v2.57.3 — Rule 22 denies again when an edit's own line is not yet written

The Rule 22 gate denies again when an edit is the first tool call of its response. Measured on Claude Code 2.1.308: such an edit has its own transcript line written only after its own PreToolUse hook returns (a transcript watcher saw it land 0.12 s after the hook exited). pre-edit-check.sh waited up to 40 s for that line, could never find it, and failed open — every such edit without a recorded heredoc carrier was allowed after a 40 s stall, with "could not verify this edit". Census of the last 250 project transcripts (events dated 2026-10-10 or later): 143 of those fail-opens against 130 real denials, and every fail-open since 2026-10-02 ran the full 40 s.

• Tail mode: the hook waits 5 s (ARIA_R22_FLUSH_WAIT_MS, was 40 s) for its own line — measured landing times are 1.18 s (a Bash-less subagent's first-tool Write) and ~1.6 s (a later tool in a response), and a too-short wait would falsely deny a compliant edit — then judges the transcript already on disk — the same window walk as before, started at end-of-file. A readable transcript now always yields allow or deny.
• Previous-edit anchor: each check records its tool_use_id in ${TMPDIR}/aria-r22-lastedit-<session><agent>. In tail mode, a previous edit that has not landed yet returns the loud "could not verify" rather than a verdict whose window could reach that edit's marker (one marker, one edit).
• Expected effect, measured by replay: run over the 144 recorded fail-opens with each edit's own line removed, the new hook denies 143 and allows 1. Of the 143: 108 had no marker at all, 23 only post-edit [Rule 22 · Scope] lines, and 12 wrote the marker mid-line (echo "[Rule 22] …"), a form the gate has never accepted — the heredoc carrier (cat <<'R22') is the reliable path. Bound: carriers and anchors were not reconstructed, and the replay sees slightly more of the transcript than the live hook did.
• Still a loud fail-open: a missing or unreadable transcript, or a python failure.
• Docs: the change-decision framework (all five ports), the always-on rules digest and the SessionStart RULE 22 ORDERING message no longer promise a 40 s wait for visible text.
• Tests: r22-tool-input-marker-and-flush.sh H2a/H2b/H4/H5/H6, 4-8-thinking-and-failopen.sh E/E2, r22-circuit-breaker.sh test 3, r22-carrier-side-channel.sh W1/R3, and the new r22-tail-anchor.sh (T1–T6). The final suites run against the 2.57.2 hook fail exactly the eleven predicted controls.
• Ports: Antigravity (2.57.3) regenerated; its hook is byte-identical to canonical. Codex, Cursor and Cowork have no transcript wait and are unchanged.