This is a pipeline-plumbing change: declined-candidate rows in the model packet drop the mechanical dedupe key (used only by code in src/domain/intake.ts, not read by any model) and clip observed/reason/reopenIf to 120/160/120 characters, with every clip explicitly marked ' [… N more]'. Nothing in §3 is loosened: the packet size bound stays at 600,000 rather than being raised to accommodate the overflow, the 'renders within the bound' check and the refuse-to-truncate check remain, and a new test enforces the compaction and the clip markers on every live case. No AGENTS.md edit, no content/ records, no citations added (verification found none), no confidence material, no assessment or ratification machinery touched — the change only shortens what reviewing models see, and marks the shortening honestly rather than hiding it. The DECISIONS entry is stamped with date and channel ('founder (direction in session, 2026-09-20)'), which §3.15 permits for founder directions. One nit I would want reconciled on the record: the PR body says the same packet renders 'under 480,000' while the DECISIONS entry says 'about 498,000', and the body cites 220,000 characters of declined rows where the code comment says 185,000 — harmless rounding in a rationale, not a provenance stamp, so it does not rise to a violation, but the numbers should be made consistent.

The change compacts only the packet’s derived view of declined candidates; it does not alter or erase the underlying disposition records or their histories. Every truncated field receives an explicit omitted-character marker, so the packet does not present clipped text as complete, and the added test enforces that behavior. The fail-closed 600,000-character packet bound remains unchanged, and no citation, ratification gate, provenance-completeness check, or multi-model concurrence check is weakened. The decisions entry records the operational change with a date and supervision channel, and the diff adds no citations or confidential material.

The change compacts declined disposition candidates within ephemeral LLM prompt packets to respect the 600,000-character packet bound while explicitly marking clipped strings. It leaves the underlying case ledger and disposition records intact. The operational direction is properly recorded in docs/DECISIONS.md per §3.15, and no constitutional rules or verification checks are weakened.

The diff only compacts how declined dispositions are projected into the edition packet (drop unused mechanical key; clip observed/reason/reopenIf with explicit “[… N more]” markers), adds a live-case test for those bounds, and records a founder-supervised decisions entry. Full disposition records are untouched; no AGENTS.md edit, no citations, no confidence material, and no check in §3.15 is loosened—the 600k send bound and fail-closed refusal stay. Clip marking and the stated rationale match honest packet accounting rather than hidden truncation or provenance mislabeling.

This is a packet-format change, not a content change: declined intake candidates in the model-facing packet are compacted (observed to 120 chars, reason to 160, reopenIf to 120), the mechanical key is dropped (it is read by code, not by models, per the comment and the test asserting its absence), and every truncation is marked with an explicit '[… N more]' clip marker, so nothing is hidden silently — the honest-labeling discipline of §3.8 is applied to the packet itself. The full disposition records remain in the repository's append-only files, so this does not suppress disagreement or weaken any §3.15 check: the PACKET_MAX_CHARS bound stands unchanged, and the existing 'refuses to truncate' test is preserved alongside the new compaction test. AGENTS.md is untouched; the added DECISIONS.md entry is the permitted decisions-log record with date and supervision stamp (§3.14's change-history discipline), not a constitution edit. The minor numeric discrepancies between the PR body (220,000 chars, <480,000 rendered) and the diff comment/DECISIONS entry (185,000 chars, ~498,000) are packet-size bookkeeping, not provenance, citation, or assessment facts, and rise only to ordinary imperfection. No citations are added, so the verification section is empty by construction.