swayfx-mosaic/tooling/probes/step20c-commit-present-input/README.md

STEP20C commit/present/input ordering — EXPERIMENTAL

The protocol was written before this fixture. Everything here is a source fixture for a disposable private laboratory; no code enters the production compositor, SceneFX ABI, package, cumulative Mosaic patch or daily session.

Replay

prep=$(sh tooling/probes/step20c-commit-present-input/prepare.sh)
LAB=$(printf '%s\n' "$prep" | sed -n 's/^LAB=//p')
for mode in control private_standard private_mosaic; do
  PYTHONDONTWRITEBYTECODE=1 python tooling/probes/step20c-commit-present-input/run.py \
    "$LAB" "$mode" --cycles 5
done
PYTHONDONTWRITEBYTECODE=1 python tooling/probes/step20c-commit-present-input/analyze.py "$LAB"

The prepare chain rebuilds the pinned private SceneFX/Mosaic fixtures and fetches wlroots 0.20.2 at commit d783533489e1f75d6886c2ab5c5960090ef268f8 inside the new lab. Its two audited scene files are hash-checked by the Mutation Coverage preparator. surface-ordering.patch adds a passive top-level surface commit trace after wlroots applies the client input region. Existing private traces provide build, commit, present, hit and seat notification. The client uses ordinary wl_surface.set_input_region and wl_surface.commit requests.

The runner reuses the frozen headless runtime, explicit IPC socket, process group cleanup and geometry checks. It defaults to five bounded cycles and ASI_MODE=0 (trace-only). Standard controls are unmodified SwayFX splith and private SwayFX splith. Native output events are uninstrumented in the unmodified control. Camera requests in the private Mosaic case supply ordinary scene damage; no backend scheduling is changed. A missing natural commit/mutation/present interleaving must be reported INCONCLUSIVE.

The runner is sequential: command() waits for IPC response, client_command() waits for a client log confirmation, and inject() waits for pointer confirmation. A Mosaic camera reply precedes submission of the next input-region request. Those process round trips can exceed the native commit-to-present interval. The executed callbacks and delivery remain valid; the timing question requires a separately recorded nonblocking experiment.

For a complete evidence manifest outside Git, hash every regular file under the three runs/ directories plus the ten structured results, excluding the two older source/artifact manifests and the new manifest itself. The historical frozen lab's raw-evidence-sha256.txt contains 94 entries and hashes to 9b74584936d44be5accb1f3882482fbd1eb85bc5c696b4551d45076b76ff4f8d. The independent replay manifest has the same scope and hashes to 7d4e5c3860cc5ea3b63668441efa58bf31ef3197c6d22890cbaccee70543d41a.

Set CPI_SURFACE_TRACE=0 before run.py for a paired workload with only the additional ORDER_SURFACE_COMMIT log site disabled. Keep these paired runs in a separate lab from the three analysis runs; analyze.py deliberately rejects ambiguous duplicate mode runs. The old private build, hit and commit/present traces still execute in this comparison.

Raw logs, screenshots, generated runner, binaries, build trees, sockets and structured JSON analysis remain in LAB, outside Git. Stop or terminate the runner normally to remove its private runtime. A killed runner may leave its lab and sockets for manual, targeted cleanup; never touch the live Sway socket. The real rejected-commit backend path remains NOT TESTED.