Re-resolve the graph and poll preconditions across canvas assertions

[coneilen]

Problem

#443 fixed the workspace chrome mount loop, and the gate then advanced past that block — and failed further downstream at line 1739:

switching to the mounted background tab changed terminal tab identity

This is the same defect, in its fourth location. Rather than fix a fifth site later, this change removes the pattern everywhere it remains.

The shape is:

$someControl.Invoke()
Start-Sleep -Milliseconds 150
$items = @(Get-DirectChildren $graph $rawWalker | Where-Object { ... })
Require (<assertion about $items>) "..."

Both halves are unsound:

1. The parent is stale. The fragment provider recreates the graph and its children when the canvas re-lays out, which is exactly what the preceding Invoke() triggers. Get-DirectChildren on a dead parent returns nothing for as long as it is asked, so no sleep length can outlast it.
2. The read is a single sample. One read after a fixed delay captures one arbitrary moment, so a fast runner passes and a loaded one fails.

In the tab-identity case these combine into a particularly quiet failure: the read returns an empty list, $workspaceTabs[0] is $null, and $null.Current.AutomationId yields $null without throwing — so the identity comparison fails with no indication that nothing was ever read.

This also explains why #440, #441, #442 and #443 each appeared to fix this region and then failed somewhere adjacent. There was never more than one defect; each fix corrected the site that happened to be reached first, and the next run surfaced the next occurrence.

Fix

Wait-ForGraphChildren re-resolves the graph parent on every attempt and polls until the caller's precondition holds, replacing all seven remaining sites.

The important constraint: it waits on preconditions only, never on an assertion. For the tab-identity checks it waits for two tabs to exist, then asserts their ids are unchanged — so a genuine identity regression still fails exactly as before. On timeout it returns the last observed state so each Require reports the real condition with its original message, rather than masking a failure as a timeout.

A latent vacuous pass, fixed

$initialWorkspaceTabId was captured from a tab list that had not necessarily remounted. When that read came back empty the baseline id was $null, and if the later read was also empty the assertion compared $null -eq $null and passed — reporting success while verifying nothing. This adds a Require that the initial tab exists before its id is captured, which is the one new assertion in this change.

Evidence

RED: CI run 35958493192 -> FAIL switching to the mounted background tab changed terminal tab identity at uia-live-gate.ps1:447, downstream of #443's fix and caused by the same stale-parent read.

GREEN: the windows-shell live gate now advances past the terminal tab identity block -> see the windows-shell check on this PR.

REGRESSION: Tools\windows\Tests\WindowsShell.Tests.ps1 -ZigExecutable $env:GRAPHCODE_ZIG0152 -> All 266 tests passed, Windows shell scaffold contract: PASS (40 source files executed).

Verification

• Parses clean.
• Require predicates compared against main with Compare-Object: zero lost, 379 -> 380, the single addition being the vacuous-pass guard described above. Every existing assertion and failure message is preserved verbatim.
• This check earned its keep: an intermediate revision of this change silently dropped Require ($workspaceTabs.Count -ge 2) "New Tab did not expose a mounted background tab", and the comparison caught it before it was committed.
• Zero remaining instances of fixed-sleep-followed-by-cached-$graph-read in the file.
• No C# touched, so the shared Add-Type block and its declaration counts are unaffected.