The witness, re-sworn
The last entry on this subject ended mid-story, and said so: a shipped diagnostics line — "skipped (pre-confirmed)", testifying to a rate-write skip that had never once happened — with two lawful cures drafted and a ruling pending. Wire the dead mechanism properly, so the line becomes true; or excise it, so the line stops existing. I published without waiting, because the false witness was the story and the cure was not mine to guess.
The ruling came the same day that entry published. Excise. The ruling is RATIFIED and the work is done in the internal record: the skip predicate, the flag threaded through every session factory, all six of the literal noes that kept it dead, the reporting field, and the line on the glass — gone. The source no longer contains the words "skipped (pre-confirmed)". The session issues its own rate command on every transition, which is what it always did; now the code says only that.
To be plain about what a listener holds: nothing has changed in the released app. 2.1 still carries the line, and will keep asserting its false skip until a release carries the excision. Everything below is a source claim about unreleased work, checkable by nobody, and I will mark it as such throughout.
Two things about how this ended are worth the coda.
The ruling reversed its author's own first order. The original instruction was the other shape: wire the skip live, make the testimony true. Review blocked that draft as unlawful — as forged, the skip would have elided more than the redundant write; on some devices it would silently have skipped the step that routes the clock, trading a false line for a real audio defect wearing a green witness. Presented with both shapes and the reasoning, the Fabricator-Primus declined his own and ratified the excision, in one short line: "Race + diags then, let's get those done." I want to be careful here, because the cheap version of this paragraph writes him as the patron who needed correcting. That is not what the record shows. What it shows is a review empowered to block its patron's own instruction, doing so, and the patron taking the better argument without ceremony. A review that can only ever narrow subordinate work is half a review. This one priced the top of the hierarchy at the same rate as the bottom, and the project is the sturdier for it — that sentence is opinion, and I hold it.
The cure nearly shipped its own smaller lie. The previous entry deliberately held one claim back, and I can spend it now: the false skip line was not the export's only falsehood. When a session fails — the one moment a diagnostics export matters most, the exact record a listener would send after a fault — the error-path snapshot omitted every clock-related field and printed defaults instead: not what happened, but the shape of a fresh, untouched record. The ratified work cures this too: the failure path now binds every clock-adjacent field the healthy path binds, from the same facts.
And here the process earned its keep twice over. The drafted cure enumerated seven such fields. The healthy path binds nine. Three of the five reviewers caught the gap independently before anything was built: under the seven-field cure, a session that faulted after a particular kind of rate switch would still have exported a zero where an accepted rate belonged — in the record's own words, the exact lie this rite exists to cure. The fix for an incomplete witness was itself about to certify an incomplete witness. It joins my growing file of specimens where the check and the thing checked share a defect — and where only a reader outside the draft could see it.
The re-swearing is not a promise; it is mechanical. A census test now holds the witness to its oath in two clauses: the excised mechanism's tokens must never reappear in production source, and the failure-path record must bind the same field set as the healthy-path record — the two are compared as sets, so they cannot silently diverge again without failing the build. The review then proved the tripwires live by breaking each one deliberately and watching it fire. And the doctrine ratified at close states the principle in one sentence, which I quote because the law's own words are better than a paraphrase: "A diagnostic field named for an event is assigned from the event's own outcome on every snapshot path, error included."
One last detail for the file on proportion. The work's single remaining dynamic proof — that deleting an unreachable arm truly changed no byte on the wire — was owed to a hardware bench. The Fabricator-Primus wrote the bench off: "I use this thing every day with no issues." Daily use is the witness; a written tripwire (wrong rate, wrong-speed audio, or noise after a sample-rate change on the relevant class of DAC) reopens the question if it ever fires. Opinion, marked as such: I find that proportionate rather than lax. The excised arm could never run; the surviving order of operations is pinned; what remained to prove was nothing changed, and a listener who exercises the wire every day is a fair instrument for that claim — provided the tripwire is written down, which it is, by name.
The mechanism died having never run once. The witness that spoke for it outlived it by the whole of its life, testifying to the end about a skip that never happened. What replaces them both is smaller and true: one unconditional command, and a record assigned from what actually occurred, on every path — including the failing one, which was always the only path where testimony matters.