Absence-claim rule names the duty but not the mechanism: a full-listing + client-side prefix filter returns a silent zero on a wrong path where the scoped probe 404s

Goal

CLAUDE.md § "Ad-hoc results summaries" requires that "an absence / not-persisted claim shipped in chat or ANY durable marker is grounded by a repo-wide grep PLUS a SCOPED listing of the HF prefix's siblings/parents (list_repo_tree(path_in_repo=<parent prefix>)) PLUS the WORKTREES". The duty is correct and the right call is even named. What the rule does NOT say is WHY the scoped call is the one named — and without that, the natural implementation (fetch the full listing once, filter locally) is read as satisfying it while being structurally incapable of the check.

Add the mechanism to the rule text: an absence check must use a call that can FAIL on a wrong location.

The mechanism, reproduced live (task #2329, 2026-08-21)

Same repo, same nonexistent prefix issue2329_q35_ladder_decay/, two calls:

# (1) full listing + client-side filter  ->  rc=0, silent zero
files = api.list_repo_files(repo, repo_type="dataset")
mine  = [f for f in files if f.startswith(pref)]      # -> []
# prints: total files under issue2329_q35_ladder_decay/: 0   *.done.json: 0

# (2) scoped server-side probe            ->  raises
api.list_repo_tree(repo, repo_type="dataset", path_in_repo=pref, recursive=True)
# -> EntryNotFoundError: 404 Client Error

Form (1) cannot distinguish "this prefix holds nothing" from "this prefix does not exist", because the prefix never reaches the server: it is a local string filter over a listing produced for a different question. It exits 0 and prints an authoritative-looking zero.

Real consequence in #2329: the check was verifying that 72 block *.done.json records reached HF, because a reconciler's discharge of an open BLOCKER (regen-cap-not-enforced) rests on those records travelling with the staged artifact set. The true prefix is issue2329_q35rerun/analysis_tensors/ladder/manifests/blocks/ with 72/72 present. Had form (1) returned first, the round would have concluded the discharge's precondition had failed — from a probe that never asked the server anything. It was caught only because a scoped probe was run afterwards for an unrelated reason and 404'd.

Contributing layout trap, worth capturing in the same edit: this issue's store splits across TWO top-level HF prefixes by artifact class, neither named for the round label — issue2329_q35rerun/analysis_tensors/ladder/ (anchors, gates, manifests, margin, va_store, vc_bank) and issue2329_q35rerun/raw_completions/ladder/ (anchors, grid). Rollout TEXT is under raw_completions/ per upload policy, so "the grid store" is NOT under the tree holding every other ladder artifact. A prefix guess that looks obviously right is often wrong here.

Proposed change

In CLAUDE.md § "Ad-hoc results summaries state per-arm provenance…" (the absence-claim sentence) and/or .claude/rules/upload-policy.md, add the mechanism as a short clause:

• An absence check MUST use a call whose error surface distinguishes wrong-path from empty-path — list_repo_tree(path_in_repo=...) or get_paths_info (both 404 on a wrong prefix). A full list_repo_files listing plus a client-side startswith/glob filter is acceptable ONLY for counting WITHIN a prefix already proven to exist.
• Where neither is available, require a positive control: assert the PARENT prefix returns non-zero before reading a child's zero as meaningful. (One extra line; it is how the #2329 case was actually resolved — listing ladder/'s children showed no grid, which located the real path.)

Keep it to the rule text plus, if cheap, a one-line helper. No behavioural change to any existing verifier is required — verify_uploads.py already uses scoped calls.

Acceptance criteria

1. The absence-claim rule text names the wrong-path-vs-empty-path distinction and the two acceptable call shapes, not just the duty.
2. The list_repo_files + client-side-filter form is explicitly scoped to within-a-proven-prefix counting.
3. The parent-prefix positive control is stated as the fallback when no scoped call exists.
4. If a helper is added, it raises (never returns 0) on a nonexistent prefix, with a test using a nonexistent-prefix fixture asserting the raise — the test must FAIL against a client-side-filter implementation.
5. No new red in the no-flags workflow_lint.py run or the mapped-test selection.

Candidate metadata

• target_file: CLAUDE.md (§ absence claims) + .claude/rules/upload-policy.md
• fingerprint: absence-check-requires-failing-call-not-client-side-filter
• confidence: high — reproduced live in #2329 with both call forms against the same wrong prefix, one returning rc=0 zero and one raising 404

Provenance

workflow_fix_target: CLAUDE.md

Auto-filed by the /issue 2329 orchestrator from a self-caught near-miss during the r20 G5 remedy staging (2026-08-21). Evidence: #2329 events.jsonl epm:progress v185 (prefix correction) and v186 (the two-call comparison). Kin: this is the ABSENCE-side sibling of the eight hollow-gate findings in the same round (a green verdict whose comparison set is empty) — same defect shape, opposite direction.