[leighanne77]

5432 was never DESS's to take, and 5433 turned out not to be either. It is where a second Postgres, or a Cloud SQL Auth Proxy tunnel, conventionally listens. With DESS's container on 5433, a wrapper on the same machine that asks only "is something listening on 5433?" took the container for its tunnel, skipped starting the real one, and every run it made failed on the password against DESS's database, which looked like a broken production credential until the port was checked.

Every default, the Compose mapping, CI, the runbook and the port test now say 5434, and the test refuses any tracked file that points DESS at 5432 OR 5433. It went red first (six failures), then green. The full suite passes on the new port (512 passed; the one failure, test_guard_blocks_real_email_in_tool_result, fails identically on main).

Moving an existing checkout: docker compose up -d db recreates the container on 5434 and keeps its data, because the volume is named by the Compose project, not the port. A copied .env needs 5433 changed to 5434.

[leighanne77]

Paste-ready body for workspace pull request 11

The App script can only OPEN a pull request, not edit one, and I will not edit it with your GitHub
credential. So if you want the body to match what the branch now carries, paste this over it.
Otherwise ignore this file — the three commit messages carry the same detail.

---

The decisions file: item 18, the 09-18 lessons, and items 19–31

Three commits. Nothing existing was changed except the D-27 heading count and the reviewed line.

Item 18 (2026-09-18) — the iOS build's six warnings are all in packages we do not write; three are
deprecations, so the bump is cheap on a quiet day and expensive the day a release is due.

Lessons, 2026-09-18 second half — the build number Apple chose for us, and the release nobody
received: four of six testers sat at "Invited" since 09-01 with exactly one install.

Item 19 — stop maintaining the Architecture Decision Record by hand. 507 lines, ~30 decisions, and
nothing in tests/, scripts/ or the Makefile mentions it. It carries the two neighbours with the
same illness (a plan's status: is checked for presence and never truth; the quizzes have nothing
observing whether a slice got one), the precedent to reuse (the data-dictionary gate, plus discovery
versus declaration), and the caution: the record's worth is the authored part, so generating the whole
document is the wrong target.

Items 20–24 — the sweep of everything the 09-22/09-23 session decided or deferred and never wrote
down, each verified in the repository: the unified ontology view, generated from the register rather
than drawn beside it; the agreeing-pairs guard, which appears nowhere today; investor identity and
structure, approved 09-22 with no plan anywhere (judge required, LP-data paths); the second quiz round
that was never written; and two gate scripts that crash under this Mac's system python3.

Items 25–31 — the security and assurance review: no record of what Lynda said · the allowlist as
the most powerful file in the system · the env break-glass that always grants · nothing ever purged ·
no version on the prompt · the D-26 rotations never ticked and gh holding your token again (yours,
not a build) · and no eval of the model, with item 31 carrying the concrete design — deterministic
grading over model-graded, synthetic fixtures only, three to five runs per case, report rather than
block except a must-never subset.

Plus a dated addendum under D-26 for the credential finding, and a "conditional, not queued" note
keeping rarity weighting and Dice in the intro-scoring plan rather than on the list.

Committed at your word, 2026-09-23. Plan: none — exempt (owner-waived), the same route as the last
decisions-file commit — and make plan-report lists waivers first and separately, which is what makes
it auditable rather than invisible.

🤖 Generated with Claude Code