runtests: clear read eofchar on multiproc capture reads (8.6 ctrl-z truncation)

A native windows Tcl 8.6 runner's text channels default the READ eofchar
to ctrl-z (\x1a; the default 9.x removed), so the multi-process child
capture read in runtests_run_child_process silently truncated child
stdout at the first \x1a. The MULTISHELL polyglot deliberately carries a
ctrl-z tailblock and the scriptwrap/dtplite suites echo polyglot content
into their output, so every such file misreported as missing-cleanupTests
with zero observed tests under an 8.6 -jobs run (children completed and
exited 0; byte-proven: text read 8689 chars of a 256523-byte capture,
first \x1a at offset 8762 - full via binary read). Fix: fconfigure
-eofchar {} on both the stdout and stderr capture reads, making 8.6 match
the 9.x read semantics (translation stays auto). Immune paths that
obscured the mechanism: 9.x parent, MSYS/unix-mode 8.6, singleproc
(in-process runx capture). An apparent file-count/concurrency dependence
during diagnosis was a red herring - the 'passing' repros had -jobs
placed after the trailing file-tail globs, silently running singleproc.

New regression pin runner/testsuites/capture/ctrlz_capture.test emits a
\x1a byte before its test result, so a truncating runner turns it into a
missing-cleanupTests warn in any future multiproc run.

Verified: scriptwrap family under tclsh86ts -jobs 8 went from 5 warned
files / 0 counted tests to 10 counted (9 pass + pre-existing isunix
skip); scriptwrap+punkboot+binscripts neighbourhoods now count 123 tests
on the 8.6 leg, identical to the 9.0 leg (18 tests across 6 files were
previously invisible); runner suites incl the pin pass on 86ts and 90s;
9.0 behaviour unchanged.

Assisted-by: harness=claude; primary-model=claude-fable-5; api-location=anthropic.com

src/tests/runner/testsuites/capture/ctrlz_capture.test

@@ -0,0 +1,28 @@
+# -*- tcl -*-
+#added 2026-08-09 (agent) - regression pin for the runtests.tcl multiproc capture read
+#(runtests_run_child_process): a native windows Tcl 8.6 runner's text channels default to
+#ctrl-z (\x1a) as the READ eofchar (9.x removed the default), which silently truncated the
+#child stdout capture at the first \x1a in output. The MULTISHELL polyglot deliberately
+#carries a ctrl-z tailblock and the scriptwrap/dtplite suites echo polyglot content, so
+#every such file misreported as missing-cleanupTests with zero observed tests under an
+#8.6 -jobs run. The capture reads now clear -eofchar (matching 9.x semantics).
+#
+#This file deliberately emits a ctrl-z byte on stdout BEFORE its test result and summary
+#line: under a truncating runner the summary vanishes and THIS file reports
+#missing-cleanupTests (warn status) - the pin is this file completing with its pass
+#counted in any multiproc run on any supported runner interpreter. Singleproc runs
+#capture in-process (shellrun runx) and were never affected; the fixture is harmless there.
+package require tcltest
+
+namespace eval ::testspace {
+    namespace import ::tcltest::*
+
+    #the ctrl-z byte on stdout IS the fixture - flushed so it precedes all tcltest output
+    puts stdout "ctrlz-capture-fixture: \x1a <- ctrl-z byte emitted before test output"
+    flush stdout
+
+    test runtests_capture_survives_ctrlz {a ctrl-z byte in child output does not truncate the runner's capture of this file (8.6 read-eofchar class)} -body {
+        return 1
+    } -result 1
+}
+tcltest::cleanupTests

src/tests/runtests.tcl

@@ -419,13 +419,25 @@ proc runtests_run_child_process {exe exemodeargs bootstrap payloadfile testfile
     }
     set out ""
     set err ""
+    #-eofchar {}: a native windows Tcl 8.6 runner's text channels default to ctrl-z (\x1a)
+    #as the READ eofchar (removed in 9.x), silently truncating the capture at the first
+    #\x1a in child output. The MULTISHELL polyglot deliberately carries a ctrl-z tailblock
+    #and the scriptwrap/dtplite suites echo polyglot content, so their captures always
+    #contain \x1a - the truncation ate the tcltest summary line and every file reported
+    #missing-cleanupTests with zero observed tests under an 8.6 -jobs run (2026-08-09;
+    #proven by byte-level read diagnostics: text read 8689 chars of a 256523-byte capture,
+    #first \x1a at offset 8762). Clearing the eofchar makes 8.6 match the 9.x read
+    #semantics; translation stays auto (crlf handling unchanged). Pinned by
+    #runner/testsuites/capture/ctrlz_capture.test.
     if {[file exists $outfile]} {
         set fd [open $outfile r]
+        fconfigure $fd -eofchar {}
         set out [read $fd]
         close $fd
     }
     if {[file exists $errfile]} {
         set fd [open $errfile r]
+        fconfigure $fd -eofchar {}
         set err [read $fd]
         close $fd
     }