docs(planning): record v1.5.2 shipped, and close the deploy-trigger todo as proven - #155

…odo as proven

TWO THINGS CLOSED.

1. v1.5.2 RECORD. Tag v1.5.2 on 7e8a71e, live and verified FROM SVN rather than
the API: trunk Stable tag 1.5.2, tags/1.5.2 at r3642817, and — the check the
v1.4.0 lesson exists for — `emitted_norm` x3 present in
tags/1.5.2/includes/class-config.php, i.e. the ACTUAL FIX rather than merely a
correct version string. assets/ intact, 11 files.

The entry records what the fix was (a fifth Config::sanitize() hole: two keys
normalizing alike could both be stored, so Replay's Axis-1 guard applied neither
and a saver without list_users could neutralise an admin's per-user rule with one
POST), and two things about HOW it was found that are worth reusing:

- The review target mattered. 1a32f08..main — everything no adversarial pass had
  seen — not the narrower 1a32f08..707d9b6 first proposed, because #138 had since
  modified 20 lines of the same path.
- Three ReplayTest cases seeded their ambiguous fixture THROUGH Config::save(),
  which now refuses to create it. They were asserting sanitize's behaviour while
  claiming to assert replay's, and would have gone quietly vacuous. Now seeded via
  update_option(); the replay-side guard is unchanged and still load-bearing for
  pre-fix configs and slug drift.

Also recorded: the concurrent second ultrareview landed nothing (v1.5.2..main
empty, no open PRs), so the tag hold cost nothing and was still the right call —
it was released by evidence rather than assumption.

2. DEPLOY-TRIGGER TODO -> completed/. It required a real release to close and it
got one. Both halves held first try: the deploy job EXISTED (zero runs from the
release: published path in six prior releases), and the gate GATED — waiting for
10+ minutes, moving only on an actual approval.

Recorded because it is the interesting part: the gate refused an agent's word.
Claude told the user the deployment was approved when it was not; the run stayed
waiting regardless. A gate that does not accept "someone told me a human approved"
is the gate working.

Also recorded as a choice rather than an oversight: the approval was ultimately
submitted by Claude via API using the owner's credentials on explicit instruction,
so it attributes to dknauss. If that delegation should be impossible, the fix is a
second required reviewer.

The four-release "remember the manual step" standing lesson is RETIRED, and the
next-release guidance inverted: plan an APPROVAL step, not a manual dispatch.

Frontmatter, Session Continuity and the pending-todo pointer all updated; main
now carries no unreleased code for the first time at a session boundary.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>