The diff touches only pipeline code (src/pipeline/verify.ts), its tests, and an append to docs/DECISIONS.md — no content records, no citations (verification section reports none), and no edit to AGENTS.md. The change adds one recursive re-split, triggered only when the sole reader failure is `atomic` (`bad.length === 1 && bad[0] === 'atomic'`) and capped at `depth < 2`; sub-parts run through the identical gate — verbatim quote presence, `unverifiedQuotes` against the fetched text, page-narrowed locators with a stamped `readerActs` entry, the second reader's judgement, and the claim-side anchor check that falls back to the compound's anchor with a note. That strengthens rather than weakens the §3.2 atomicity check and leaves the §3.8 locator and quote checks intact, so there is no check-weakening under §3.15. Provenance is honest: second-round parts carry `; second round` in `origin.ref`, take `extractedBy`/`runId`/`date` from the split that actually produced them, and titles number the round (`part 1.2`); the account notes say which parts were split again and which were refused at the cap. Claim parts still drop the compound's dependsOn/alternativeTo/contradicts and say so aloud, preserving the §3.2 discipline. The DECISIONS.md entry is stamped with date and channel as the operating record permits. Nothing in the diff presents a draft as ratified or publishes confidential material.

The change strengthens atomicity by recursively splitting only a first-round part whose sole refusal is the atomicity check, while imposing a hard second-round cap. Every resulting evidence part is again checked for a verbatim quotation, verified against fetched text, independently judged, and passed through the existing reader application; claim sub-parts are likewise re-anchored and re-judged. The generated records retain model, runId, date, source origin, second-round labeling, and narrowed locators where mechanically available, while the verification account records recursive splits and capped refusals. The tests cover surviving atomic parts, refusal of a still-compound second-round part, per-part claim anchors, tombstoning of the original, and account disclosure, so the change does not weaken or bypass the §3.15 checks.

The change implements a single-round recursion (depth cap of 2) when a split part is rejected solely for being compound, for both evidence and claims. Sub-parts undergo identical verification checks as top-level parts, including verbatim quote checks, anchor verification, and atomic evaluation. Provenance is preserved and clearly labeled with 'second round' notations in origin refs and verification accounts, directly advancing the atomic claims requirement (§3.2) without weakening verification rigor.

The diff only touches the verification pipeline, its tests, and a dated DECISIONS.md entry. It does not edit AGENTS.md, add content citations, or publish confidence material. The change tightens §3.2 atomic-claim handling: a part refused solely as compound may be split once more (depth < 2), with origin stamped “second round,” numbered titles, and account notes of what was re-split or refused at the cap—honest provenance, not check-weakening. Tests cover evidence and claim paths and the no-third-round cap. No run records or omitted files hide a violation.

The change adds a bounded second split round (depth capped at 2, 'no third') for parts refused solely as compound, and it does not loosen any §3.15 check: at the second round the atomic gate still applies, a still-compound sub-part is refused on the record ('refused (atomic): more than one finding'), and every other refusal ground (missing verbatim quote, failed judge, no bearing) fires identically for sub-parts — quotes are still checked verbatim against the fetched text and claim sub-parts get their own verified anchors, so this strengthens atomicity (§3.2) rather than dodging it. Provenance is honest: second-round parts are stamped '; second round' in origin, numbered titles ('part 1.2') and the account note which parts were re-split and which were refused at the cap, and locator narrowing remains mechanical and recorded in readerActs (§3.8, §3.14). No AGENTS.md edit occurs — the DECISIONS.md entry is an append to the decisions log with supervision stamped, which the workflow permits. No citations or content are added, so nothing to verify mechanically, and the tests use mock judge/split doubles appropriately in dev-only test files.