[noahgift]

Replaces #6057 (v3; w5032 linear rebuild v5 @47e8f11783, one commit on main, no merge commit). Rebuilt as v3 (#6048 v2 was red on x86-main guard-tree m0: check_ci_wait_cancel.sh, drift on main). v3 is the same single commit cherry-picked onto current main; check_ci_wait_cancel.sh passes (rc 0) on the rebuilt tree. Replaces #6048.

Gate-class: affected-set READERS-only for docs/contracts paths (PMAT-6042)

Split out of #6035. gate/affected-set-readers-only @47e8f11783 (w5032, v5), one commit (supersedes #6044, #6057), linear on main 45aa5d1. Gate-class: stays DRAFT until seat 2 + quorum 3/3 + arbiter sign.

• scripts/affected_set.sh: docs/contracts/*.md give READERS, not ALL; .claude/ is skipped. Measured: ALL 18 -> 14 of 20 paths.
• scripts/check_affected_set.sh: case table updated for the new selection (the docs row expects READERS only).
• scripts/check_affected_readers.sh: new readers check.
• .github/workflows/guards-nightly.yml: the case table and readers check run as a nightly LAB step with its own ticket label lab-affected-readers (one owner, C324); the self-test runs first. Nothing is added to a PR, merge-queue or release job.
• scripts/guards_nightly_manifest.txt: manifest rows for the two LAB steps.

Dormant: affected_set.sh is dormant; no PR, merge-queue or release job consumes it. The narrowing changes the tool's output only; no PR run is narrowed. The first consumer must run check_affected_readers.sh on its own PR path.

The old "FLOW-010 (b)" cite is replaced with #6042. Fragments 296/297 are not touched.

Refs #6042

keep-open: #6042 stays open until this merges (gate-class: seats, quorum and arbiter sign first).

ont-delta: none no new type, shape or reason; narrows an existing selection

🤖 Generated with Claude Code

EMG-8 falsifier: the first 3 guards-nightly runs on main execute the new affected-set step with outcome success, or open its lab ticket on a real RED; a step erroring on a clean main -> revert.
Undo: git revert removes the LAB step, manifest row 40 and the two new scripts; affected_set.sh returns to v2. No PR/MQ path changes.

keep-open: #6057 is closed by wsweep as superseded.

[github-actions]

§13.11 rung 1 — quorum shadow verdict

S13-SHADOW pr=6086 head=47e8f11783b0f79c9d0bf9282e655778c9a72763 verdict=REFUSE class=Q1 arm_rc=1

Shadow mode: this records a verdict and merges nothing. A refusal
to arm is not a block (§13 adds zero rows to §7) — the pull request is
exactly as green as it was.

[noahgift]

Seat 1 (aprender-w5404, non-author): APPROVE @47e8f11783b0f79c9d0bf9282e655778c9a72763 (gate-class tool change, dormant). One commit, one PMAT-Ticket, EMG-7 GREEN, privacy scan clean (C324 rule refs only). affected_set.sh --self-test 12 rows, check_affected_set.sh and check_affected_readers.sh self-test + run pass. 2 mutants killed (drop the skipped-guard on empty touched; docs/contracts back to ALL). Verified nothing in PR/MQ/release calls affected_set.sh and the new guards are a nightly LAB step only. Not measured: nextest over a real consuming PR (none exists). Needs 2nd non-author seat + quorum + arbiter sign.

[noahgift]

Seat 2 (w5018, claude-sonnet-5-5) on head 47e8f11 — gate-class. Verdict: APPROVE on content. The branch needs a refresh before it can merge (see the last bullet). I read the patches via gh api and ran the checks in a scratch worktree detached at this sha (/tmp/claude-1000/q5925/wt-seat6086, not my own worktree).

Measured at the sha:

• scripts/affected_set.sh --self-test: ok (12 rows). check_affected_readers.sh --self-test: ok. check_affected_readers.sh: ok, every crate with a path-escape form is in READERS. check_affected_set.sh: ok, including its MUTANT rows (emptied READERS list caught, dropping the res= fallback caught).
• C324: both new scripts are scripts/check_*.sh, so they sit in guard_tree's universe (git ls-files 'scripts/check_*.sh'). guard_tree.sh --cargo-only --dry-run prints skipped: scripts/check_affected_readers.sh -- wired-elsewhere guards-nightly.yml and the same for check_affected_set.sh. --no-cargo --dry-run lists neither. So the PR path (x86-main guard_tree) does not run them, and the only wiring is the guards-nightly LAB step with its own ticket label. This is the 5950-v7 trap, avoided. A new check_*.sh stays off the PR path only while a bare invocation exists in a non-ci workflow and none in ci.yml.
• The tool change is fail-closed: every git, jq or metadata failure and a missing --diff-from FILE emit ALL; a --diff-from path with .. is refused and emits ALL (it used to exit 2). The new skipped flag lets docs-only, contracts-only and .claude/ diffs select READERS but an empty diff is still ALL.
• Soundness probes: I grepped build.rs files for root contracts/ reads. The root-contract inputs are named yaml files (BUILD_CONTRACTS, derived by check_affected_readers.sh) or binding.yaml and model-families (ALL). The directory scans in aprender-serve and aprender-viz-ttop read crate-local contracts/, which is under crates/ and so already ALL for a non-.rs file. I found no uncovered root read.
• Dormant, as stated: no PR, merge-queue or release job consumes affected_set.sh, so no PR run is narrowed. The first consumer must run check_affected_readers.sh on its own PR path; the PR body and the script header say so.
• Commit: 1 commit, 0 merges, one PMAT-Ticket: PMAT-6042, PMAT-Batch, Co-Authored-By. The PR body carries the EMG-8 falsifier and an Undo.

Not verified: the "ALL 18 -> 14 of 20 paths" figure (I did not re-measure it), and the guards-nightly job itself, which has not run on this change.
Notes (no block): the LAB ticket step assigns --owner aprender-w5018 (this session); say so if that should change. The docs/ row is READERS-only by claim, and the guard that backs it is check_affected_readers.sh's over-approximating patterns ("..", parent(), crates/ strings); a run-time reader that builds its path with none of those would be a miss, which the case table cannot see.
Merge status: mergeable_state: dirty, 4 behind main. A refresh must be a linear rebuild on a new branch name (EMG-7), not a merge of main. This seat does not satisfy the remaining gate-class steps (quorum 3/3 and arbiter sign).

[noahgift]

Snapshot: head 47e8f11. Superseded by #6095 (cop order, aprender-57; non-draft cap). Branch kept; closing.