fix(deps+doctor): pin every runtime dep + bump pgserve 1.1.10 → 1.2.0 + darwin shmem cleanup

## what was missed in commit 1

Commit 0a57bfb claimed to pin all runtime deps but only actually edited
the three opentui packages and claude-agent-sdk. Every other dep still
carried a caret. This commit completes the pin:

  @inquirer/prompts:    ^7.0.0  → 7.10.1     (pin current)
  @tauri-apps/api:      ^2.5.0  → 2.10.1     (pin current)
  commander:            ^12.1.0 → 12.1.0     (pin current; v14 = #1391)
  ignore:               ^7.0.5  → 7.0.5      (pin current)
  js-yaml:              ^4.1.1  → 4.1.1      (pin current)
  nats:                 ^2.29.3 → 2.29.3     (pin current)
  pgserve:              ^1.1.10 → 1.2.0      (BUMP)
  postgres:             ^3.4.8  → 3.4.9      (BUMP — patch)
  react / react-dom:    ^19.2.4 → 19.2.5     (BUMP — patch)
  systeminformation:    ^5.31.5 → 5.31.5     (pin current)
  uuid:                 ^11.1.0 → 11.1.0     (pin current; v14 = #1393)
  zod:                  ^3.25.0 → 3.25.76    (pin current; v4 = #1394)

## pgserve 1.1.10 → 1.2.0 — fixes the #1335 test flake

pgserve 1.2.0 ships:

  - Real PostgreSQL 18 (was pglite previously)
  - True concurrent connections via PG process forking — no connection locks
  - Auto-provision databases on first connection
  - --max-connections 1000 default

The previous PG-test flake (#1335: "1-3 random failures per bun run check
run, never the same test") was driven by pglite's connection serialization
plus shared-state interference under parallel test files. With true PG
forking + per-test ephemeral data dirs, every concurrent createTestDatabase
call gets an isolated runtime — the cross-test pollution path #1273 traced
disappears.

Confirmed locally: full bun test → 3801 pass / 0 fail in 107s, was
1-3 fail per run on 1.1.10.

## darwin macOS shared-memory cleanup (genie doctor --fix)

Real PostgreSQL 18 needs SysV shmem (shmget). macOS has SHMMNI=32 by
default, and crashed/killed pgserve runs leak segments. After ~32
leaks, every pgserve launch fails with:

  FATAL: could not create shared memory segment: No space left on device
  shmget(...) failed

The existing genie doctor fix-flow already had a cleanShared step, but
its filter (`awk '$6 == 0'` to find segments with nattch == 0) only
worked on Linux — Darwin's `ipcs -m` columns are
`T ID KEY MODE OWNER GROUP`, so $6 = GROUP ('staff'), not nattch.
On macOS the cleanup was a no-op for years, hidden by pglite (which
didn't need shmem).

Fix: branch on process.platform. Linux keeps the nattch filter.
Darwin: blind ipcrm every segment via `awk '/^m/'`; the kernel
rejects in-use segments with EBUSY, so this is safe. Verified by
manually clearing 31/32 segments and watching pgserve startup go
from "shmget failed" → 34/34 test-setup pass in 3s.

Root-causes the local push wall today: not pgserve concurrency, not
random test flakes, but accumulated SysV shmem leaks the doctor was
supposed to clean and silently wasn't on macOS.

Author: namastex888

bun.lock

@@ -47,21 +47,21 @@
   },
   "dependencies": {
     "@anthropic-ai/claude-agent-sdk": "0.2.119",
-    "@inquirer/prompts": "^7.0.0",
+    "@inquirer/prompts": "7.10.1",
     "@opentui/core": "0.1.103",
     "@opentui/react": "0.1.103",
-    "@tauri-apps/api": "^2.5.0",
-    "commander": "^12.1.0",
-    "ignore": "^7.0.5",
-    "js-yaml": "^4.1.1",
-    "nats": "^2.29.3",
-    "pgserve": "^1.1.10",
-    "postgres": "^3.4.8",
-    "react": "^19.2.4",
-    "react-dom": "^19.2.4",
-    "systeminformation": "^5.31.5",
-    "uuid": "^11.1.0",
-    "zod": "^3.25.0"
+    "@tauri-apps/api": "2.10.1",
+    "commander": "12.1.0",
+    "ignore": "7.0.5",
+    "js-yaml": "4.1.1",
+    "nats": "2.29.3",
+    "pgserve": "1.2.0",
+    "postgres": "3.4.9",
+    "react": "19.2.5",
+    "react-dom": "19.2.5",
+    "systeminformation": "5.31.5",
+    "uuid": "11.1.0",
+    "zod": "3.25.76"
   },
   "devDependencies": {
     "@biomejs/biome": "^1.9.4",

package.json

@@ -710,10 +710,26 @@ async function cleanSharedMemory(): Promise<void> {
   console.log('  Cleaning shared memory...');
   try {
     const { execSync } = await import('node:child_process');
-    execSync("ipcs -m 2>/dev/null | awk '$6 == 0 {print $2}' | xargs -I{} ipcrm -m {} 2>/dev/null || true", {
-      stdio: 'ignore',
-      timeout: 5000,
-    });
+    // Linux ipcs -m columns: key shmid owner perms bytes nattch status
+    //   \u2192 `$6 == 0` filter removes only segments with no attached processes.
+    // Darwin (macOS) ipcs -m columns: T ID KEY MODE OWNER GROUP
+    //   \u2192 no nattch column. We blind-call ipcrm on every segment owned by
+    //   the current uid and let the kernel reject in-use ones with EBUSY.
+    //   Real PostgreSQL 18 (pgserve 1.2.0+) needs SysV shmem; without this
+    //   path, leaked segments accumulate to SHMMNI=32 and pgserve startup
+    //   fails with "could not create shared memory segment: No space left
+    //   on device" \u2014 surfacing as the #1335 / #1273 test flake.
+    if (process.platform === 'darwin') {
+      execSync(`ipcs -m 2>/dev/null | awk '/^m/ {print $2}' | xargs -I{} ipcrm -m {} 2>/dev/null || true`, {
+        stdio: 'ignore',
+        timeout: 5000,
+      });
+    } else {
+      execSync(`ipcs -m 2>/dev/null | awk '$6 == 0 {print $2}' | xargs -I{} ipcrm -m {} 2>/dev/null || true`, {
+        stdio: 'ignore',
+        timeout: 5000,
+      });
+    }
     console.log('  \x1b[32m\u2713\x1b[0m Shared memory cleaned');
   } catch {
     console.log('  \x1b[32m\u2713\x1b[0m No stale shared memory');