Issue #832: bound TermWidth probes and honor $COLUMNS

[Yaminyam]

Fixes #832.

Implements both fixes suggested in the issue discussion, scoped to tools/TermWidth.lua so the global capture() callers are unaffected (per @mrcawood's note in #832).

Changes

• l_askSystem honors $COLUMNS before the stty/tput probes. Interactive shells already export COLUMNS from their SIGWINCH handler, so the common path skips the subprocess probes entirely. This is the cheapest fix and also covers wedged-PTY environments where the shell still has a sensible COLUMNS.

• stty size / tput cols capture() calls are wrapped with timeout (GNU coreutils) when the binary is on PATH. Detection happens once, lazily, and the result is cached. A broken PTY whose ioctl(TIOCGWINSZ) never returns can no longer hang module indefinitely — the probe is killed after one second and the next fallback (or the caller's default) is used. When timeout is not on PATH the prefix is empty, so the pre-fix behavior is preserved on systems without coreutils (no regression).

Both changes are local to TermWidth.lua; capture() and its other callers are untouched.

Regression test

Adds rt/termWidthFallback/. It installs a fake stty/tput on PATH that report a sentinel width (999) and log every invocation, then runs module -D avail twice — once with $COLUMNS set, once without. The log is expected to be empty in the COLUMNS-set scenario (subprocess probes skipped) and non-empty when COLUMNS is unset (probes executed and returned the sentinel).

Note on goldens: my local macOS doesn't currently have a full Lmod build environment ready (lua + luaposix + luafilesystem + Hermes), so the out.txt / err.txt for the new RT aren't included in this commit yet. Happy to follow up with a bless commit once CI shows the actual output, or if a maintainer can run tm -bless rt/termWidthFallback on their end I can fold the result in.

Verification done

• Reproduced the original hang in the issue's docker setup (stty size → sleep 999999), and verified that module use ... returns normally after this patch, with both LMOD_TERM_WIDTH unset and COLUMNS unset (relying on the timeout wrapper) — fallback to default 80 columns.
• With COLUMNS=120 set, confirmed the fake stty/tput probes are not invoked at all.
• Re-read the existing call sites of capture() to confirm nothing outside TermWidth.lua changes behavior.

[Yaminyam]

Thanks @mrcawood — pulled the updated branch and tested it in a clean Rocky Linux 9 container (lua-posix 35.0, the EL9-packaged luaposix). Two findings: the width-bounding logic works, but the branch as-is (dbca31f7f) does not load on EL9. I've pushed a small follow-up (d097496) that fixes the load and kept everything else as you wrote it.

1. Load-time break on luaposix 35 (EL9/Rocky/Alma 9)

dbca31f7f indexes posix.signal.kill on the merged posix table:

lua: tools/TermWidth.lua:41: attempt to index a function value (field 'signal')
stack traceback:
	tools/TermWidth.lua:41: in main chunk
	[C]: in function 'require'
	src/lmod.in.lua:108: in main chunk

On luaposix 35 posix.signal is the signal() function, not a table, so .kill errors at require time and lmod --version aborts before doing anything — Lmod won't start at all on EL9. (Interestingly the same file already calls posix.signal(SIGALRM, ...) as a function at lines 117/134, so the two uses were inconsistent.)

My follow-up requires the posix.signal submodule and pulls kill / signal / SIGALRM / SIGKILL from it — a table exposing both the functions and the constants across the luaposix versions Lmod supports:

local psignal = require("posix.signal")
local kill    = psignal.kill
local SIGALRM = psignal.SIGALRM
local SIGKILL = psignal.SIGKILL
...
local oldSig = psignal.signal(SIGALRM, function() ... end)
...
psignal.signal(SIGALRM, oldSig)

2. With that fix, the bounding works

Reproducer = my original wedged stty from the issue (hang on size), timing real lmod bash avail against a 15 s hard kill:

case	elapsed	result
naive stty size under the wrapper (control)	killed at 3 s	wrapper genuinely hangs
wedged stty, COLUMNS and LMOD_TERM_WIDTH unset	~1 s	SIGALRM fires → 80-col default ✅
wedged stty, COLUMNS=120	0 s	probes skipped entirely ✅
wedged stty, LMOD_TERM_WIDTH=100	0 s	short-circuits as before ✅

Before the fix, case 1 hung indefinitely; after, it's bounded to the 1 s alarm. $COLUMNS and LMOD_TERM_WIDTH are both honored.

The follow-up commit is on the branch. 1 s felt fine in my docker / ttyd / etty setups. Happy to adjust anything else you'd like before merge.