daily-held: deferred sweep backlog 2026-07-25

Why this needs you

Filed by the /daily 2026-07-25 problem sweep as a route-3 tracked item so nothing from an unusually large sweep is silently dropped.

The sweep mined 28 in-scope transcripts across 21 autonomous /issue sessions and 7 interactive ones and surfaced ~117 distinct problems — roughly 4× a normal night (the #1667–#1689 wave merged 19 tasks in one day, and #1689 alone burned 8 implementer/code-review rounds). 14 route-2 workflow-fix tasks and 4 route-3 judgment-call tasks were filed. This task holds the remainder: real, fix-derivable items deliberately NOT filed tonight, so the ~5-session infra concurrency cap is not flooded by a single night's sweep.

Your decision: drain this list at your own pace (task.py new per item, or point a future /daily at it), or tell me to file them in a later batch. Nothing here is urgent; several are cosmetic. They are recorded so they do not evaporate.

Deferred items, grouped by target

CLAUDE.md — chat/writeup register

• Low-cosine similarity twin. The #1310 "similarity-statistic semantics" rule governs direction-aware vs spectrum-invariant reads; it has no clause for the negative direction. A near-zero operator cosine (raw −0.013, aligned 0.068) was narrated as "essentially orthogonal / degenerate / uninformative", then refuted ~2 h later by the same session's own ladder run (a single context mean shift took direct transfer from −1982 to −0.06; context reparameterization alone reached 93–95% of ceiling). Proposed: a low cosine may not be narrated as "unrelated" without first checking whether an affine correction recovers transfer. (dffde9b6 @ 22:10:29Z → 23:24:09Z / 00:22:53Z.)
• Gloss on FIRST use, not on challenge. "Guarded rerun" was used across two full writeups before you asked what it meant. Existing item (3) says every compressed claim gets a gloss; sharpen it to bind on first use. (63122023 @ 18:11:19Z — your message: "what do you mean by guarded rerun".)
• Opaque condition codes reach chat. M_instruct_user_chat, S_instruct_chat, Track-M, r1~r2, b2i/i2b appeared in 4 assistant text blocks (11 occurrences) in 63122023; the no-opaque-codes rule is currently scoped to plans and clean-result bodies. Extend to chat writeups and interim tables.
• Plots-over-tables when a figure already exists. A 7-row cross-condition markdown table led the final #825 summary with an existing figure merely linked below.
• Detach-lifecycle recurrence. You asked twice in one session — "will this continue running if i shut down the terminal (in tmux?)" (18:36:57Z) and "so i can close this terminal now?" (19:08:13Z) — after Phase 0 dispatched at 18:17 and #1689 spawned at 18:55 with no lifecycle statement. This is a recurrence of the SOUL.md detach-transparency rule written 2026-07-24. Proposed: make the line a mandatory field of the inline-round dispatch ack and the epm:progress dispatch note.
• Structural/ordering claims need the same compose-time check as numeric ones. Two bullets committed in docs/mapping_similarity_metrics.md were wrong (a "predicted through W_s" claim, and a nesting claim ordering two model families as strong/weak when they are siblings). Committed 13f00f85a1 07-24 19:57 PT, corrected 67f8e798db 07-25 10:40 PT — ~15 h live. (51e59f07 @ 17:38:43Z.)
• Methodology-doc gloss debt. Four consecutive clarification re-asks from you on the same two sections of that note in a 14-minute span; the assistant's own verdict was "those two sections were the muddiest part of the note". Extend the plain-English gloss duty from chat writeups to committed docs/*.md methodology notes.

CLAUDE.md § Orchestrator vs subagent / .claude/rules/gotchas.md — process hygiene

• Self-matching pkill/pgrep. A compound pkill -f "issue1345_ladder_rung[s]"; …; nohup <relaunch> killed its own shell before the relaunch ran (exit 144, nothing restarted); separately a watcher loop's pgrep matched its own command line and would have spun forever (reaped with SIGKILL). Extend the bracket-a-character rule: (a) kill and relaunch never share one Bash invocation, (b) a watcher's own command line must not contain the literal it greps for.
• Backgrounded sleep is not elapsed time. Several "wait for X" polls were dispatched run_in_background=true and read immediately, so ~2 min of believed progress never happened (etimes=122 disproved it). Job age must be read from ps -o etimes= / /proc/<pid>/stat.
• Stale monitors outlive their jobs. 6 separate triage turns across two sessions spent dismissing monitors armed on killed runs; two "wait for X to finish" loops survived 3.5 h past their work and were killed by the harness. Proposed: killing a monitored process cancels its Monitor in the same step, and a session-wrap duty reaps own background wait-loops.
• Foreground-vs-background sleep boundary. A foreground sleep 60 && cat was hard-blocked while background sleep 540 && backend_poll.py chains ran fine 5× in the same session — state the distinction rather than rediscovering it by tripping it.
• py-spy dump needs sudo on this VM (Permission Denied … elevated permissions); document the sudo env "PATH=$PATH" form and the /proc/<pid>/stat fallback in the hang-diagnosis rule.
• spec_from_file_location + @dataclass raises AttributeError unless the module is registered in sys.modules before exec_module — hit while probing a scripts/ module's internals.

.claude/rules/code-style.md / .claude/skills/paper-plots/SKILL.md — new-script authoring

• Heavy-import-before-load_dotenv() shipped into 4 newly authored scripts across two sessions, caught only by the commit-time gate (#847 assertion). Proposed: an explicit new-VM-entrypoint import-order template.
• Bare hf_hub_download bypassing hub.retry_transient (the #1547 shared-commit-budget rule) in a new script; the rule currently reads as upload-centric.
• paper_plots entrypoint guessed wrong (apply_paper_style → the real set_paper_style), then the auto-formatter dropped the corrected import → NameError. Put the exact import boilerplate at the top of the skill.
• Symlog default: an all-arms figure rendered unreadable (prefix panels flattened by a −17,500 outlier); caught only by the mandatory PNG read. Propose a symlog default when a series spans >~3 orders of magnitude.
• Edit-after-formatter: two String to replace not found failures from constructing old_string against pre-formatter state. Note that a Write to a Python file may be reflowed, so an Edit needs a fresh read of the region.
• Heredoc over python -c for multi-statement one-liners with nested quotes (one unterminated-quote failure).

.claude/skills/issue/SKILL.md — pipeline papercuts

• Canonical inline lint gate. Three sessions ran the gate's components by hand, were blocked by guard_root_code_commit.sh for an uncertified payload, then had to re-run scripts/inline_lint_gate.py (~4–10 min each). CLAUDE.md names the checks but never the invocation (grep -c 'inline_lint_gate' CLAUDE.md → 0; it appears only in SKILL.md at ~line 7041). Also show the --map-files argument shape (git diff --cached --name-only > /tmp/files.txt), which three calls got wrong.
• Bare /issue-tick with no <N> loaded 21,752 chars of skill body into a live interactive session and was interrupted 2.4 s later; the #1629 human-active screen lives inside tick_triage.py, which never runs. Propose a first-line short-circuit.
• Step 0 ToolSearch preload should include TaskOutput,Monitor (one InputValidationError from an unloaded deferred schema).
• Draft-PR recipe should ship the guard-compliant redirect form (gh pr create … > /tmp/…log 2>&1; PR_RC=$?) — the piped form is blocked.
• Per-lane on-pod repo root (/workspace/eps-issue-<N> on GCP vs /workspace/explore-persona-space on RunPod) and the working gcloud incantation for the eps/phase metadata key (the / breaks a naive projection) — both cost a turn after a failover.
• Staggered critic verdicts cost 2–3 orchestrator turns each, including 5 bare true no-op Bash calls purely to end a turn, plus one redundant ScheduleWakeup backstop that had to be swept at teardown.
• Mergeability double-read: #1680 burned an attempt on a stale post-push mergeable read (3 attempts total). Propose two consecutive agreeing non-UNKNOWN reads ≥5 s apart.

Other named targets

• scripts/task.py / scripts/spawn_session.py hygiene — new --body-file silently drops the body's goal: frontmatter (documented, but silent); set-goal --by defaults to user, so an agent-set goal was recorded as yours; the crash-recovery registry wrote session_id: null / auto: null despite an --auto spawn; spawn_session.py list does not mark the calling session's own row (5 probes to disambiguate); pod.py sync results accepts none of --pull/--pod/--paths and errors under the sub-tool's name.
• .claude/rules/workflow-fix-on-bug.md grep recipe prescribes grep -rln '<pattern>' .claude/ CLAUDE.md scripts/ at 3 places; .claude/worktrees/ now holds 29 checkouts, so the command auto-backgrounded at 2 min and had to be re-run scoped. Add --exclude-dir=worktrees --exclude-dir=__pycache__.
• Recursion-guard exit latency distorts scope. #1668's orchestrator explicitly reasoned that parking a candidate costs ~a day and therefore expanded the current round's scope instead. Proposed: extend the #1681 urgent-park grammar beyond urgency: main-red to a gate-blocking class (a park whose bug demonstrably red-verdicted this session's own gate).
• .claude/agents/planner.md — §10 should record the actual git log origin/main..origin/issue-<M> output per parent commit rather than a bare "merged on main" (a consistency-checker BLOCK on #1689 proved 11 such assertions false-in-evidence though right-in-conclusion); and a plan assumption naming a pytest node's PASS/FAIL should be marked unverified unless actually executed (a #1679 fact-checker found a second test breaks and the count was 18, not 17).
• .claude/agents/implementer.md:174 cites tests/test_workflow_lint.py at 319–771 s; a #1679 verification stage ran ~24 min against that expectation. Re-measure and date the figure.
• scripts/step9c_baseline.py — trigger the ledger refresh on merge-count since refresh as well as wall-clock age (six concurrent merges aged it mid-afternoon).
• .claude/agents/planner.md §9 / GPU width — #1689 held a 4×H100 pod at gpu_util 0,0,0,0 in 7 of 7 numeric poll samples through CPU (corpus, render) and pure-API (haiku_u2) phases, repeated across ~4 relaunch cycles. The per-phase GPU-width right-sizing rule exists; it was not applied at plan time.
• docs/api_throughput_guidelines.md — a Haiku 100 k row was quoted twice in one plan as a Sonnet figure (3–5 min vs the correct 13–17 min), near the ~200 k sync-vs-batch crossover. Re-key each latency band unambiguously on (model family, call count).
• .claude/rules/plan-compute-sizing.md — a #1689 bootstrap battery projected 140 GPU-h against a 4 h plan estimate (35×) before batching brought it to 2.5×; the measured 1-cell pilot basis the rule already mandates was evidently not used.
• .claude/skills/adversarial-planner/SKILL.md — plan_patch.py --verify-contains needs "pick a fragment with no backticks and no line break" (one failed patch chain).
• GCP boot-loop diagnostics. #1689's GCP instance boot-looped twice and the #1029 auto-failover absorbed it with no serial-console capture and no root cause; the task was then hard-pinned to RunPod, abandoning the free-credits lane. Propose persisting the dying instance's serial tail + last eps/phase to issue<N>_partial/boot_death/ before the failover fires.
• Pre-existing baseline debt. Gates are baseline-subtracted: 35 ruff errors and 15 lint WARNs on main are permanently invisible to the pipeline. A one-off cleanup + a ratchet would retire them.
• RunPod account-key preflight WARN — the VM's key is absent from the shared team account key list (mutated by fellows-cluster onboarding). Latent: fresh pods are fine via PUBLIC_KEY boot injection, but a pod resume could fail. Touches a shared account, so it is yours either way.

Not filed, and why — items the sweep deliberately dismissed

Recorded so a later pass does not re-raise them:

• "Code-reviewer returned PASS with 7 open BLOCKERs" — REFUTED at compose time. .claude/agents/code-reviewer.md rule 11 prescribes raising a BLOCKER "even on a PASS verdict" because "the Step 5c-ter dispatch gate reads concerns.jsonl, not verdict prose". The orchestrator's bounce was the designed mechanism working.
• git status --porcelain --cached — an improvised flag, not present anywhere in the workflow surface (grep → 0 hits). Nothing to fix.
• The piped-git guard blocking the session that was fixing it (#1675) — self-resolving; the merged fix removes the class.
• Read overflow on scripts/sync_repo_root.py (28,240 tokens vs a 25,000 cap) — one extra tool call to recover; not worth a rule.
• #1689's ~3.5 h spawn-to-progress delay from an API capacity wall — the crash-recovery watcher behaved as designed. The mis-classification half (a 529 labelled boot-refusal) IS filed as a route-2 task tonight.
• ~3 h and 2.3–3.5 M subagent tokens per single-file workflow fix — deliberate review-depth policy, not a defect. The tractable sub-question (whether the Step 9c invariant remainder could run once per wave rather than once per session) is a cost/coverage tradeoff for you, not an autonomous call.