countWindowHeight calls update and can re-enter UI/socket handlers

[adkins17]

Summary

::potato::countWindowHeight calls Tk's global update, allowing a geometry/NAWS
measurement to execute unrelated pending UI, input, timer, or socket callbacks.

Working stock patch with latency tester:
https://github.com/potatomushclient/potato/pull/539.patch

Pull request: #539

Why this matters

countWindowHeight is reached from <Configure> through connMaybeNAWS. Calling
update from inside that path processes the complete Tk event queue before the
current handler returns. That permits nested configure handling and can re-enter
socket reads, input dispatch, timers, or other UI work during resize and layout.

This is observable as intermittent UI stalls and creates unsafe event ordering even
when an individual height calculation appears inexpensive.

Stock reproducer

The attached test queues an unrelated zero-delay callback immediately before calling
countWindowHeight:

• Unmodified stock: the callback fires inside countWindowHeight and the regression
  assertion fails.
• Patched stock: the callback remains pending until the normal event loop resumes,
  and all assertions pass.

The test also verifies that the function still returns a positive visible row count.

Latency results

Windows, Tcl/Tk 8.6.18, upstream commit
46d18387bd6b156b85176fa588a60991cbba6235:

Build	Re-entry assertion	p50	p95	max
Stock	FAIL	0.011 ms	0.060 ms	0.600 ms
Patched fresh worktree	PASS	0.005 ms	0.006 ms	0.237 ms

The primary fix is removal of nested event-loop processing. The focused 500-call
measurement also showed an approximately 90 percent p95 reduction.

Included tester

The patch adds:

• tests/Test-WindowHeightLatency.ps1
• tests/window-height-latency.test.tcl

Run with:

./tests/Test-WindowHeightLatency.ps1 -Runtime C:/path/to/tclkit-8.6.18-win64-intel-tk.exe

Validation

• Built directly against stock potatomushclient/potato:master
• Fresh-worktree patch apply check: PASS
• Downloaded GitHub PR .patch apply check: PASS
• Reverse patch check: PASS
• PowerShell parser: PASS
• Patched Tcl/Tk assertions: PASS
• No split-screen, packet-debugger, telemetry, or other custom-client code

[adkins17]

Repeated stock latency report

I reran the patch's stock tester five times against untouched upstream and five
times against the patched stock tree. Each run performs 500 calls to
::potato::countWindowHeight, for 2,500 measured calls per build.

Environment:

• Windows
• Tcl/Tk 8.6.18
• Stock baseline: 46d18387bd6b156b85176fa588a60991cbba6235
• Identical runtime and test workload for both builds
• Working patch and tester: https://github.com/potatomushclient/potato/pull/539.patch

Build	Run	p50	p95	Maximum	Re-entry assertion
Stock	1	0.006 ms	0.013 ms	0.274 ms	FAIL
Stock	2	0.012 ms	0.042 ms	0.149 ms	FAIL
Stock	3	0.006 ms	0.009 ms	0.565 ms	FAIL
Stock	4	0.007 ms	0.011 ms	0.059 ms	FAIL
Stock	5	0.013 ms	0.021 ms	0.747 ms	FAIL
Patched	1	0.002 ms	0.004 ms	0.026 ms	PASS
Patched	2	0.003 ms	0.004 ms	0.023 ms	PASS
Patched	3	0.003 ms	0.004 ms	0.119 ms	PASS
Patched	4	0.003 ms	0.004 ms	0.033 ms	PASS
Patched	5	0.004 ms	0.005 ms	0.041 ms	PASS

Aggregate comparison

Build	Calls	Median run p95	Mean run p95	Worst maximum	Re-entry passes
Stock	2,500	0.013 ms	0.019 ms	0.747 ms	0/5
Patched	2,500	0.004 ms	0.004 ms	0.119 ms	5/5

The patched median run-level p95 is approximately 69% lower. More importantly,
unmodified stock executed the unrelated queued callback from inside
countWindowHeight in every run. The patched function never did.

This microbenchmark measures an otherwise quiet event queue. In normal client use,
the cost of stock's nested update is unbounded by the height calculation itself:
it also executes whatever socket, input, timer, redraw, or configure work happens
to be queued at that moment. The deterministic re-entry assertion is therefore the
primary regression result; the timing reduction is supporting evidence.

Raw run data was retained locally as latency-runs.csv. The tester is included in
PR #539 as:

• tests/Test-WindowHeightLatency.ps1
• tests/window-height-latency.test.tcl