This is a tooling-only change to `src/lib/arbiter-core.mjs`, the arbiter script, and tests, plus an append-only DECISIONS.md entry; AGENTS.md is untouched and no content/ records or citations are added (verification section finds no identifiers). The effect on the §3.15 checks is strengthening, not weakening: the account cap rises 150k→250k, the head edition and the assessment it adopts are now unconditionally first, never dropped and never cut mid-way, an added/modified record digest lets seats check atomicity, direction and locators for files the diff budget drops, modified records name changed fields with new values, and superseded assessments retain their reasoning with differing verdicts marked against the head's — all mechanical extraction, with the header stating every cap and every drop named. The OMITTED FILES annotation via `omittedNotes` marks which omitted files were read into the account rather than claiming they were shown in the diff, so it does not misrepresent what a seat saw. Reducing a superseded edition to its non-shared paragraphs and giving up claim reasoning past 30,000 characters are stated in place, so nothing is silently hidden (§3.9, §3.14 provenance stays visible). The only blemish I see is a bookkeeping inconsistency: the DECISIONS.md entry describes a fourth round and review note #379 while the PR body describes three rounds and notes #377/#378 — a narrative mismatch about the operating record, not an invented citation or provenance stamp, and the code comment for #379 (reading an adopted-but-unchanged assessment) is in fact present. That is ordinary imperfection; I would ask the operator to reconcile the round count, but it does not park the change.

The generated account can misstate what was retained and how the review packet was assembled. Every assessment that is not classified as superseded is assigned rank 0, including check-role assessments and assessments in cases with no changed head edition, so these sections can never be dropped; nevertheless, the overflow notice says that only “head edition(s) and the assessment(s) they adopt” remain and that “nothing else is included.” The header likewise does not disclose that these unrelated assessments share the protected rank, which can materially change both account size and the diff budget available to the panel. Fix this by either assigning non-adopted assessments an accurately documented droppable rank or describing all protected section classes truthfully in the header and overflow notice.

This change improves the tooling for generating the run account passed to the arbiter panel when PR diffs exceed size budgets. By prioritizing the head edition and adopted assessment, digesting added or modified records from evidence and claim ledgers, and annotating omitted files that were read into the account, it enables panel seats to review atomic claims, locators, and assessments rather than forcing 'unsure' votes due to diff budget truncation. No epistemic checks in §3.15 are weakened, AGENTS.md is untouched, decisions are logged in DECISIONS.md, and comprehensive unit tests are provided.

The diff changes only arbiter tooling (runAccount assembly, omitted-file notes, tests) and appends a dated founder-supervised entry to docs/DECISIONS.md; AGENTS.md is untouched. No content citations, sources, or ledger claims are added, so fabrication and confidence rules are not in play. The change strengthens §3.15 reviewability: head edition and adopted assessment are rank-0 and never dropped, records are digested with explicit clips and modified-field values, superseded material keeps its own wording, and OMITTED FILES marks what the account already carries—addressing prior unsure parks without loosening any check. Caps, drop order, and provenance wording are stated in the generated header and covered by tests.

This change edits only the panel-packet tooling (arbiter scripts, tests, and an appended DECISIONS.md entry), not AGENTS.md, and adds no citations or content records, so §3.15's constitution-edit and fabrication rules are not implicated. Far from weakening the §3.15 checks, it strengthens them: the run account now carries the head edition and adopted assessment whole, digests every added or modified canon record (surfacing, per the PR body, a by-hand correction to VASO-R012 that the old diff budget hid), keeps every claim verdict even when reasoning must give way, names each dropped section, and never silently cuts the head — all of which directly addresses the failure mode (three seats voting unsure on #372) that previously parked legitimate work. The new digest's honest labeling of what is whole versus clipped, the marked OMITTED FILES annotations, and the stated 250,000-character cap with per-field clip marks preserve the provenance-transparency requirements (§3.8, §3.14); the DECISIONS.md entry is append-only and candid about four rounds of review corrections. The founder-supervision stamp in DECISIONS.md is an operational record I cannot independently verify, but nothing in the material contradicts it, and the change itself is exactly the kind of check-hardening the constitution favors. No rule broken; routine imperfection is absent even in style — tests cover the digest, ordering, superseded handling, and the cap invariant.