Record the F1 and F2 physical observations against their exact image

[RobVanProd]

The work list still described F1 and F2 as awaiting evidence that has since been collected, so anyone picking it up would either re-run work already done or assume nothing had been tried.

Both notes are dated and name the installed image explicitly, because evidence here is only ever about one binary.

F1 — favourable, not closed

Ten completed turns on d631da56… each ran the full sequence:

response_started → response_end_deferred → reply_pending → response_end_after_playback_complete

with the end gated on real playback_complete rather than estimated speech duration. Ten replies played, no response left open, no 90-second stall.

Recorded explicitly as not closure. The stated gate is a qualification reporting host-response-wire-clean across cancellation, model/TTS failure, and owner-loss. Those error paths were not exercised and no such report exists. F1 stays open, and the note says so.

F2 — the zero-delta evidence it asks for

bridge_uplink_errors stayed at 0 across twelve captures and ten completed turns over 66 minutes, alongside bridge_uplink_queue_failures: 0, mww_uplink_submit_failed: 0, wake_capture_discontinuities: 0, wake_capture_timeouts: 0.

The counter was never reset — it has read zero since the OTA (boot_count: 1, reset_reason: software), which is the specific thing F2 warns against manufacturing.

Ten of ten utterances matched declared byte and chunk totals exactly, longest 217 chunks (10.85 s), with zero audio_count_mismatch and zero orphaned audio_without_utterance.

That last part is worth the extra sentence in the doc: this image also carries the fix for a separate ordering defect found while qualifying it — utterance_end could be queued ahead of a PCM chunk it had already counted, because the socket writer drains queued text before queued binary.

Documentation only. Evidence detail is on #221.