The exact public-text probe at c3d8881fbec4554a79a863b07a97335a9727c376 exited 0 on 2026-10-16, with outer containment stop confirmed at 05:00:49.611929 UTC. Genuine Firefox AT events, an accepted Orca51.1/Speech Dispatcher utterance and espeak-ng PCM growth satisfy this finite environment prerequisite. Actual packaged-console announcements, native suppression and accessibility/device certification are not established by this result.

DATA below means /run/media/kmosoti/data/FabricO11y. Raw evidence remains in DATA/results/browser-announcement-native-orca51-01; no owner desktop, bus, audio, network configuration or trust store was changed.

The frozen protocol requires genuine AT-SPI event → Orca → accepted SSIP225/message ID → exact espeak-ng module/mapped library → growing generated PCM, with original role=status, aria-live=polite and focus elsewhere. One eight-second action interval lies within the original150-second inner/180-second environment bound. No DOM-only or Orca debug-text-only result can pass. This is synthesis evidence, not a listener/acoustic test or an elapsed endurance measurement.

Pre-flight 6207769886 was posted04:00:15 UTC. The append-only ordered-cleanup binding and complete PCM diagnostic binding retained its05:00:15 window. Fresh full-board review found no findings, exact pins matched, and the sole scheduler cleared only this native positive. The guarded command started05:00:35.602734 UTC and stopped05:00:49.611929 UTC:

python3 -B /run/media/kmosoti/data/FabricO11y/results/candidate-build-fresh-8fb6f413179b-01/run-step-browser-announcement-corrective-15.py browser-announcement-native-orca51-01 1073741824 python3 -B tools/ui/announcement_probe.py --mode xvfb --xvfb-receipt /run/media/kmosoti/data/FabricO11y/results/accessibility-xvfb-preparation-01/receipt.json --orca-receipt /run/media/kmosoti/data/FabricO11y/results/accessibility-orca-preparation-02/receipt.json --output /run/media/kmosoti/data/FabricO11y/results/browser-announcement-native-orca51-01

The guard performs the strict namespace DATA census and delegates execution to python3 tools/resource_group.py; no uncapped fallback ran. Admission observed 67,767,996,651 bytes with a fresh1,073,741,824-byte reservation under the95GB stop threshold. Source was clean before and after execution.

Direct exact-source controls, independently of later caller source, ran at this same c3d pin: env PYTHONPATH=tools/ui python3 -B -m unittest -v test_announcement.AnnouncementTests, eight tests/exit0/confirmed stop04:13:20. They include inclusive8MiB PCM,4MiB exact daemon-log and1MiB other-stream boundaries, each+1 rejected; strict marked-SSML and lifecycle refusals remain. Direct documentation check exited0/confirmed stop04:13:28. These source results alone were not substituted for native evidence.

The owned HTTP server served the unchanged167-byte public HTML, SHA256 c217125053d695c0b7e3cb9d72118caea81743fcc43b0118c31bd38b87bcaf1a. Actual AT document attributes for Firefox namespace PID85 include DocURL:http://127.0.0.1:48752/probe. Original role/live/focus prerequisites were true. Monotonic action bounds were207654410112432–207662434949242ns.

Two genuine events from PID85 carried the exact target text: object:children-changed:add:system at207654449500919ns and object:text-changed:insert:system at207654450274517ns. Their role was status bar, with xml-roles:status, live:polite and container-live:polite. Real Orca PID54 returned SSIP225/message14 from moduleespeak-ng between 207654538790300 and207654538911293ns, with this retained raw utterance:

<speak><mark name="0:6"/>Fabric <mark name="7:13"/>public <mark name="14:27"/>accessibility <mark name="28:33"/>probe <mark name="34:42"/>complete</speak>

The strict pre-registered XML extractor preserves the actual interword mark tails as Fabric public accessibility probe complete; it does not drop meaningful characters or accept unknown markup. Genuine backend PID9 mapped libespeak-ng. PCM grew from909,942 to1,072,022 bytes in the action window. All bytes are retained; no sample rate, audio duration or performance claim is inferred.

The actual loaded runtime witness says Orca51.1, module DATA/results/accessibility-orca-preparation-02/extract/usr/lib/python3.15/site-packages/orca/orca_platform.py, executed by installed Python3.14. The privately prepared signed Fedora orca-51.1-1.fc45.noarch RPM SHA256 is 9ebafb81bcb3adce127f4a308ca04215a8e6e590c4953f380928c2552e7b3b8c. Its corrected source files were byte-equal to the official 51.1 archive, SHA256 22829394474e9fde7fdd46b03a98c9a98eb595106b21fb71afdbe2f82436077b. Native admission rechecked the complete prepared regular-file inventory and rejected executable source/schema symlinks. No host installation, patched presentation routine or synthetic speech event was used.

Upstream's status-role correction2f8d804 fixes the UI_UPDATE/live-region ordering and adds a native regression. The retained50.3 run had real status events and PCM but omitted the target utterance. This successful corrected-release execution supports the identified limitation; it is not a benchmark or proof about every Orca/Gecko version. Other prospective cleanup/diagnostic changes are separately registered and must not be hidden in a claim that only one byte/version changed.

Firefox157.0.1 and GeckoDriver0.37.1 remained explicit. Private Xvfb21.1.24 used authenticated UNIX sockets, TCP disabled and GLX disabled after the retained GLX crash. Complete native tool hashes follow:

Actual PID/network/mount/IPC namespaces all differed from the controller. A private user/UTS namespace, DATA-backed /tmp/runtime/config/cache/data, private DBus/AT-SPI registry and file/null ALSA output isolated the fixture. CPU affinity was exactly2-3, CPU quota200%, tasks512, memory high1GiB/max2GiB, no swap. Task peak was337 and pids.events max stayed0. Actual memory.peak was 688,214,016 bytes; all memory/OOM event counters and swap use were0. The full voice inventory was retained; exact PCM8MiB/daemon-log4MiB/other1MiB, 4KiB records and512MiB aggregate bounds were unchanged for this admitted case.

After WebDriver session deletion, dependencies stopped and were actually waited for in reverse launch order. HTTP server/observer/registry exited by TERM(-15); WebDriver, Orca, Speech Dispatcher and Xvfb exited0. All tool commands had exited0. No persistent process had exited early; no forced kill or cleanup error occurred. The namespace leader was reaped, owned configuration/profile/runtime removed, and the independent outer resource receipt confirmed stop. The raw ordered process_stops and exits remain in inside.json; its hash binds these cleanup facts separately from prose.

Every row below is an actual exit1 with confirmed outer stop. Its original output/scratch and inner error remain, rather than being relabelled by51.1. Paths are underDATA/results/; error names describe observed refusal, not a successful negative-control detection.

The original public page, eight-second target oracle and every actual failure remain. No discarded voice tree, presentation patch, DOM semantic relaxation, clock/deadline reset or fabricated event supplied this positive.

The source guard was frozen at c3d, not the later interactive caller module. Full original argv/environment and tools/configuration witnesses remain in the native receipt/inside command list. Prepared GSettings schema SHA256 was a14ebd3855c2dc3a8a033fbf3f59aadae678953d2bc16323025ab0cb35b21485.

The native suppression cell is not run. Actual console caller, keyboard/target inventory and package-bound speech cells remain unrun, requiring successful exact replacement#41, canonical staging, same-package browser positive, merged runtime admission and fresh sole scheduler clearance. Signed-security/key-format,32-row controlled expiry and the rest of the existing browser matrix keep their own prerequisites and receipts. This public-text result closes only the working private AT environment prerequisite; it does not close#45 or supply soak credit.