fix: bootstrap PostgreSQL without a working service manager (#823)

.claude/skills/fix-issue.md

@@ -473,3 +473,4 @@ extracting `OffsetExpression`/`LimitExpression`.
 | An association's line anchors — where the connector attaches to the entity boxes in the domain model editor — are absent from `DESCRIBE ASSOCIATION`, and manual adjustments made in Studio Pro do not survive an mxcli round trip | `DomainModels$Association.ParentConnection`/`ChildConnection` (the string `"x;y"`) were **hardcoded** to `"0;50"`/`"100;50"` in BOTH writers and never read by either parser. Because every association write rebuilds the whole element, this was not an omission but active destruction: a documentation-only `alter association … set comment` reset them | `sdk/domainmodel/connection.go` (new: `ParseConnectionPoint`/`FormatConnectionPoint`, `Default*Connection`), `sdk/domainmodel/domainmodel.go` (fields → `*model.Point`), `sdk/mpr/parser_domainmodel.go` + `sdk/mpr/writer_domainmodel.go`, `mdl/backend/modelsdk/domainmodel.go` + `domainmodel_write.go`, `mdl/executor/cmd_associations.go` (`describeConnectionPoints`) | **A feature request that says "X is not exposed" may be hiding "X is destroyed"** — check the write path before scoping the read path. The A/B that settled it: a blank 11.13 app's own `Administration.AccountPasswordData_Account` stores `0;54/100;54`, so a Studio-Pro-authored association is a free fixture for "did mxcli overwrite this?" — no Studio Pro needed. **Learn the value's constraints from the LOADER, not from the shape**: hand-patch and run `mx check` — `"0.5;50"` dies with `StorageLoadException` (integers required) while `"0;500"` and `"-20;50"` load with 0 errors (no range check), so out-of-range values must round-trip untouched. **A zero value is not an absent value** — `{0,0}` is a real anchor (top-left), which forces the field to be a POINTER; a plain `model.Point` cannot distinguish "unset" from "top-left" and would silently rewrite it. **Fix both engines**: they share the semantic model, and a fix in one is invisible to a user on the other. **Emit unauthorable data as a COMMENT** — DESCRIBE output must stay re-executable, and inventing syntax (`@anchor(parent: bottom-left, …)`) would bake in a vocabulary the storage does not have: the pair is CONTINUOUS, not 8 named anchors (observed x values 0 9 11 17 18 47 49 50 65 77 78 84 87 100). **The marketplace is the sample** when you need to know what Studio Pro actually writes: `mxcli marketplace download <id>` gives real Mendix-authored models, and a module .mpk holds either a raw BSON `project.mpr` or an MPR v1 SQLite one — 88 coordinate pairs from three modules turned "looks like percentages" into a measurement (all 0..100; 85 of 88 pin one coordinate to exactly 0 or 100). **Rule a unit out from the model, not the values**: pixels is impossible because `DomainModels$EntityImpl` stores only `Location` and NO size — the box is sized by the editor from the name and attribute list, so a pixel anchor would have nothing to measure against. Not applicable to `CrossAssociation`, which has no connection properties and crashes Studio Pro if given them (#50). Tests `sdk/domainmodel/connection_test.go`, `mdl/backend/modelsdk/association_connection_test.go`, `sdk/mpr/writer_domainmodel_test.go`, example `mdl-examples/bug-tests/872-association-line-anchors.mdl`. upstream #872 |
 | An association's line anchors can be preserved but not AUTHORED — a scripted domain model cannot lay out its own connector lines, so `@Position(x, y)` gets you boxes and nothing gets you the lines between them | Feature gap, not a defect. `DomainModels$Association.ParentConnection`/`ChildConnection` had no MDL surface | `mdl/grammar/domains/MDLDomainModel.g4` (`SET ANCHOR`/`anchorPoint` — the ONLY grammar change), `mdl/visitor/visitor_association.go` (`anchorAnnotation`, `annotationParenPoint`, `anchorCoord`), `mdl/ast/ast_association.go` (`FromAnchor`/`ToAnchor` on both create and alter), `mdl/executor/cmd_associations.go` (`applyAnchors`, `describeConnectionPoints`) | **Look for an existing annotation before inventing one** — `@anchor(from:, to:)` already existed for microflow sequence flows, asking the same question (where does the connector attach), and `annotationParamName` already admitted FROM and TO, and `(x, y)` was already `annotationParenValue`: CREATE needed **zero** grammar. The two forms cannot be confused because the microflow one names its inner params (`(from: right, to: left)`) while a coordinate pair is positional. **Let the storage pick the value type**: the measured pair is continuous (x takes 14 distinct values across 88 samples), so named anchors were never an option — see the preservation row above for how that was established. **Silence must mean "preserve", not "default"** — naming one end sets it and omitting one keeps what is stored, which is what stops a `create or modify association` about the delete behaviour from flattening a hand-tuned line; the AST carries POINTERS so "not mentioned" and "mentioned as (0, 0)" stay distinguishable. **Reject what the LOADER rejects, at check time**: a fractional coordinate must error, not be truncated to 0 — Mendix refuses to open such a project, and a silently-wrong value in a file that still loads is the worse failure. **Prove DESCRIBE round-trips by parsing its own output** — asserting on a string literal passes against a formatter emitting something nothing can read. Tests `mdl/visitor/visitor_association_anchor_test.go`, `mdl/executor/cmd_associations_anchor_test.go`, example `mdl-examples/bug-tests/872-association-line-anchors.mdl`. upstream #872 |
 | `ALTER PAGE` over `--mcp` fails against Studio Pro **11.13** with `pg_patch_page: … PROP_NOT_PRIMITIVE: Property 'widgets' is not a primitive property`. `CREATE PAGE` is fine; the page itself is left intact | 11.13 gave `pg_read_page` a **`depth` argument defaulting to 4**, replacing anything deeper with the literal string `"..."`. ALTER PAGE is read-modify-**replace-whole-page**, so the truncated read went straight back as the new page body. Measured live: `Administration.Account_Overview` read 32,594 bytes at full depth but **1,052 bytes** at the default, its entire tree reduced to `{"widgets":["...","..."]}`. Every ordinary page truncates — three of three PgTest pages did | `mdl/backend/mcp/page.go` (`pgReadPage`, `pgReadFullDepth`, `hasTruncationSentinel`), `mdl/backend/mcp/client.go` (`SupportsToolArg`) | Request the full depth, and **guard rather than trust it**: refuse a read still carrying the sentinel instead of letting a partial page reach a write (ADR-0005 guard-don't-drop). Two traps. (1) **Do not send `depth` unconditionally** — 11.11/11.12 declare `pg_read_page` `additionalProperties:false` without it, so the whole call fails; gate on a live `tools/list` probe of the tool's input schema, because `serverInfo.version` is frozen at `1.0.0` across 11.11/11.12/11.13 and cannot discriminate releases. (2) **Match the sentinel only as an array element** — a caption or title legitimately reading `"..."` is real content, and a naive substring scan rejects valid pages. The release notes announced none of this, exactly as 11.12 silently removed `pg_write_page` (#697): on any Studio Pro upgrade, re-probe `tools/list` and diff the input schemas, not just the tool names. Tests `mdl/backend/mcp/page_depth_test.go`; controls: stub the depth arg (full-depth test fails) and stub the guard (truncation test fails) |
+| `mxcli run --ensure-db` fails to start PostgreSQL when no service manager becomes ready (e.g. Arch): `exec: "pg_ctlcluster": executable file not found in $PATH`, though `initdb`/`pg_ctl`/`psql` are present | `startLocalPostgres` only knew the `service`/`pg_ctlcluster` helpers (the latter a placeholder that could never run), and role/database provisioning assumed `sudo -u postgres`, which a user-owned cluster does not need | `cmd/mxcli/docker/ensuredb.go` (`startLocalPostgres`, `startUserCluster`, `resolveSuperuser`) | Fall back to a user-owned `initdb`/`pg_ctl` cluster under `~/.mxcli/postgres`, but never initialize a competitor when the requested TCP port is already owned by a service-started server that is still recovering. Enforce the passwordless-superuser boundary instead of assuming it: explicitly tighten an existing socket directory to `0700`, set the socket itself to `0700`, and persist `listen_addresses`/`port`/`unix_socket_directories` in `postgresql.conf` so a plain later `pg_ctl start` stays safe. Refuse, with a cleanup path, a cluster made by the earlier development revision if its host authentication is still `trust`. Reuse a running cluster only when both the port and socket directory in `postmaster.pid` match. Keep `sudo -u postgres` for system clusters. Tests stub every external tool and exercise slow service startup, legacy-cluster refusal, permissions, persisted settings, and repeated/running cases. Issue #823 |

.claude/skills/mendix/run-local.md

@@ -50,8 +50,14 @@ association catalog only at startup; behavioural changes are hot-reloaded.
 - A **PostgreSQL** database (defaults: `127.0.0.1:5432`, user `mendix`, db derived
   from the project name; override with `--db-host/--db-name/--db-user/--db-password`).
   - **`--ensure-db`** provisions it for a fresh session: starts local Postgres if the
-    port is down and creates the role + database if missing (local superuser via
-    `sudo -u postgres`). Remote hosts are only checked, not provisioned.
+    port is down and creates the role + database if missing. It uses a service
+    manager, or a user-owned `initdb`/`pg_ctl` cluster under `~/.mxcli/postgres`
+    when no service becomes ready (e.g. Arch) — needing no `postgres` OS account or `sudo`.
+    Remote hosts are only checked, not provisioned.
+    The user-owned cluster persists across sessions; its server log is
+    `~/.mxcli/postgres/server.log`. Stop it with
+    `pg_ctl -D "$HOME/.mxcli/postgres/data" stop`. To remove it, stop it first and
+    then delete `~/.mxcli/postgres` (this permanently deletes its databases).
   - Without `--ensure-db`, create it once and the command errors if it's unreachable:
 
     ```bash

CHANGELOG.md

@@ -43,6 +43,7 @@ Headline: **A full Mendix build-and-test loop that fits on an iPad** — you can
 - **`run --local` robustness** — the web client bundle is verified after boot, not before; stale ports are refused and the holding process is named; a relative project path resolves where MxBuild is called; the browser bundle is re-served after a structural `--watch` change; boot uses live-preview flags so `mxcli oql` can reach the app; the hub tunnel retries forever and re-registers on a heartbeat 404.
 - **Theme** — `mxcli theme remove` with no name now reads the installed theme from the `mxcli:theme` markers instead of removing the built-in default and reporting a silent no-op; switching themes no longer orphans the previous theme's block in `_mxcli-atlas-map.scss`; the topbar language selector, filter-operator popovers, login page, and Data Grid 2 now follow the palette (the language selector went from 1.13:1 to 17.79:1 / 19.47:1 measured contrast).
 - **JavaScript action sources were written to the wrong directory** — `CREATE JAVASCRIPT ACTION` wrote to `javascriptsource/<ModuleName>/actions/`, but Mendix reads a **lowercased** module directory. MxBuild found nothing there, generated a stub whose body throws `JavaScript action was not implemented`, and bundled that — so the action parsed, passed `mxcli check`, built cleanly, and threw the moment it ran. Only reproduced on a case-sensitive filesystem, which is why it went unnoticed on macOS and Windows.
+- **`mxcli run --ensure-db` can start PostgreSQL without a working service manager (#823)** — on hosts that ship neither `service` nor Debian's `pg_ctlcluster` (e.g. Arch Linux), `--ensure-db` failed with `exec: "pg_ctlcluster": executable file not found in $PATH` even though the portable `initdb`/`pg_ctl`/`psql` tools were present. `startLocalPostgres` now falls back to a user-owned cluster under `~/.mxcli/postgres` when no service becomes ready, but never starts a competitor while another process owns the requested port. The cluster is idempotent and needs neither a `postgres` OS account nor passwordless `sudo`: its listen address, port, private socket directory, and `0700` socket permissions persist in PostgreSQL's own configuration; loopback TCP uses SCRAM while role/database provisioning uses local trust only through that private socket. Anyone who ran an earlier development revision of this fix must stop PostgreSQL, remove `~/.mxcli/postgres`, and rerun `--ensure-db`; reuse detects and refuses its insecure host-trust records. The existing `sudo -u postgres psql` path remains for system clusters.
 - **Widget sync wrote duplicate GUIDs and corrupted the project** — reconciliation now removes stale properties and syncs attributes via `AugmentTemplate` so the widget `Type` is exact.
 
 ## [0.16.0] - 2026-07-12