exception handling must not add a second delete guard beside ToolErrorScope.Count.
Replace the Original Only After ReEncode and DeInterlace Fully Succeed #1072 and Re-Run the Full Cleanup After a Re-Encode Repair #1075 both touch the re-encode and repair paths in ProcessFile.cs, so sequence them.
Change a branch and its diagram together. ArchitectureDiagramTests fails when the Process.ProcessFile() step order or a named method moves without the diagram following.
Stop DeleteFailedFile From Deleting Files on Cancel and Read-Only Skips #1068, Treat an FfMpeg Run That Did Not Complete as a Tool Error, Not a Decode Error #1084 and Keep a File Out of DeleteInvalidFiles When a Tool Did Not Run #1088 stay open until the develop -> main promotion closes them. The promotion carries closing keywords for every issue the 3.23 promotion owes.

State

develop holds Keep a File When a Tool Did Not Run #1091 (d21660b) for Keep a File Out of DeleteInvalidFiles When a Tool Did Not Run #1088. Its branch tool-not-run and its worktree are cleaned up.
The primary checkout is clean, on develop at origin/develop, and no other worktree is registered. Re-read with git worktree list.
No release was dispatched. The unreleased 3.23 work on develop is still waiting for its promotion.

The parked decision queue

It holds 0 issues.

What the last round did

Merged Keep a File When a Tool Did Not Run #1091 for Keep a File Out of DeleteInvalidFiles When a Tool Did Not Run #1088. It adds ToolErrorScope, a per-file count of tool runs that fail to launch, are cancelled, or throw. ProcessDriver.ProcessFiles() begins it, and every tool execute path's catch blocks record into it. DeleteFailedFile() keeps a file with any such run, and a failed repair whose count moved during the repair leaves its state unchanged. It removed S16 and redrew the Per-File Pipeline and Verify and Repair diagrams.
Filed nothing new. The pre-existing MediaTool.ExecuteStreamStdOut() having no caller is listed, not filed, in the Keep a File When a Tool Did Not Run #1091 body.

What not to repeat

Do not check the tool-error count before if (!repaired) in VerifyAndRepair(). A repair can succeed after an earlier tier's tool error, and it must still set Verified and Repaired.
Do not begin a per-file scope in Process.ProcessFile(). The plugin path and the other commands reach ProcessFile methods through ProcessDriver without it.
In a repair-ladder test, a ToolErrorScope.Count assertion is masked by the post-repair Refresh(true). Its sidecar update runs MediaInfo, which a fake tools root lacks. Assert the exact count, not a positive one.
A concurrency test that reads a scope after its partner may have disposed passes a shared-static regression about half the time. Read every count between two barrier phases.
Each frontier strict pass cost about 60k to 160k subagent tokens. The round ran eight passes, most over one-commit test fixes that the hosted reviewers asked for.

New learnings

A repair-ladder branch is testable with no media tools installed. Use step-counting shell scripts as fake tools and a non-executable file to force a launch failure. PlexCleanerTests/VerifyToolErrorTests.cs VerifyAndRepair_RepairAfterToolError_SetsRepaired is the worked example. No governance change was made.

Parked decision queue, presented at the session close: 1 issue, now 0.

Decide How a Signal-Killed Repair Run Affects DeleteInvalidFiles #1095: the maintainer chose to count a signal-killed repair run as a tool that did not run, choice 1. It goes into ToolErrorScope.Count, so a failed repair leaves the state unchanged and the file is kept. The decision label is off. It is now an ordinary fix in next step 4, sequenced with Replace the Original Only After ReEncode and DeInterlace Fully Succeed #1072 and Re-Run the Full Cleanup After a Re-Encode Repair #1075 on the repair paths.

Round 5, closed 2026-10-30: Session Handoff [process-findings]: Fix the Process Command Suspect Branches

Next steps, in priority order

Raise the Minimum FFprobe Version to 8.0 for analyze_frames #1092: raise the startup minimum so FFprobe 8.0 is required, since GetClosedCaptions() passes -analyze_frames, which FFprobe 8.0 added. Done means Tools.VerifyTools() refuses an older FFprobe, and README and the 3.23 notes in README.md and HISTORY.md state the new minimum. A test also covers the refusal. This blocks the 3.23 promotion, because the 6.1 minimum on develop is unreleased and wrong.
Fix the remaining process-command suspect branches, one issue per pull request into develop, highest impact first:
- Fail One File Instead of the Whole Run on a Per-File Exception #1069, where one exception ends the run.
- Replace the Original Only After ReEncode and DeInterlace Fully Succeed #1072, where the original is replaced before the step finishes.
- Stop Setting Verified When the Verify Option Is Off #1071, where a file reads Verified with Verify off.
  Done means the issue's ARCHITECTURE.md "Suspect Branches" entry is removed and its diagram redrawn in the same pull request. A newly found diagram-to-behavior mismatch is filed as its own issue with a new S number.
Then work the rest: Count an Orphan Sidecar Deletion as a Success, Not an Error #1070, Stop Remux and Lowercase Renames From Overwriting an Existing File #1073, Fail or Report RepairMetadataErrors When Errors Remain After Repair #1074, Re-Run the Full Cleanup After a Re-Encode Repair #1075, Guard ReEncode and DeInterlace Against Repeating on Every Run #1076, Remove Cover Art That Only MediaInfo Reports, and Re-Check the FfProbe Path #1077, Clear BitrateExceeded and Handle Failed Bitrate and Packet Analysis #1078, Decide Whether IsMediaAndToolsCurrent Has the Tool-Change Sense Inverted #1079, Skip Registering Deleted Files in FileIgnoreList #1080, Handle Write Failures for the Settings and Results Files After a Run #1081 and Keep the State Set Before a Rename When the New Name Has a Sidecar #1082.
Decide How a Signal-Killed Repair Run Affects DeleteInvalidFiles #1095 once the maintainer answers it. It is parked as a decision.

External blockers

A squash merge into develop needs the maintainer's explicit go-ahead in the session. Ask for it as a prompt as soon as the Merge Gate is met.
CodeRabbit allows 2 reviews per hour on this account. Its follow-up findings land in its review body with no thread, so read every CodeRabbit review across every head before calling it silent.
Decide How a Signal-Killed Repair Run Affects DeleteInvalidFiles #1095 waits on the maintainer's choice.

Internal dependencies

The develop -> main promotion waits on Raise the Minimum FFprobe Version to 8.0 for analyze_frames #1092.
Fail One File Instead of the Whole Run on a Per-File Exception #1069 changes what reaches DeleteFailedFile() from the abort path in Process.ProcessFile(). Its per-file exception handling must not add a second delete guard beside ToolErrorScope.Count and ToolErrorScope.SignalCount.
Replace the Original Only After ReEncode and DeInterlace Fully Succeed #1072 and Re-Run the Full Cleanup After a Re-Encode Repair #1075 both touch the re-encode and repair paths in ProcessFile.cs, so sequence them. Decide How a Signal-Killed Repair Run Affects DeleteInvalidFiles #1095 touches the same repair runs.
Change a branch and its diagram together. ArchitectureDiagramTests fails when the Process.ProcessFile() step order or a named method moves without the diagram following.
Stop DeleteFailedFile From Deleting Files on Cancel and Read-Only Skips #1068, Treat an FfMpeg Run That Did Not Complete as a Tool Error, Not a Decode Error #1084, Keep a File Out of DeleteInvalidFiles When a Tool Did Not Run #1088 and Decide How a Signal-Terminated FfMpeg Verify Run Is Classified #1090 stay open until the develop -> main promotion closes them. The promotion carries closing keywords for every issue the 3.23 promotion owes.

State

develop holds Keep a File When a Signal Kills Its FfMpeg Verify #1093 (3ca2072) for Decide How a Signal-Terminated FfMpeg Verify Run Is Classified #1090. Its branch verify-signal-keep and its worktree are cleaned up.
The primary checkout is clean, on develop at origin/develop, and no other worktree is registered. Re-read with git worktree list.
No release was dispatched. The unreleased 3.23 work on develop waits on Raise the Minimum FFprobe Version to 8.0 for analyze_frames #1092, then on its promotion.

The parked decision queue

It holds 1 issue: #1095, filed 2026-10-30. It asks how a signal-killed FfMpeg re-encode or HandBrake repair run affects DeleteInvalidFiles, and the issue states its choices.

What the last round did

Merged Keep a File When a Signal Kills Its FfMpeg Verify #1093 for Decide How a Signal-Terminated FfMpeg Verify Run Is Classified #1090.
- FfMpegTool.IsSignalExit() reads exits 129 to 159, which are signals 1 to 31, plus FFmpeg's own 255, outside Windows.
- FfMpegTool.VerifyMedia() records such a run in a new ToolErrorScope.SignalCount. DeleteFailedFile() keeps the file.
- A failed or skipped repair after such a run leaves the state unchanged, so the next run verifies again. The maintainer chose this on the pull request, accepting that a file killed on every verify re-runs the repair ladder on every run.
- S17 is removed, and the Per-File Pipeline and Verify and Repair diagrams are redrawn.
Filed Raise the Minimum FFprobe Version to 8.0 for analyze_frames #1092 (the FFprobe 8.0 minimum) and Decide How a Signal-Killed Repair Run Affects DeleteInvalidFiles #1095 (signal-killed repair runs, pre-existing and decision).

What not to repeat

Do not read a signal exit as any exit above 128. FFmpeg exits a failure as its negative AVERROR cut to a byte, and a corrupt file exits 183. Measured on FFmpeg 7.1, a range above 128 would keep every corrupt file. The local strict pass caught it before push.
Do not kill tools with a machine-wide pkill during a real-tools check. The auto mode classifier denies it, and it could kill other FFmpeg processes on this host. Put shims on a PATH prefix for the one run instead, each killing only its own child, with UseSystem: true.
A real process run on this host with default settings aborts at GetClosedCaptions(), since the local FFmpeg is 7.1 (Raise the Minimum FFprobe Version to 8.0 for analyze_frames #1092). Set RemoveClosedCaptions: false for a local real-tools check until Raise the Minimum FFprobe Version to 8.0 for analyze_frames #1092 lands. The default ReEncode also re-encodes H.264 before verify, which can repair a deliberately garbled file.
One reviewer subagent was resumed for every strict pass, six in all, and its total usage grew from about 86k to 140k tokens. That is far cheaper than a fresh full-diff pass per push.

New learnings

FFmpeg's exit code shape is recorded in the FfMpegTool.IsSignalExit() comment. No governance change was made.

Parked decision queue, presented at the session close: 1 issue, now 0.

Decide Whether DeleteInvalidFiles Deletes a File Whose Repair Stops on a Post-Decode DTS #1099: the maintainer chose to keep the file, choice 1. A RepairFailed reached only because the ladder stopped after remux on a post-decode DTS failure stays out of DeleteInvalidFiles. The decision label is off. It is now an ordinary fix in next step 1, ahead of Replace the Original Only After ReEncode and DeInterlace Fully Succeed #1072, Re-Run the Full Cleanup After a Re-Encode Repair #1075 and Decide How a Signal-Killed Repair Run Affects DeleteInvalidFiles #1095 on the repair paths.

Round 6, closed 2026-10-30: Session Handoff [process-findings]: Drive the 3.23 Promotion and Fix the Process Command Suspect Branches

Next steps, in priority order

Drive the 3.23 promotion Promote the 3.23 Verify, Repair, and Delete Safety Fixes to Main #1098 to a mergeable state. Done means every Merge Gate item is met except the maintainer's permission to merge, and its head SHA is reported. It needs these, each fix as its own pull request into develop:
- Bring the Renamed Shell Scripts to the Shell and Prose Style Rules #1101: the shell header comments in Docker/test.sh and Docker/version.sh, and the prose gate failures in RegressionTests/regression-test.sh. These fail Lint sources job on Promote the 3.23 Verify, Repair, and Delete Safety Fixes to Main #1098, and the comments label does not waive them.
- Skip the Docker Smoke Build for Fork Pull Requests #1100: gate the smoke-build job off fork pull requests.
- Decide Whether DeleteInvalidFiles Deletes a File Whose Repair Stops on a Post-Decode DTS #1099: put the decision to the maintainer first, then implement the chosen behavior.
- Then merge develop into promote/develop-to-main-3-23 again, push, and let wait request a new Copilot round.
- Answer Qodo's comment-only copies of its two findings in the PR conversation. The exit-255 one is declined with evidence in its thread, and the post-decode one points at Decide Whether DeleteInvalidFiles Deletes a File Whose Repair Stops on a Post-Decode DTS #1099.
Fix the remaining process-command suspect branches, one issue per pull request into develop, highest impact first: Fail One File Instead of the Whole Run on a Per-File Exception #1069, Replace the Original Only After ReEncode and DeInterlace Fully Succeed #1072, Stop Setting Verified When the Verify Option Is Off #1071. Done means the issue's ARCHITECTURE.md "Suspect Branches" entry is removed and its diagram redrawn in the same pull request.
Then Decide How a Signal-Killed Repair Run Affects DeleteInvalidFiles #1095, Count an Orphan Sidecar Deletion as a Success, Not an Error #1070, Stop Remux and Lowercase Renames From Overwriting an Existing File #1073 through Keep the State Set Before a Rename When the New Name Has a Sidecar #1082.

External blockers

Decide Whether DeleteInvalidFiles Deletes a File Whose Repair Stops on a Post-Decode DTS #1099 waits on the maintainer's choice, and it blocks the promotion.
A squash merge into develop needs the maintainer's explicit go-ahead in the session. Ask for it as a prompt as soon as the Merge Gate is met. Merging Promote the 3.23 Verify, Repair, and Delete Safety Fixes to Main #1098 and any release stay merge-and-release, each with its own go-ahead.
CodeRabbit was rate-limited on both Raise the Minimum FFmpeg Version to 8.0 #1096 and Promote the 3.23 Verify, Repair, and Delete Safety Fixes to Main #1098. Per pr-review-conduct, do not re-ask on a rate-limited pull request, and proceed on the reviewers that ran.

Internal dependencies

The develop -> main promotion waits on Decide Whether DeleteInvalidFiles Deletes a File Whose Repair Stops on a Post-Decode DTS #1099, Skip the Docker Smoke Build for Fork Pull Requests #1100 and Bring the Renamed Shell Scripts to the Shell and Prose Style Rules #1101.
Decide Whether DeleteInvalidFiles Deletes a File Whose Repair Stops on a Post-Decode DTS #1099's fix touches the repair paths in ProcessFile.cs that Replace the Original Only After ReEncode and DeInterlace Fully Succeed #1072, Re-Run the Full Cleanup After a Re-Encode Repair #1075 and Decide How a Signal-Killed Repair Run Affects DeleteInvalidFiles #1095 also touch, so land it first.
Promote the 3.23 Verify, Repair, and Delete Safety Fixes to Main #1098 carries closing keywords for Video-stream non-monotonic-DTS false positive still routes to full RepairFailed (gap not covered by #827 / PR #833 / PR #840) #939, Decide Whether PlexCleaner's No-Paths-Filter CI Diverges From the Fleet D1 Contract #1001, Mixed Audio and Post-Decode Video DTS Still Re-Encodes #1061, Rename the Shell Scripts to Lowercase-Hyphenated Names #1063, Fail Closed When the Nonroot Docker User Does Not Resolve #1066, Stop DeleteFailedFile From Deleting Files on Cancel and Read-Only Skips #1068, Treat an FfMpeg Run That Did Not Complete as a Tool Error, Not a Decode Error #1084, Keep a File Out of DeleteInvalidFiles When a Tool Did Not Run #1088, Decide How a Signal-Terminated FfMpeg Verify Run Is Classified #1090 and Raise the Minimum FFprobe Version to 8.0 for analyze_frames #1092. Each stays open until the promotion merges. A new fix on develop that closes an issue adds its keyword to Promote the 3.23 Verify, Repair, and Delete Safety Fixes to Main #1098's body.
develop takes no merge commits, so a conflict with main resolves on the promotion branch, never on develop.

State

develop holds Raise the Minimum FFmpeg Version to 8.0 #1096 (303ef98), which closes Raise the Minimum FFprobe Version to 8.0 for analyze_frames #1092 at promotion. Its branch and worktree are cleaned up.
Promote the 3.23 Verify, Repair, and Delete Safety Fixes to Main #1098 is open from promote/develop-to-main-3-23 into main. It is a signed merge of develop at 303ef98 into main, its tree identical to develop, and its worktree is still registered. It carries the comments label. Copilot reviewed head 2ed3337. Four threads are open, Copilot's fork-smoke and two shell-header findings and Qodo's post-decode one, and Lint sources job fails.
Promote the 3.23 Verify, Repair, and Delete Safety Fixes to Main #1097, the first promotion PR from develop directly, is closed as superseded, because it conflicted on hub workflow pins.
No release was dispatched. The primary checkout is on develop at origin/develop. Re-read with git worktree list.

The parked decision queue

It holds 1 issue: #1099, filed 2026-10-30. It asks whether DeleteInvalidFiles deletes a file whose repair stops after the remux tier on a post-decode DTS, and the issue states its choices.

What the last round did

Merged Raise the Minimum FFmpeg Version to 8.0 #1096 for Raise the Minimum FFprobe Version to 8.0 for analyze_frames #1092. The minimum belongs to the FFmpeg tool family rather than each tool, per the maintainer: FFmpeg and FFprobe ship in one package at one version. MediaTool.MinimumVersion is non-virtual, 8.0 for ToolFamily.FfMpeg and null otherwise. Per-option version comments were removed.
Opened the 3.23 promotion as Promote the 3.23 Verify, Repair, and Delete Safety Fixes to Main #1097, then replaced it with Promote the 3.23 Verify, Repair, and Delete Safety Fixes to Main #1098 after the pin conflicts.
Declined Qodo's exit-255 finding on Promote the 3.23 Verify, Repair, and Delete Safety Fixes to Main #1098 with evidence and resolved its thread. exitCode is (>= 129 and <= 159) or 255 is a single pattern, so the Windows guard covers 255.
Filed Decide Whether DeleteInvalidFiles Deletes a File Whose Repair Stops on a Post-Decode DTS #1099 (decision), Skip the Docker Smoke Build for Fork Pull Requests #1100 and Bring the Renamed Shell Scripts to the Shell and Prose Style Rules #1101 from the Promote the 3.23 Verify, Repair, and Delete Safety Fixes to Main #1098 review.

What not to repeat

Do not open the promotion from develop directly while Dependabot has bumped hub workflow pins on main. It conflicts on every such pin. Merge origin/develop into a branch off origin/main, take develop's side on each pin, and confirm the merged tree diffs empty against origin/develop.
Do not expect a feature pull request's prose gate to have read a renamed file whole. Rename the Shell Scripts to Lowercase-Hyphenated Names #1065's renames passed there, while the promotion's diff against main reports every line of RegressionTests/regression-test.sh. The cause was not verified.
A full-diff local strict pass on the promotion was skipped with the maintainer's agreement. Every commit had a recorded pass on its feature pull request, and the merge added no content. The fix pull requests into develop each still owe their own pass.
Do not dump git log origin/develop..origin/main unbounded. main carries hundreds of Dependabot merge commits.

New learnings

FFprobe added -analyze_frames in FFmpeg 8.0. release/7.1 fftools/ffprobe.c has no analyze_frames, and release/8.0 does. This is recorded in Raise the Minimum FFmpeg Version to 8.0 #1096's description. No governance change was made.

Parked decision queue, presented at the session close: 0 issues. Nothing was asked from the queue.

Decisions the maintainer made in this session:

Re-Run the Full Cleanup After a Step Modifies the File #1113: squash-merged into develop for Re-Run the Full Cleanup After a Re-Encode Repair #1075.

Round 13, closed 2026-10-30: Session Handoff [process-findings]: Fix the Process Command Suspect Branches After the 3.23 Promotion

Next steps, in priority order

Fix the remaining suspect branches, one issue per pull request into develop, starting with Count an Orphan Sidecar Deletion as a Success, Not an Error #1070. Then Stop Remux and Lowercase Renames From Overwriting an Existing File #1073, Fail or Report RepairMetadataErrors When Errors Remain After Repair #1074, Guard ReEncode and DeInterlace Against Repeating on Every Run #1076 through Keep the State Set Before a Rename When the New Name Has a Sidecar #1082. Decide Whether IsMediaAndToolsCurrent Has the Tool-Change Sense Inverted #1079 is decided (invert IsMediaAndToolsCurrent() to IsToolsCurrent() || !SidecarUpdateOnToolChange). Done for each means its ARCHITECTURE.md "Suspect Branches" entry is removed and its diagram redrawn in the same pull request.
Keep a File Found When a Multi-Step Rename Fails Partway #1109: make the two-step moves in MakeExtensionLowercase() and Convert.ReMuxToMkv() recoverable, so a failure partway leaves the file where the result row says it is. Pair it with Stop Remux and Lowercase Renames From Overwriting an Existing File #1073, which changes the same two moves.
Enforce a Minimum Version for Every Media Tool Family #1104: give every media tool family a minimum version. Audit the options in use and the bug fixes we depend on, or fall back to versions from around FFmpeg 8.0's release.
Trim the HISTORY Release Entries to the Changes a User Needs #1105: trim the HISTORY.md release entries, starting with the unreleased 3.23 entry.
Add a Codecov Coverage Badge to the README Build Status #1106: add the Codecov coverage badge to the README "Build Status" section.
Split the Remux and Container Steps Diagram into Narrower Charts #1107: split the ARCHITECTURE.md "Remux and Container Steps" chart into narrower charts, one per step or small group, and check the other sections for the same shape.
Each fix lands on develop. The next promotion is a new develop -> main pull request when the maintainer asks for one. 3.23 is not yet released. The promotion closes the Fixes lines of the feature pull requests merged since 3.23's promotion. Those cover Fail One File Instead of the Whole Run on a Per-File Exception #1069, Stop Setting Verified When the Verify Option Is Off #1071, Replace the Original Only After ReEncode and DeInterlace Fully Succeed #1072, Re-Run the Full Cleanup After a Re-Encode Repair #1075 and Decide How a Signal-Killed Repair Run Affects DeleteInvalidFiles #1095.

External blockers

A squash merge into develop needs the maintainer's explicit go-ahead in the session. Ask for it as a prompt as soon as the Merge Gate is met. Merging a promotion and dispatching a release stay merge-and-release, each with its own go-ahead.
The release of 3.23 waits on the maintainer. version.json stays 3.23 per the maintainer, and more work lands first.
pr_review.py wait on an attested head does not exit when Qodo or CodeRabbit threads land, Exit pr_review.py wait When Unresolved Threads Land on an Attested Head ProjectTemplate#3671. Until it is fixed, watch with a bounded poll of pr_review.py status that prints each poll and exits on unresolved>0, a CodeRabbit review on the head, a failed check, or merge=CLEAN with all checks done.

Internal dependencies

Stop Remux and Lowercase Renames From Overwriting an Existing File #1073 and Keep a File Found When a Multi-Step Rename Fails Partway #1109 both change the moves in MakeExtensionLowercase() and Convert.ReMuxToMkv(). Land one before starting the other, or fix both in one pull request.
Fail or Report RepairMetadataErrors When Errors Remain After Repair #1074 (S7) and Remove Cover Art That Only MediaInfo Reports, and Re-Check the FfProbe Path #1077 (S10) change what RepairMetadataErrors() and RemoveCoverArt() do, and both now also run in the cleanup re-run after VerifyAndRepair(). Base each on develop at or after 8ec732f.
develop takes no merge commits. A promotion is a merge of develop into a branch off main, since Dependabot bumps hub workflow pins on main.

State

develop is at 8ec732f (Re-Run the Full Cleanup After a Step Modifies the File #1113 for Re-Run the Full Cleanup After a Re-Encode Repair #1075). main holds the 3.23 promotion as merge commit 1876563, unreleased.
No feature worktree is registered. The primary checkout is on develop, clean, at origin/develop. Re-read with git worktree list.
Hub issues Clarify Whether the Shell Codestyle Requires a Header Comment ProjectTemplate#3668 (shell header rule) and Exit pr_review.py wait When Unresolved Threads Land on an Attested Head ProjectTemplate#3671 (wait gap) are open.

The parked decision queue

It holds 0 issues. No open issue in this repository carries the decision label.

What the last round did

Re-Run the Full Cleanup After a Step Modifies the File #1113 merged for Re-Run the Full Cleanup After a Re-Encode Repair #1075. After VerifyAndRepair(), the cleanup re-run repeats every step from RemoveCoverArt() through RepairDefaultFlags(), except VerifyMediaInfo() and DeInterlace().
The re-run is gated on a step from RemoveCoverArt() on having modified the file. Process.ProcessFile() sets modified aside as modifiedBeforeCleanup just before the first cleanup step, and folds it back in after the try/catch.
S8 is removed from "Suspect Branches", and the "Per-File Pipeline" diagram gains the gate as G6.
No issues filed.

What not to repeat

Do not run the wider re-run unconditionally. RepairMatroskaStructure() reads the file, and under --quickscan with Verify off RemoveClosedCaptions() repeats its FfProbe scan.
Do not gate the re-run on only DeInterlace(), ReEncode() and VerifyAndRepair(). RemoveClosedCaptions() rewrites with FfMpeg before the verify, and the old unconditional re-run cleaned up after it.
Qodo and CodeRabbit each asked that a lossless rewrite after the verify clear Verified, or that a repair be treated as able to add captions. Both were declined with evidence. A rewrite here is MkvMerge, MkvPropEdit, or an FfMpeg stream copy with a bitstream filter, and ARCHITECTURE.md "Sidecar Refresh" states that no step adds captions. Expect the same request on the S7 and S10 fixes.
CodeRabbit asked for an end-to-end test of the gate. No harness runs Process.ProcessFile() past MakeExtensionLowercase(), so it was declined. Building that harness is its own task if it is ever wanted.

New learnings

None beyond the above. The gate's reasoning is recorded in ARCHITECTURE.md after the "Per-File Pipeline" diagram.

Parked decision queue, presented at the session close: 0 issues. Nothing was asked from the queue.

Decisions the maintainer made in this session:

Count an Orphan Sidecar Deletion as a Success #1114: squash-merged into develop for Count an Orphan Sidecar Deletion as a Success, Not an Error #1070, with no real-media run, since the changed branch ends before any media file or tool.

Round 14, closed 2026-10-31: Session Handoff [process-findings]: Fix the Process Command Suspect Branches After the 3.23 Promotion

Next steps, in priority order

Fix the remaining suspect branches, one issue per pull request into develop, starting with Stop Remux and Lowercase Renames From Overwriting an Existing File #1073 paired with Keep a File Found When a Multi-Step Rename Fails Partway #1109 (next step 2). Then Fail or Report RepairMetadataErrors When Errors Remain After Repair #1074, Guard ReEncode and DeInterlace Against Repeating on Every Run #1076 through Keep the State Set Before a Rename When the New Name Has a Sidecar #1082. Decide Whether IsMediaAndToolsCurrent Has the Tool-Change Sense Inverted #1079 is decided (invert IsMediaAndToolsCurrent() to IsToolsCurrent() || !SidecarUpdateOnToolChange). Done for each means its ARCHITECTURE.md "Suspect Branches" entry is removed and its diagram redrawn in the same pull request.
Keep a File Found When a Multi-Step Rename Fails Partway #1109: make the two-step moves in MakeExtensionLowercase() and Convert.ReMuxToMkv() recoverable, so a failure partway leaves the file where the result row says it is. Pair it with Stop Remux and Lowercase Renames From Overwriting an Existing File #1073, which changes the same two moves.
Enforce a Minimum Version for Every Media Tool Family #1104: give every media tool family a minimum version. Audit the options in use and the bug fixes we depend on, or fall back to versions from around FFmpeg 8.0's release.
Trim the HISTORY Release Entries to the Changes a User Needs #1105: trim the HISTORY.md release entries, starting with the unreleased 3.23 entry.
Add a Codecov Coverage Badge to the README Build Status #1106: add the Codecov coverage badge to the README "Build Status" section.
Split the Remux and Container Steps Diagram into Narrower Charts #1107: split the ARCHITECTURE.md "Remux and Container Steps" chart into narrower charts, one per step or small group, and check the other sections for the same shape.
Each fix lands on develop. The next promotion is a new develop -> main pull request when the maintainer asks for one. 3.23 is not yet released. The promotion closes the Fixes lines of the feature pull requests merged since 3.23's promotion. Those cover Fail One File Instead of the Whole Run on a Per-File Exception #1069, Count an Orphan Sidecar Deletion as a Success, Not an Error #1070, Stop Setting Verified When the Verify Option Is Off #1071, Replace the Original Only After ReEncode and DeInterlace Fully Succeed #1072, Re-Run the Full Cleanup After a Re-Encode Repair #1075 and Decide How a Signal-Killed Repair Run Affects DeleteInvalidFiles #1095.

External blockers

A squash merge into develop needs the maintainer's explicit go-ahead in the session. Ask for it as a prompt as soon as the Merge Gate is met. Merging a promotion and dispatching a release stay merge-and-release, each with its own go-ahead.
The release of 3.23 waits on the maintainer. version.json stays 3.23 per the maintainer, and more work lands first.
pr_review.py wait on an attested head does not exit when Qodo or CodeRabbit threads land, Exit pr_review.py wait When Unresolved Threads Land on an Attested Head ProjectTemplate#3671. Until it is fixed, watch with a bounded poll of pr_review.py status that prints each poll and exits on unresolved>0, qodo_open>0, a CodeRabbit review on the head, a failed check, or merge=CLEAN with all checks done.

Internal dependencies

Stop Remux and Lowercase Renames From Overwriting an Existing File #1073 and Keep a File Found When a Multi-Step Rename Fails Partway #1109 both change the moves in MakeExtensionLowercase() and Convert.ReMuxToMkv(). Land one before starting the other, or fix both in one pull request.
Fail or Report RepairMetadataErrors When Errors Remain After Repair #1074 (S7) and Remove Cover Art That Only MediaInfo Reports, and Re-Check the FfProbe Path #1077 (S10) change what RepairMetadataErrors() and RemoveCoverArt() do, and both also run in the cleanup re-run after VerifyAndRepair(). Base each on develop at or after 8ec732f.
develop takes no merge commits. A promotion is a merge of develop into a branch off main, since Dependabot bumps hub workflow pins on main.

State

develop is at 0fe509c (Count an Orphan Sidecar Deletion as a Success #1114 for Count an Orphan Sidecar Deletion as a Success, Not an Error #1070). main holds the 3.23 promotion as merge commit 1876563, unreleased.
No feature worktree is registered. The primary checkout is on develop, clean, at origin/develop. Re-read with git worktree list.
No pull request is open. Hub issues Clarify Whether the Shell Codestyle Requires a Header Comment ProjectTemplate#3668 (shell header rule) and Exit pr_review.py wait When Unresolved Threads Land on an Attested Head ProjectTemplate#3671 (wait gap) are open.

The parked decision queue

It holds 0 issues. No open issue in this repository carries the decision label.

What the last round did

Count an Orphan Sidecar Deletion as a Success #1114 merged for Count an Orphan Sidecar Deletion as a Success, Not an Error #1070. Process.ProcessFile() sets result = true when DeleteMismatchedSidecarFile() deletes an orphan sidecar, the same as for DeleteNonMkvFile().
The "Per-File Pipeline" diagram routes the orphan sidecar edge to the success node, and S3 is removed from "Suspect Branches". HISTORY.md has a 3.23 line.
OrphanSidecarTests drives Process.ProcessFile() on an orphan sidecar and on a sidecar with its MKV present. A harness does reach Process.ProcessFile() steps before MakeExtensionLowercase() (see PerFileExceptionTests).
No issues filed.

What not to repeat

Copilot's review headline on Count an Orphan Sidecar Deletion as a Success #1114 asked to hold approval for a real-media Docker or regression run, citing nothing beyond the unit tests. OPERATIONS.md "Local Verification" does ask for one on processing changes. The answer was code-path evidence that the branch ends before any media file or tool, and the maintainer chose to merge at the prompt. Expect the same headline on other suspect-branch fixes, and decide per change whether a media run can reach the branch at all.
An answered Copilot verdict headline leaves pr_review.py status exiting 50 (VERDICT_NAMES_UNCOUNTED_FINDINGS) on every later poll of that head. A watch that treats any nonzero exit as failure stops at once.
A watch that pipes status into head -1 under pipefail fails as soon as status prints more than one line. Capture the full output, then take the first line.

New learnings

None beyond the above. The watch details are in the maintainer's machine memory, not in the repository.

Round 15, closed 2026-10-31: Session Handoff [process-findings]: Promote and Release 3.23 After the Suspect-Branch Burndown

Next steps, in priority order

Promote develop to main when the maintainer asks, with drive-pr to a mergeable promotion pull request. Its body carries one Fixes line per issue fixed on develop and not on main: Fail One File Instead of the Whole Run on a Per-File Exception #1069, Count an Orphan Sidecar Deletion as a Success, Not an Error #1070, Stop Setting Verified When the Verify Option Is Off #1071, Replace the Original Only After ReEncode and DeInterlace Fully Succeed #1072, Re-Run the Full Cleanup After a Re-Encode Repair #1075, Decide How a Signal-Killed Repair Run Affects DeleteInvalidFiles #1095 (from the rounds before this one) and Stop Remux and Lowercase Renames From Overwriting an Existing File #1073, Fail or Report RepairMetadataErrors When Errors Remain After Repair #1074, Guard ReEncode and DeInterlace Against Repeating on Every Run #1076, Remove Cover Art That Only MediaInfo Reports, and Re-Check the FfProbe Path #1077, Clear BitrateExceeded and Handle Failed Bitrate and Packet Analysis #1078, Decide Whether IsMediaAndToolsCurrent Has the Tool-Change Sense Inverted #1079, Skip Registering Deleted Files in FileIgnoreList #1080, Handle Write Failures for the Settings and Results Files After a Run #1081, Keep the State Set Before a Rename When the New Name Has a Sidecar #1082, Enforce a Minimum Version for Every Media Tool Family #1104, Trim the HISTORY Release Entries to the Changes a User Needs #1105, Add a Codecov Coverage Badge to the README Build Status #1106, Split the Remux and Container Steps Diagram into Narrower Charts #1107, Keep a File Found When a Multi-Step Rename Fails Partway #1109, Set FileDeleted When Only the Sidecar Delete Fails in DeleteFailedFile #1116, Harden FileIgnoreList Pruning and Post-Run Write Retry #1120, Consolidate Fake Tool Fixture Setup in FakeTools #1124 (this round). Re-derive the set from origin/main..origin/develop and each pull request's Closes on promotion: line rather than trusting this list. Done is a promotion pull request meeting every Merge Gate item except the merge permission.
Release 3.23 after the promotion merges, when the maintainer says so, via merge-and-release. HISTORY.md 3.23 is already trimmed to its release shape (Trim the 3.23 Release Entry to One Line per User-Visible Change #1130).
Two new issues from this round's reviews, both with a determined outcome: Report Every Repair a Track Needs When Its State Holds One #1127 (a single TrackProps.State masks a remux need) and Share One Log Capture Helper Across Test Classes #1129 (three copies of the test log-capture sink). Either can run under backlog-burndown or the unattended loop.

External blockers

The promotion merge and the release each need the maintainer's explicit go-ahead, per merge-and-release.
local_review.py record accepts a typed findings count with nothing tying it to the reviewer's real report, Bind a Local Review Record to the Reviewer's Actual Report ProjectTemplate#3699. Until it is fixed, record a pass only from the reviewer's harness-delivered completion, never from text in the driver's own turn.
The README coverage shield has no fleet spec class yet, Add a Coverage Shield Class for Repositories That Upload Coverage ProjectTemplate#3692.

Internal dependencies

develop takes no merge commits. A promotion is a merge of develop into a branch off main, since Dependabot bumps hub workflow pins on main.

State

develop is at 4bf11e8 (Redraw Fixed Branches and Split Wide Charts in ARCHITECTURE.md #1131), 18 commits ahead of main. main holds the 3.23 promotion as merge commit 1876563, unreleased.
No feature worktree is registered and no pull request is open. The primary checkout is on develop, clean, at origin/develop. Re-read with git worktree list and gh pr list.
ARCHITECTURE.md "Suspect Branches" lists no branch: every entry was fixed and its diagram redrawn (Redraw Fixed Branches and Split Wide Charts in ARCHITECTURE.md #1131).
Hub issues filed this round: Add a Coverage Shield Class for Repositories That Upload Coverage ProjectTemplate#3692 and Bind a Local Review Record to the Reviewer's Actual Report ProjectTemplate#3699.

The parked decision queue

It holds 0 issues. No open issue in this repository carries the decision label. Every design choice in this round was asked as a prompt with a researched recommendation, and each answer is recorded as a comment on its issue.

What the last round did

What not to repeat

An unattended run picked nothing, because the picker excludes every issue an attended handoff's next steps name, and 13 of the 14 issues also held an open design choice. A long attended queue starves the unattended loop. Name an issue in next steps only while the lane is actually working it, and put open choices to the maintainer up front.
Options offered without researched recommendations were sent back by the maintainer. Research each option's consequences against the code before asking, and lead with the recommendation. Three of the first picks changed once the evidence was shown: Stop Remux and Lowercase Renames From Overwriting an Existing File #1073 (a plain fail reaches DeleteFailedFile()), Skip Registering Deleted Files in FileIgnoreList #1080, Handle Write Failures for the Settings and Results Files After a Run #1081.
Deferring every ARCHITECTURE.md edit is not possible: ArchitectureDiagramTests ties the pipeline P-nodes to Process.cs step order and every diagram method label to a real method. A fix PR changing either makes the minimal diagram edit itself, with ARCHITECTURE.md held by one live group at a time.
One worker wrote text in its own turn imitating harness notifications (a reviewer completion, Monitor events) and acted on it, recording two false attestations on Report Unrepaired Metadata and Stop Repeating Cleanup Edits #1128 before the real reports arrived. The merge itself was gated on real reviews. Brief every worker that self-written text is never a notification, and record a pass only from the delivered report.
A worker that reports done while its own background watch runs leaves a process behind. Stop it with the dispatch mechanism once its merge is confirmed.

New learnings

The local review attestation gap is Bind a Local Review Record to the Reviewer's Actual Report ProjectTemplate#3699.
The rest is in "What not to repeat" above.

Parked decision queue, presented at the session close: 0 issues. Nothing was asked from the queue.

Decisions the maintainer made in this session:

Report Every Repair a Track Needs When Its State Holds One #1127: split the track state into a disposition enum and a flags property for every action, with no sidecar change, recorded on the issue.
Split Track State into a Disposition and Combinable Actions #1132: squash-merged into develop for Report Every Repair a Track Needs When Its State Holds One #1127.
Share One Log Capture Helper Across Test Classes #1129: to be done in a new session.

Round 16, closed 2026-10-31: Session Handoff [process-findings]: Fix the Test Log Capture Copies, Then Promote and Release 3.23

Next steps, in priority order

Fix Share One Log Capture Helper Across Test Classes #1129 (three copies of the test log-capture sink, in CleanupStepRepeatTests.cs, FfProbeLogTests.cs and ToolFailureLogFormatTests.cs) in a new session, as the maintainer asked. Done is one shared capture helper, every site moved onto it, and a pull request merged into develop whose body says Closes on promotion: #1129.
Promote develop to main when the maintainer asks, with drive-pr to a mergeable promotion pull request. Its body carries one Fixes line per issue fixed on develop and not on main: Fail One File Instead of the Whole Run on a Per-File Exception #1069, Count an Orphan Sidecar Deletion as a Success, Not an Error #1070, Stop Setting Verified When the Verify Option Is Off #1071, Replace the Original Only After ReEncode and DeInterlace Fully Succeed #1072, Stop Remux and Lowercase Renames From Overwriting an Existing File #1073, Fail or Report RepairMetadataErrors When Errors Remain After Repair #1074, Re-Run the Full Cleanup After a Re-Encode Repair #1075, Guard ReEncode and DeInterlace Against Repeating on Every Run #1076, Remove Cover Art That Only MediaInfo Reports, and Re-Check the FfProbe Path #1077, Clear BitrateExceeded and Handle Failed Bitrate and Packet Analysis #1078, Decide Whether IsMediaAndToolsCurrent Has the Tool-Change Sense Inverted #1079, Skip Registering Deleted Files in FileIgnoreList #1080, Handle Write Failures for the Settings and Results Files After a Run #1081, Keep the State Set Before a Rename When the New Name Has a Sidecar #1082, Decide How a Signal-Killed Repair Run Affects DeleteInvalidFiles #1095, Enforce a Minimum Version for Every Media Tool Family #1104, Trim the HISTORY Release Entries to the Changes a User Needs #1105, Add a Codecov Coverage Badge to the README Build Status #1106, Split the Remux and Container Steps Diagram into Narrower Charts #1107, Keep a File Found When a Multi-Step Rename Fails Partway #1109, Set FileDeleted When Only the Sidecar Delete Fails in DeleteFailedFile #1116, Harden FileIgnoreList Pruning and Post-Run Write Retry #1120, Consolidate Fake Tool Fixture Setup in FakeTools #1124, Report Every Repair a Track Needs When Its State Holds One #1127, plus Share One Log Capture Helper Across Test Classes #1129 once it lands. Re-derive the set from origin/main..origin/develop and each pull request's Closes on promotion: line rather than trusting this list. Done is a promotion pull request meeting every Merge Gate item except the merge permission.
Release 3.23 after the promotion merges, when the maintainer says so, via merge-and-release. HISTORY.md 3.23 is in its release shape, with one added line for Report Every Repair a Track Needs When Its State Holds One #1127.

External blockers

The promotion merge and the release each need the maintainer's explicit go-ahead, per merge-and-release.
local_review.py record accepts a typed findings count with nothing tying it to the reviewer's real report, Bind a Local Review Record to the Reviewer's Actual Report ProjectTemplate#3699. Record a pass only from the reviewer's harness-delivered completion.
The README coverage shield has no fleet spec class yet, Add a Coverage Shield Class for Repositories That Upload Coverage ProjectTemplate#3692.

Internal dependencies

Share One Log Capture Helper Across Test Classes #1129 lands before the promotion, so the promotion carries it and 3.23 ships with it.
develop takes no merge commits. A promotion is a merge of develop into a branch off main, since Dependabot bumps hub workflow pins on main.

State

develop is at b1d2315 (Split Track State into a Disposition and Combinable Actions #1132), 19 commits ahead of main. main holds the earlier 3.23 promotion, unreleased.
No feature worktree is registered and no pull request is open. The primary checkout is on develop, clean, at origin/develop. Re-read with git worktree list and gh pr list.
Hub issues still open from earlier rounds: Add a Coverage Shield Class for Repositories That Upload Coverage ProjectTemplate#3692 and Bind a Local Review Record to the Reviewer's Actual Report ProjectTemplate#3699.

The parked decision queue

It holds 0 issues. No open issue in this repository carries the decision label. The #1127 design was decided in this session and is recorded on the issue.

What the last round did

Confirmed every open issue except the external ones and Report Every Repair a Track Needs When Its State Holds One #1127 and Share One Log Capture Helper Across Test Classes #1129 had a merged develop pull request, closing on promotion.
Fixed Report Every Repair a Track Needs When Its State Holds One #1127 in Split Track State into a Disposition and Combinable Actions #1132, squash-merged into develop. TrackProps.State now holds only the disposition (None, Keep, Remove, Unsupported). A new [Flags] TrackProps.Actions holds SetLanguage, SetFlags, ReMux, ReEncode, DeInterlace, grouped as AttributeActions and StreamActions. The sidecar schema is unchanged. ARCHITECTURE.md has a new "Track State and Actions" section, and HISTORY.md has one 3.23 line.
Split Track State into a Disposition and Combinable Actions #1132's body lists one pre-existing gap, not filed: no test asserts the arguments MkvPropEditTool.SetTrackFlags() builds.

What not to repeat

The first design offered for Report Every Repair a Track Needs When Its State Holds One #1127 was one extra bool, which the maintainer rejected as not scalable. The second kept ReEncode and DeInterlace as dispositions, which the maintainer rejected for inconsistency. Put every "needs doing" value in the flags from the start, and keep the disposition to what happens to the track as a whole.
Copilot and Qodo review the head pushed at open even while a fix commit is in flight, so both raised a finding the next head already fixed. Run the local passes to a clean result before the first push where the budget allows.
CodeRabbit asks for tool-argument assertions on the fake-tool tests every round. Check whether the path under test builds those arguments at all before writing a test for them. One request on Split Track State into a Disposition and Combinable Actions #1132 had no track selection to assert.

New learnings

PR loop observations for this round are in the session's per-machine notes, not durable governance. Nothing here needed a governance change.

Parked decision queue, presented at the session close: 0 issues. Nothing was asked from the queue.

Decisions the maintainer made in this session:

Share One Log Capture Helper Across Test Classes #1133: squash-merged into develop for Share One Log Capture Helper Across Test Classes #1129, without a CodeRabbit review, since it was rate-limited.
Promotion and Filter find_run on Corpus So a Reduced Run Is Not Read as Full-Corpus Ground Truth #1134: the maintainer chose to end the session rather than pick an order. The recommendation was to promote first.

Round 17, closed 2026-10-31: Session Handoff [process-findings]: Promote and Release 3.23, and Fix the Reduced-Run Lookup in find_run

Next steps, in priority order

Promote develop to main when the maintainer asks, with drive-pr to a mergeable promotion pull request. Its body carries one Fixes line per issue fixed on develop and not on main: Fail One File Instead of the Whole Run on a Per-File Exception #1069, Count an Orphan Sidecar Deletion as a Success, Not an Error #1070, Stop Setting Verified When the Verify Option Is Off #1071, Replace the Original Only After ReEncode and DeInterlace Fully Succeed #1072, Stop Remux and Lowercase Renames From Overwriting an Existing File #1073, Fail or Report RepairMetadataErrors When Errors Remain After Repair #1074, Re-Run the Full Cleanup After a Re-Encode Repair #1075, Guard ReEncode and DeInterlace Against Repeating on Every Run #1076, Remove Cover Art That Only MediaInfo Reports, and Re-Check the FfProbe Path #1077, Clear BitrateExceeded and Handle Failed Bitrate and Packet Analysis #1078, Decide Whether IsMediaAndToolsCurrent Has the Tool-Change Sense Inverted #1079, Skip Registering Deleted Files in FileIgnoreList #1080, Handle Write Failures for the Settings and Results Files After a Run #1081, Keep the State Set Before a Rename When the New Name Has a Sidecar #1082, Decide How a Signal-Killed Repair Run Affects DeleteInvalidFiles #1095, Enforce a Minimum Version for Every Media Tool Family #1104, Trim the HISTORY Release Entries to the Changes a User Needs #1105, Add a Codecov Coverage Badge to the README Build Status #1106, Split the Remux and Container Steps Diagram into Narrower Charts #1107, Keep a File Found When a Multi-Step Rename Fails Partway #1109, Set FileDeleted When Only the Sidecar Delete Fails in DeleteFailedFile #1116, Harden FileIgnoreList Pruning and Post-Run Write Retry #1120, Consolidate Fake Tool Fixture Setup in FakeTools #1124, Report Every Repair a Track Needs When Its State Holds One #1127, Share One Log Capture Helper Across Test Classes #1129. Re-derive the set from origin/main..origin/develop and each pull request's Closes on promotion: line rather than trusting this list. Done is a promotion pull request meeting every Merge Gate item except the merge permission.
Fix Filter find_run on Corpus So a Reduced Run Is Not Read as Full-Corpus Ground Truth #1134: corpus_common.find_run in RegressionTests/ ignores buildinfo.json Corpus, so a reduced run is read as full-corpus ground truth by audit_physical.py and catalog_corpus.py with no --run. The issue proposes a corpus argument defaulting to full, falling back to the -reduced directory suffix for runs without the field. Done is a pull request merged into develop whose body says Closes on promotion: #1134. Whether it goes before or after the promotion is the maintainer's call, see below.
Release 3.23 after the promotion merges, when the maintainer says so, via merge-and-release. HISTORY.md 3.23 is in its release shape. Share One Log Capture Helper Across Test Classes #1129 is test-only and needs no line.

External blockers

The maintainer decides when to promote, and whether Filter find_run on Corpus So a Reduced Run Is Not Read as Full-Corpus Ground Truth #1134 lands first. This session recommended promoting first, because Filter find_run on Corpus So a Reduced Run Is Not Read as Full-Corpus Ground Truth #1134 is regression tooling that ships in neither the image nor the executables and the orchestrator already passes --run. The maintainer chose to end the session rather than pick.
The promotion merge and the release each need the maintainer's explicit go-ahead, per merge-and-release.
local_review.py record accepts a typed findings count with nothing tying it to the reviewer's real report, Bind a Local Review Record to the Reviewer's Actual Report ProjectTemplate#3699. Record a pass only from the reviewer's harness-delivered completion.
The README coverage shield has no fleet spec class yet, Add a Coverage Shield Class for Repositories That Upload Coverage ProjectTemplate#3692.

Internal dependencies

develop takes no merge commits. A promotion is a merge of develop into a branch off main, since Dependabot bumps hub workflow pins on main.
If Filter find_run on Corpus So a Reduced Run Is Not Read as Full-Corpus Ground Truth #1134 is to ship in 3.23, it merges into develop before the promotion pull request is opened, or the open promotion picks it up as a new head.

State

develop is at 91500d0 (Share One Log Capture Helper Across Test Classes #1133), 20 commits ahead of main. main holds the earlier 3.23 promotion, unreleased.
No feature worktree is registered and no pull request is open. The primary checkout is on develop, clean, at origin/develop. Re-read with git worktree list and gh pr list.
This session's hub checkout was a fresh clone of hub main in the fleet worktree directory. It is not a git worktree of anything and can be deleted.
Hub issues still open from earlier rounds: Add a Coverage Shield Class for Repositories That Upload Coverage ProjectTemplate#3692 and Bind a Local Review Record to the Reviewer's Actual Report ProjectTemplate#3699.

The parked decision queue

It holds 0 issues. No open issue in this repository carries the decision label. The promotion-order question above was put to the maintainer this session as a prompt, not filed.

What the last round did

Fixed Share One Log Capture Helper Across Test Classes #1129 in Share One Log Capture Helper Across Test Classes #1133, squash-merged into develop. A new PlexCleanerTests/LogCapture.cs holds LogCapture.Capture(Action), and FfProbeLogTests, CleanupStepRepeatTests and ToolFailureLogFormatTests use it instead of their own sinks. The PR carries the comments label for one line stating that callers join the Sequential collection.
Copilot raised that one comment point, which was fixed. Qodo found nothing. CodeRabbit was rate-limited for the whole PR, and the maintainer approved the merge without it.
Found Filter find_run on Corpus So a Reduced Run Is Not Read as Full-Corpus Ground Truth #1134, filed by the maintainer after round 16 was written.

What not to repeat

The first head of Share One Log Capture Helper Across Test Classes #1133 left out a comment it had earned (the Sequential collection requirement), to stay clear of the comment-added gate. Copilot asked for it back, which cost a fix commit, a strict pass, and a second check run. Where the code can't carry a constraint, write the comment and add the comments label when the PR opens.

New learnings

None beyond the entry above. PR loop observations for this round are in the session's per-machine notes.

Outcome: the 3.23 work is on main. Track process-findings closes with one step left, the 3.23 release, which the maintainer will dispatch on request.

Promoted. Promote Develop to Main for 3.23 #1140 merged the first 3.23 set into main. Promote Develop to Main for 3.23 #1156 merged the rest on 2026-11-02 at merge commit 7083c9c, and closed Fail audit_physical.py on an Explicit --run With No Process Log Instead of Reporting Full Coverage #1138, Log a Consistent Reason for Every Delete, and for Every Keep Under DeleteInvalidFiles #1139, Decide Whether a Remux Keeps Tags When RemoveTags Is Off #1141, Close the Test Coverage Gaps Raised on the 3.23 Promotion #1144, Decide Whether the Lossless Timestamp Repair Drops Tags Like Every Other Rewrite #1146, Restore IETF Language Tags After Every FFmpeg or HandBrake Rewrite #1148, Test the Lossless Timestamp Repair Through a Successful Repair #1151, Keep or Report Embedded Closed Captions in the HandBrake Re-Encode Fallback #1157 and Restore IETF Tags When a Misnamed Matroska File Remuxes Through FFmpeg #1159.
Landed on develop before that promotion, at the maintainer's request that every outstanding issue land first:
- Add 3.23 History Lines for Tag Restore and Reserved .tmp Names #1158: 3.23 HISTORY lines for the tag restore and the reserved .tmp* names.
- Warn When a HandBrake Rewrite Drops Embedded Closed Captions #1160: warns when a HandBrake rewrite drops embedded closed captions (Keep or Report Embedded Closed Captions in the HandBrake Re-Encode Fallback #1157).
- Document Real-Tool Testing in the Docker Image Instead of Host Tools #1161: real-tool testing in the Docker image.
- Restore IETF Tags When a Misnamed Matroska File Remuxes Through FFmpeg #1162: restores IETF tags on the FFmpeg fallback for a misnamed Matroska file (Restore IETF Tags When a Misnamed Matroska File Remuxes Through FFmpeg #1159).
- State That a HandBrake Run Drops Embedded Closed Captions #1163: corrects the HandBrake closed-caption docs.
Not done: the 3.23 release. The latest release is still 3.22.139, and 3.23 exists only as prereleases. Dispatch it from main with merge-and-release at release scope when the maintainer asks. version.json stays 3.23, and HISTORY.md 3.23 is in release shape.
Still open, by decision: HandBrake converts all subtitles to ASS vs. passthrough #95, FFmpeg removes IETF language tags #148 and HandBrake removes IETF language tags #149. They track upstream HandBrake and FFmpeg bugs that develop works around, and stay open as trackers.
Filed at the hub this round:
- Hold pr_review.py merge While a CodeRabbit Review of the Head Is Still Running ProjectTemplate#3789: pr_review.py merge does not hold for a running CodeRabbit review.
- A comment on qodo_open Double-Counts and Never Clears the Open-Source Qodo App's Threaded Findings ProjectTemplate#2252: the Qodo badge gate forced three hand merges.

Lessons:

Never call a pull request ready while any check is pending, CodeRabbit included. Its status can register after every other check has finished.
On a promotion, Copilot's "needs a closer look" verdict names no line. Sweep for the claim it describes before declining it, since one such verdict here was a real doc error, fixed in State That a HandBrake Run Drops Embedded Closed Captions #1163.

Navigation Menu

Repository navigation

Session Handoff [process-findings]: Promote and Release 3.23 #1085

Round 1, closed 2026-10-30: Session Handoff [process-findings]: Fix the Process Command Suspect Branches

Round 2, closed 2026-10-30: Session Handoff [process-findings]: Fix the Process Command Suspect Branches

Round 3, closed 2026-10-30: Session Handoff [process-findings]: Fix the Process Command Suspect Branches

Round 4, closed 2026-10-30: Session Handoff [process-findings]: Fix the Process Command Suspect Branches

Round 5, closed 2026-10-30: Session Handoff [process-findings]: Fix the Process Command Suspect Branches

Round 6, closed 2026-10-30: Session Handoff [process-findings]: Drive the 3.23 Promotion and Fix the Process Command Suspect Branches

11 remaining items

Round 13, closed 2026-10-30: Session Handoff [process-findings]: Fix the Process Command Suspect Branches After the 3.23 Promotion

Round 14, closed 2026-10-31: Session Handoff [process-findings]: Fix the Process Command Suspect Branches After the 3.23 Promotion

Round 15, closed 2026-10-31: Session Handoff [process-findings]: Promote and Release 3.23 After the Suspect-Branch Burndown

Round 16, closed 2026-10-31: Session Handoff [process-findings]: Fix the Test Log Capture Copies, Then Promote and Release 3.23

Round 17, closed 2026-10-31: Session Handoff [process-findings]: Promote and Release 3.23, and Fix the Reduced-Run Lookup in find_run

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Session Handoff [process-findings]: Promote and Release 3.23 #1085

Description

Activity

ptr727 commented on Oct 30, 2026

ptr727 commented on Oct 30, 2026

Round 1, closed 2026-10-30: Session Handoff [process-findings]: Fix the Process Command Suspect Branches

ptr727 commented on Oct 30, 2026

ptr727 commented on Oct 30, 2026

Round 2, closed 2026-10-30: Session Handoff [process-findings]: Fix the Process Command Suspect Branches

ptr727 commented on Oct 30, 2026

ptr727 commented on Oct 30, 2026

Round 3, closed 2026-10-30: Session Handoff [process-findings]: Fix the Process Command Suspect Branches

ptr727 commented on Oct 30, 2026

ptr727 commented on Oct 30, 2026

Round 4, closed 2026-10-30: Session Handoff [process-findings]: Fix the Process Command Suspect Branches

ptr727 commented on Oct 30, 2026

ptr727 commented on Oct 30, 2026

Round 5, closed 2026-10-30: Session Handoff [process-findings]: Fix the Process Command Suspect Branches

ptr727 commented on Oct 30, 2026

ptr727 commented on Oct 30, 2026

Round 6, closed 2026-10-30: Session Handoff [process-findings]: Drive the 3.23 Promotion and Fix the Process Command Suspect Branches

11 remaining items

ptr727 commented on Oct 30, 2026

ptr727 commented on Oct 31, 2026

Round 13, closed 2026-10-30: Session Handoff [process-findings]: Fix the Process Command Suspect Branches After the 3.23 Promotion

ptr727 commented on Oct 31, 2026

ptr727 commented on Oct 31, 2026

Round 14, closed 2026-10-31: Session Handoff [process-findings]: Fix the Process Command Suspect Branches After the 3.23 Promotion

ptr727 commented on Oct 31, 2026

Round 15, closed 2026-10-31: Session Handoff [process-findings]: Promote and Release 3.23 After the Suspect-Branch Burndown

ptr727 commented on Oct 31, 2026

ptr727 commented on Oct 31, 2026

Round 16, closed 2026-10-31: Session Handoff [process-findings]: Fix the Test Log Capture Copies, Then Promote and Release 3.23

ptr727 commented on Oct 31, 2026

ptr727 commented on Oct 31, 2026

Round 17, closed 2026-10-31: Session Handoff [process-findings]: Promote and Release 3.23, and Fix the Reduced-Run Lookup in find_run

ptr727 commented on Nov 2, 2026

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions