companions: the PR #68 repairs, applied against the ratified 4.2 edition

[dhh1128]

Applies the repairs from the review on #68. No ratified bytes are touched; verify_kernel.py is 7/7.

Why this is a separate PR

#68 merged on 2026-08-14 without the repairs. The merged blob is byte-identical to the branch tip 76f068e — git diff 76f068e:companions/custos-for-keri-practitioners.md main:... is empty — so every defect the review raised on 2026-08-10 has been live in main since. This is that repair, arriving late.

One correction to the review's own tally, in its favor. PR68-3 (stale-base README regression) never happened. The merge diff on companions/README.md adds two lines and removes none: all six sibling rows and the owed notation-register paragraph survive. Either the branch was rebased before merge or GitHub resolved it cleanly. Nothing to repair there.

The two live MAJORs

PR68-1, stale edition. The closing section read "The most instructive section of Custos 4.1… Six commitments are fixed." The ratified edition is 4.2, and §16 enumerates seven (spec/custos-4.2.md:2886, "Seven commitments bound the governance evaluator"), in this edition's own text and with keyword force over companions. Retargeted the section and the README status cell to 4.2.

PR68-2, byte-identity overclaim. Line 165 read "Two evaluations of the same triple return byte-identical findings." The ratified text expressly declines this today at spec/custos-4.2.md:1632: findings SHALL be equal "under the one conformance predicate of section 17: semantic full-payload equality today, byte identity by construction the moment a carriage encoding ratifies." Restated in that two-grade form. This is the sentence the whole chess-scoresheet argument lands on, so it carries the qualification rather than dropping the claim.

The review was right that line 217 ("the same defeat down to the byte") is a different claim and is sound — it rides the canonical-selection SHALL at spec/custos-4.2.md:1769. Untouched.

PR68-6, and a count defect the review didn't catch

The companion's README row sat behind a blank line, which severed it from the table above; it rendered as its own stray one-row table. Blank line removed.

While there: companions/README.md:3 read "Four companions travel beside the ratified kernel" and line 7 "Two of the four". That was already wrong before #68 — the table held six rows — and #68 made it seven. Both counts corrected. Flagging it as pre-existing rather than folding it in silently.

Minors

• PR68-5. The terminality sketch said "affirmed never becomes defeated" with no exit. spec/custos-4.2.md:1712 states the one event that is an exception: a contradictory pair bearing on the same question moves either to self-convicted. One sentence added, since this audience implements from that sketch.
• PR68-7. The self-convicted gloss carried no key-tier caveat. spec/custos-4.2.md:1531: at the key tier, whether a pair bears is decided by KERI's own superseding-recovery rules, and a lawfully superseding event reconciles rather than convicts. This is the first question a Kever-holding reader asks, so it belongs in the gloss.
• PR68-8. "GLEIF's root identifier is adopted grade, exactly" now points at the GLEIF EGF mapping, which is where that reading is worked out. The kernel doesn't ground it.
• PR68-4. Under the 2026-08-10 register amendment, "the price is sovereignty" was the one hit left live (rhetorical class). The sentence no longer frames the tradeoff as a purchase. "Budget" at line 108 is left alone — it reads to me as the resource sense, but say the word and it goes.

Not done

PR68-9 and the analogy leak are not addressed. PR68-9 was a NOTE on simplifying the lexicographic sentence, and the review's closing observation was that the chess figure never confesses its second boundary — chess assumes a single, integral, complete scoresheet, with no duplicity and no completeness surface. Both are judgment calls about the author's own figure, and I'd rather not make them on his behalf in a repair PR. They're worth a follow-up.

The review report the comment cites is not in the repo. reviews/rounds/pr68-companion-review-2026-08-10/review.md doesn't exist on main. The findings survive in the PR comment on #68, but the report itself was never committed, so the evidence set has a hole where the other rounds have records.

Relates to #68.