fix: count log growth and process-tree CPU as worker progress, and report what a kill left behind

.claude/skills/orchestrate/SKILL.md

@@ -118,7 +118,7 @@ node tools/plan-queue.mjs        (--tickets ORB-1,ORB-2 | --board) [--format mar
 node tools/comment-ticket.mjs    --issue "<ticket-ref>" --body-file <path|->
 node tools/complete-ticket.mjs   --issue "<ticket-ref>" [--preflight]
 node tools/compose-prompt.mjs    --issue "<ticket-ref>" --repo <key> --out <file> [--worktree <p>] [--branch <b>] [--base <ref>]
-node tools/launch-worker.mjs     --issue "<ticket-ref>" --worktree <p> --prompt <f>
+node tools/launch-worker.mjs     --issue "<ticket-ref>" --worktree <p> --prompt <f> [--hard-ceiling-minutes <n>]
 node tools/verify-delivery.mjs   --issue "<ticket-ref>" --worktree <p> --branch <b> --repo <key> [--base <ref>] [--wait-ci <s>]
 node tools/list-bot-threads.mjs  --pr <n-or-url> --repo <key> [--wait-seconds <s>]
 node tools/resolve-bot-thread.mjs --thread <PRRT_...> --repo <key> --pr <number>   # reply body on stdin
@@ -501,13 +501,17 @@ Write it to the scratchpad. A prompt file inside the worktree gets committed by
 ## Steps 5 and 6. Spawn the worker
 
 ```bash
-node tools/launch-worker.mjs --issue "<ticket-ref>" --worktree <p> --prompt <f>
+node tools/launch-worker.mjs --issue "<ticket-ref>" --worktree <p> --prompt <f> [--hard-ceiling-minutes <n>]
 ```
 
 Headless, `stdin=NUL`, `cwd` = the worktree, log to the scratchpad.
 
-`launch-worker.mjs` runs in the FOREGROUND as its own watchdog and owns both clocks: a hard 45
-minute cap and a 10 minute no-progress cap, killing the whole process tree on either.
+`launch-worker.mjs` runs in the FOREGROUND as its own watchdog and owns both clocks: the hard
+ceiling and the no-progress cap from `.claude/orchestrator.json` (currently 45 and 10 minutes),
+killing the whole process tree on either. Pass `--hard-ceiling-minutes` at launch for a ticket the
+plan already shows outrunning the fleet-wide ceiling: a large migration, a subsystem ticket, or one
+whose test matrix is the work. Three finished workers died at the fixed 45 on 2026-08-22 for
+exactly that shape.
 
 The orchestrator launches it as a **background shell task and ends its turn.** Zero tokens burn
 while the worker runs, and the process exiting is what wakes the session.

.github/workflows/visual.yml

@@ -70,14 +70,22 @@ jobs:
         if: steps.mode.outputs.seed == 'true'
         run: |
           set -euo pipefail
-          git lfs install --local
+          # --skip-repo: lefthook owns .git/hooks, and a plain install exits 2 refusing to
+          # overwrite them. This step only needs the LFS clean filter for the PNG add.
+          git lfs install --local --skip-repo
           git config user.name "github-actions[bot]"
           git config user.email "41898282+github-actions[bot]@users.noreply.github.com"
           git add apps/web/e2e/visual/__screenshots__
           if git diff --cached --quiet; then
             echo "Baselines already up to date — nothing to commit."
           else
-            git commit -m "test(web): update hermetic visual baselines (#461 B1) [skip ci]"
+            # This push uses the workflow's GITHUB_TOKEN, and GitHub never starts workflow runs
+            # for events that token causes, so the bot head gets NO required checks either way.
+            # [skip ci] stays out because it overstated intent while adding nothing the token
+            # suppression does not already do. main requires 21 checks per head, so after a
+            # labeled re-seed the operator's next real push (or a checks re-run) is what puts the
+            # required contexts on the head. The staged-diff guard above is the seed-loop stop.
+            git commit -m "test(web): update hermetic visual baselines (#461 B1)"
             git push origin "HEAD:${{ github.head_ref || github.ref_name }}"
           fi
 

tools/README.md

@@ -23,7 +23,7 @@ name or git status. Read `CONVENTIONS.md` before adding one.
 | `complete-ticket.mjs` | Owns the explicit post-merge transition. `--preflight` proves the open ticket and configured project item without writing. Completion sets Status Done and closes the issue as completed. `--repair-status` is the inverse, for a ticket GitHub already closed from a merge commit: it reconciles the board Status with the close reason (completed to Done, not planned to Canceled, duplicate to Duplicate), writes nothing else, and is a no-op when the column already agrees. `--cancel --reason-file` closes an open ticket whose work is GONE rather than done: it posts the required reason, sets Status Canceled and closes as not planned, so deleted work is never recorded as shipped. | `node tools/complete-ticket.mjs --issue <ORB-N\|#N\|N> [--preflight \| --repair-status \| --cancel --reason-file <path\|->]` |
 | `plan-queue.mjs` | Resolves a scope (explicit tickets, a GitHub milestone, or the configured board) into ONE ordered execution plan: admission and deferrals, DAG-safe waves, transitive fan-out (`unlocks`), and stackable dependency chains. A ticket whose same-repo blockers do not form one chain has no branch that can carry them, so it defers as `UNSTACKABLE_BLOCKERS_IN_QUEUE` and the rest of the board still plans; cross-repo blockers order but never stack. A still-open blocker outside the queue defers the ticket. It plans only, and writes nothing anywhere. | `node tools/plan-queue.mjs (--tickets ORB-1,#123 \| --project <name> \| --board)` (`--format markdown`) |
 | `compose-prompt.mjs` | Builds one worker prompt: the GitHub ticket body verbatim, the orchestrator's brief (objective, scope, output, boundaries), and the finishing contract. Refuses to write inside any declared repo, because a worker that finds its own work order in the tree can commit it. | `node tools/compose-prompt.mjs --issue <ORB-N\|#N\|N> --repo <ui\|api\|landing> --out <abs path>` |
-| `launch-worker.mjs` | Spawns one headless worker and supervises it in the foreground. Both clocks kill the complete process tree, and this process is itself the run's wake source. `--measurement` swaps the no-progress clock for the measurement one and leaves the hard ceiling alone. `--dry-run` prints the resolved plan and spawns nothing. GitHub credentials are owner-selected per child environment; global account state is never switched. It launches a worker; it never merges, reviews, or moves a ticket. | `node tools/launch-worker.mjs --issue <ORB-N\|#N\|N> --worktree <path> --prompt <file>` (`--measurement`, `--dry-run`) |
+| `launch-worker.mjs` | Spawns one headless worker and supervises it in the foreground. Both clocks kill the complete process tree, and this process is itself the run's wake source. `--measurement` swaps the no-progress clock for the measurement one and leaves the hard ceiling alone. `--hard-ceiling-minutes <n>` replaces the configured ceiling for this one launch, for a ticket that legitimately outruns the fleet-wide default. `--dry-run` prints the resolved plan and spawns nothing. GitHub credentials are owner-selected per child environment; global account state is never switched. It launches a worker; it never merges, reviews, or moves a ticket. | `node tools/launch-worker.mjs --issue <ORB-N\|#N\|N> --worktree <path> --prompt <file>` (`--measurement`, `--hard-ceiling-minutes <n>`, `--dry-run`) |
 | `verify-delivery.mjs` | Derives delivery solely from bounded Git and GitHub children: clean/pushed/current PR, exact current-head CI state, and GitHub compare `behind_by=0`. `OUT_OF_DATE` reports base/head SHAs and behind count. File and line counts remain `sizeAdvisory` and never block. Completed CI passes only on the confirmed `SUCCESS`, `NEUTRAL`, or `SKIPPED` allowlist; every other or future conclusion fails closed. Failed CI retains run/check ID, details URL, workflow, name, status and conclusion; newest rerun wins. | `node tools/verify-delivery.mjs --issue <ORB-N\|#N\|N> --worktree <p> --branch <b> --repo <key>` (`--base`, `--wait-ci`, `--command-timeout-seconds`) |
 | `record-readiness.mjs` | Reads the live PR base/head/draft, its branch's required status checks, the compare, and the ticket ONCE, then persists and evaluates one repository-qualified receipt. READY requires green CI over the required contexts, a non-draft state, live `behind_by=0`, and ticket synchronization simultaneously; evidence pinned to an older head or base becomes an explicit stale verdict, and every verdict is reported together. The code review is NOT a separate axis: Pullfrog publishes `pullfrog-approval` as a required status check, so the verdict already arrives through the required contexts this tool reads. It never trusts a caller-authored status flag. | `node tools/record-readiness.mjs --repo <ui\|api\|landing> --pr <number> --delivery <file> --ticket <file>` |
 | `sync-issue-state.mjs` | Asserts the live ticket carries exactly the `repo:<key>` label matching `--repo` before either write, so a mistyped ticket key cannot move or comment on a stranger's ticket. Moves working and blocked states to In Progress and simultaneous technical readiness to In Review, never Done. Stores the last posted state to suppress duplicate comments and emits a receipt for the readiness state machine. | `node tools/sync-issue-state.mjs --issue <ORB-N\|#N\|N> --repo <key> --pr <n> --state <working\|blocked\|ready> --head-sha <sha> --base-sha <sha> --message-file <path\|->` |