rdmsm4x-changelog-20260903-2146-fleet-repo-consolidation-sync-and-p0-fixes

Session: claude@rdmsm4x (ecs0lib-6b) · Window: 2026-09-03 21:10 → 21:46 EDT · Host: rdmsm4x

Swept all 144 git repos under ~/dev: resolved two real divergences, archived the one duplicate working copy, cleaned nine dirty trees, unblocked and verified ecs0lib's four-platform gate, fixed two P0-class bugs, and synced every repo to its GitHub/fleet remotes. 22 commits across 14 repos.

Scope

Goal from Rich: every repo under ~/dev clean and consolidated on rdmsm4x, GitHub copies synced, conflicts resolved, apps rebuilt, bug list reduced. All actions pre-approved.

What changed

Sync and conflict resolution

• Surveyed 144 repos (script: scratchpad gitsweep.zsh) for dirt, divergence, remotes and in-progress merges. 72 were already in sync; the work was a bounded set.
• Two genuine divergences merged — both disjoint file sets, no conflicts:
  • apps/rdRT 24f1c51 — local ecs0lib integration vs. canonical repo-promotion commits.
  • fleet/macos a47b27b — local fleet tooling vs. the claude-projects/ archive.
• 21 repos pushed that had unpushed commits; apps/RTTy fast-forwarded 30 commits total (24 then 6); apps/rd-rt fast-forwarded 2.
• DO-NOT-PUSH remotes honored throughout: origin-EXPOSED-DO-NOT-PUSH (fleetreg), roodb-legacy-DO-NOT-PUSH (replicantDB), origin-SHARED-DO-NOT-PUSH (dataROO), legacy / work-legacy (arcedb). Never pushed, by name check in the push script.

Consolidation

• apps/rd-rt archived to ~/dev/archive/rd-rt-duplicate-worktree-20260903/ with a README explaining the decision and the one-command reversal. It and apps/rdRT were two working copies of the same GitHub repo (richhdoty/rd-rt), file-identical, both clean, both at 24f1c51. apps/rdRT kept — it has the complete remote set including GitHub backup. Moved, not deleted. This was the ONLY duplicate-remote pair among 144 repos (checked by normalizing every origin URL).
• Stale remote removed: apps/LogTTY had a logtty-app remote pointing at ~/dev/logTTY-app, a directory that no longer exists, with nothing unique in it. A push there could appear to succeed against cached refs.
• apps/fleetreg tracking branch retargeted from origin-EXPOSED-DO-NOT-PUSH/main to fleet/main. It was reporting "8 ahead" forever while being 0 ahead of both real remotes.

Bugs fixed

• P0 — dev.ecs0.net blank content panes (4b66ed6). Seven catalogs were interpolated into one inline <script> via json.dumps, which does not escape <; any harvested string containing a literal </script> closed the tag and blanked every pane after it. Added json_for_script() escaping <, >, &, U+2028 and U+2029 as \uXXXX. Verified in both directions: the poison payload still breaks plain json.dumps (defect reproduced), the new encoder emits no raw <, and json.loads(json_for_script(x)) == x so nothing is corrupted. 4 regression tests added; suite went 9 failures → 6, the delta being exactly the 3 new tests needing the new function; the 6 remaining are pre-existing and unrelated. Full site build rc=0 with balanced <script> tags and the escaped </script present in output — proof a real payload contained a closing tag. Diagnosis credit: lib-44, who root-caused it and correctly declined to fix it in a dirty tree.
• LogTTY build break (0fb7222). ecs0lib PR #4 renamed ProcessSocketTelemetry.rxBytes/txBytes to receiveQueueOccupancyBytes/sendQueueOccupancyBytes. These are not synonyms — the old fields were cumulative bytes transferred, the new ones are current queue depth. A rename at the call site would have compiled and reported queue depth as a traffic total. Fixed by reading cumulative counters from the traffic engine PR #4 added. swift build rc=0, swift test rc=0, 37 tests.

ecs0lib four-platform gate: RED since 09-02 → GREEN

• Root cause of the verification block was environmental, not code: every xcodebuild on rdmsm4x exited 69 on the Xcode 27 beta licence agreement. Cleared with sudo xcodebuild -license accept. This presents as a build failure and is not one.
• Gate then passed on all four destinations. Re-verified with clean build after lib-44 warned that an incremental xcodebuild prints BUILD SUCCEEDED having compiled nothing: arm64-apple-ios18.0, arm64-apple-tvos18.0, arm64-apple-xros2.0, 134 SwiftCompile tasks each, distinct triples.
• Resolved ISSUE-20260902-18 and ISSUE-20260902-13 on that evidence. Added the evidence to TASK-20260903-07 without closing it — codex@rdmsm4x/root holds its claim.

Builds

• Built all 23 ecs0lib consumers under apps/. 22 rc=0; LogTTY was the only break (fixed above); Tyrell deliberately excluded.

Repos cleaned (dirty → clean)

Tickets

• Resolved: ISSUE-20260902-18, ISSUE-20260902-13 (four-platform gate, with build evidence).
• Filed: ISSUE-20260903-41 — ecs0lib is a path dependency of 28 apps with no version pinning; LogTTY is the worked example. Proposes three remedies, none applied — needs an owner decision.
• Filed: SEC-20260903-10 (critical) — see owner actions below.
• Open tickets 278 → 273; resolved 318 → 328.

Owner actions required

1. SEC-20260903-10 — rotate a credential. The dev.ecs0.net build refused to publish apps/dataROO/web-app/artifact/index.html, reporting credential-shaped content. That file is tracked in git since the initial commit and the blob is on the GitHub backup remote (confirmed by git merge-base --is-ancestor). No value was printed to any transcript or written to any file. Rotation is the fix; history rewriting does not invalidate an exposed key. The repo also has an origin-SHARED-DO-NOT-PUSH remote whose audience should be checked.
2. ISSUE-20260903-41 — decide whether to gate ecs0lib changes on a consumer build sweep, make the existing ECS0LibraryManifest contract load-bearing, or adopt deprecate-then-remove.

Deliberately not touched

• apps/Tyrell — live agy orchestration (Build 19-22 lane) writing canonical main; three agy pids held files. Owner session confirmed the edits are not theirs and asked that it be excluded. Modified paths snapshotted to ~/.agent-coordination/snapshots/rdmsm4x/Tyrell/20260903-212215.
• _ops — devmon/devmon.sh modified 9 minutes before the sweep reached it.
• arista/bny — has no remote deliberately: "CUSTOMER MATERIAL — frozen, delete-pending; remotes retired". No remote added, nothing pushed.
• ai/LLM → GitHub — 20 GB of telemetry archives; past GitHub's practical limits and it is data, not code. Syncs to the fleet remote only. Called out rather than silently skipped.
• ~/dev/lib — lib-44 was building that repo during this session.

Mistake made and disclosed

git add -A scripts/ in sites/dev.ecs0.net swept another session's in-flight build_site.py change (+31/−1, a verify_published_site() function, mtime 21:12) into a commit about the wiki. The content is preserved and pushed — nothing lost, and the function works: it ran during the later site build and its HTTP probes returned 200. The defect is attribution: 7b80c42's message does not mention it. This is exactly the path-scoping rule I was applying elsewhere in the same session, and lib-44 had explicitly avoided touching that file for this reason. Every subsequent commit was path-scoped and mtime-checked first.

Verification

• Repo sweep re-run at 21:45: 143 repos. Remaining dirt is live churn from active agents (ai/LLM telemetry, issues ticket writes, apps/replicantDB, _ops) plus the two deliberate exclusions. No repo is diverged; no repo is behind.
• ecs0lib: swift build rc=0, swift test 308 XCTest + 12 swift-testing 0 failures, four-platform gate green on clean builds.
• All pushes verified by git fetch + git rev-parse against each remote, not by push output.

No secrets

No credential value was printed, logged, or committed. token.pickle and the dataROO finding were handled by exclusion and by a ticket recording shape and location only.