test: run the gfx examples — no build variant ever did (#886)

[InauguralPhysicist]

Section [97] skipped example programs by content:

if grep -qE 'gfx_|net_listen' "$f"; then EX_SKIP=$((EX_SKIP + 1)); continue; fi

That is unconditional. make gfx builds the extension and [132] exercises the real SDL renderer — but the nine gfx_ examples were skipped in every variant, so nothing covered them anywhere. The suite reported this honestly ("10 gfx skipped"); the gap was that nothing else covered them either.

What changed

The skip is now gated on build capability (the same probe [132] uses). Under a gfx build each demo runs against the dummy video driver, memory-capped — an unbounded UI run can take the whole machine.

A gfx demo ends in ui.app_loop, an interactive event loop that no quit event ever reaches headlessly. So reaching it (rc 124) is the pass signal: the program got through parse, module load, widget construction and layout without erroring, which is the failure class this section exists to catch. It deliberately does not verify loop behavior — [132] and the lib/ui sections own that — and the section header says so rather than implying more coverage than it has.

Validation

Planted the issue's own bug back in (a call to the private ui._layout in ui_hex.eigs):

[97] Example programs (examples/*.eigs; gfx demos INCLUDED)
  FAIL(1): ../examples/ui_hex.eigs (gfx demo errored before its event loop)

Both paths checked:

build	result
make gfx	81 examples run (was 72), 1 net skipped — 3863/3863
make (release)	72 run, 10 gfx/net skipped — 3804/3804

One honest correction to the issue

The broken example it names is already fixed. b49e85c — an unrelated sandbox/zlib PR — deleted the ui._layout of [root, 0, 0] line by accident, after the issue was filed:

$ git show b49e85c -- examples/ui_canvas_events.eigs
-ui._layout of [root, 0, 0]

All nine gfx demos currently reach their event loop cleanly. So this PR closes the coverage gap, not a live breakage — and that incidental fix is the issue's own argument made concrete: with nothing running these, they break and get fixed invisibly.

This is a tests-and-CI change only (tests/run_all_tests.sh, a CI step name, CHANGELOG); no runtime source is touched, so the sanitizer gates are unaffected and CI's own ASan job covers them.

Closes #886

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR closes a long-standing test coverage gap by ensuring the gfx_ example programs under examples/ are executed when the current build actually supports the gfx extension (rather than being unconditionally skipped by content). It does this by adding a gfx capability probe and running gfx demos headlessly under SDL’s dummy video driver with a short timeout and a memory cap, and updates CI and the changelog accordingly.

Changes:

• Gate example skipping on gfx build capability and run gfx demos headlessly (treating rc=124 timeout as success for interactive event loops).
• Update CI job step name to reflect gfx examples being exercised in the gfx build suite run.
• Document the change (and historical context) in CHANGELOG.md.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 2 comments.

File	Description
tests/run_all_tests.sh	Adds gfx capability probing and executes gfx examples under dummy SDL driver with timeout+memory cap instead of unconditional content-based skipping.
CHANGELOG.md	Records the new coverage behavior and the motivation/history from #886.
.github/workflows/ci.yml	Renames the gfx-suite CI step to reflect that gfx examples are now exercised.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

Copilot reviewed 3 out of 3 changed files in this pull request and generated no new comments.

Suppressed comments (2)

tests/run_all_tests.sh:3922

• The gfx-example runner hardcodes timeout 3, but this script explicitly supports macOS/BSD userlands by falling back to gtimeout (or no wrapper) via EIGS_TMO. Calling timeout directly will fail on systems that only have gtimeout, and it also ignores the existing detection logic. Reuse the detected wrapper command for the short gfx-demo timeout, and skip gfx demos if no timeout wrapper is available.

        # UI run can take the whole machine.
        EX_OUT=$( cd "$(dirname "$f")" && ulimit -v 2000000 2>/dev/null; \
                  cd "$(dirname "$f")" && SDL_VIDEODRIVER=dummy timeout 3 \
                  "$EIGS_ABS" "$(basename "$f")" </dev/null 2>&1 ); EX_RC=$?

tests/run_all_tests.sh:3890

• The section header comment implies rc=124 means the program reached parse/module-load/layout successfully, but rc=124 only indicates the timeout wrapper fired. That’s a reasonable proxy for these demos, but it’s not guaranteed (a hang earlier in startup would also be rc=124). Consider tightening the wording so it’s clear timeout is being used as a proxy rather than a definitive signal.

# the dummy driver no quit event ever arrives, so reaching it means timing out.
# That is the PASS signal here — rc 124 means the program got through parse,
# module load, widget construction and layout without erroring, which is the
# failure class this section exists to catch. It deliberately does NOT verify
# loop behavior; [132] and the lib/ui sections own that.

[Copilot]

Pull request overview

Copilot reviewed 3 out of 3 changed files in this pull request and generated no new comments.

Suppressed comments (3)

tests/run_all_tests.sh:3900

• gfx examples are treated as runnable whenever gfx is present, but the gfx-demo path later relies on timeout to terminate the headless event loop. On systems without timeout/gtimeout (which this runner explicitly supports by degrading), the suite would hang when it hits the first gfx example. Gate gfx-demo inclusion on the presence of the timeout wrapper, and emit a distinct header when gfx exists but no wrapper is available.

EX_GFX_PROBE=$(mktemp /tmp/eigs_ex_gfx_XXXXXX.eigs)
echo 'print of (gfx_text_width of ["m", 1])' > "$EX_GFX_PROBE"
EX_HAS_GFX=0
if ! ./eigenscript "$EX_GFX_PROBE" 2>&1 | grep -q "undefined variable"; then EX_HAS_GFX=1; fi
rm -f "$EX_GFX_PROBE"

tests/run_all_tests.sh:3922

• The gfx-demo execution path hard-codes timeout 3, which (a) bypasses the runner’s timeout/gtimeout detection (breaking macOS where only gtimeout may exist) and (b) doesn’t match the PR description’s stated 12s budget for reaching ui.app_loop, which could cause flakes on slower CI. Use the detected timeout wrapper and a configurable/appropriate budget.

    if grep -q 'gfx_' "$f"; then
        if [ "$EX_HAS_GFX" != "1" ]; then EX_SKIP=$((EX_SKIP + 1)); continue; fi
        # #886: reaching the event loop (rc 124) is the pass; any other
        # nonzero rc is a real setup failure. Memory-capped — an unbounded
        # UI run can take the whole machine.
        EX_OUT=$( cd "$(dirname "$f")" && ulimit -v 2000000 2>/dev/null; \
                  cd "$(dirname "$f")" && SDL_VIDEODRIVER=dummy timeout 3 \
                  "$EIGS_ABS" "$(basename "$f")" </dev/null 2>&1 ); EX_RC=$?

tests/run_all_tests.sh:3954

• The [97] PASS summary says examples “run clean” even when gfx demos are expected to time out as the success signal, and it keys off EX_HAS_GFX rather than whether gfx demos actually ran (e.g., gfx build but no timeout wrapper). Adjust the message/branching to reflect the real behavior and the new EX_CAN_RUN_GFX gate.

    if [ "$EX_HAS_GFX" = "1" ]; then
        echo "  PASS: all $EX_PASS example programs run clean (gfx demos included; $EX_SKIP net skipped)"
    else
        echo "  PASS: all $EX_PASS example programs run clean ($EX_SKIP gfx/net skipped — no gfx build)"
    fi