ADR 0310: Amendment notes consolidate into one timeless block, and supersession hygiene is gate-enforced

Status: accepted. Applies the forcing-function discipline of ADR 0051 to the decision archive itself, joining the number-uniqueness guard of ADR 0186; the currency audit that motivated it moved ADR 0227 and ADR 0066.

Agents read the decision archive as instruction material, so a current record carrying a stale claim is misinformation delivered with the archive's authority. A corpus-wide currency audit found the rot concentrated in three shapes.

Amendment stacks. A record attracts one leading blockquote per later amendment. The heavily-amended records open with several consecutive multi-paragraph banners, and a reader must merge them mentally before the record itself begins.

Drift-prone specifics. Notes that restate a count or a version go false the next time the number moves — the audit found a note asserting a bundled-set size the set had since outgrown, dependency pins behind the lockfile, a renamed config key and environment variable presented as current, and a Git-admin path that no longer exists. A note exists to keep a record honest; a note that itself rots inverts the purpose.

One-way supersession. Later records declared they superseded or amended an earlier clause while the earlier record carried no matching note — and one record whose own status declared full supersession still sat among the current records.

One amendment block per record. All amendment notes on a record live in a single leading blockquote, one entry per amending decision, each citing its ADR and stating the delta in a sentence or two. Shared boilerplate is stated once for the block, not repeated per entry.

Notes read timelessly. An amendment note states the direction of a change, never a re-stateable current value: no counts a later change would falsify, no version pins, no "is now N". Where the current value matters, the note names the live authority — the config, a registry, the lockfile — instead of restating it.

Supersession is recorded on both records in the same change. A record that supersedes, amends, or narrows an earlier clause adds the matching note to the earlier record as it lands. A record whose whole decision falls moves to _superseded/, with a banner naming its successor and whatever survives.

A guard enforces the mechanical half. The gate fails when a current record's status declares it superseded, when a status line deviates from the **Status**: form, and when a current record claims to supersede a sibling current record that carries no reference back. The editorial half — consolidation, timelessness — stays with the writer; the README states it as the format.

• Heavily-amended records regain a readable opening: one block to absorb, then the record.
• Amendment notes stop being a second source of drift. The price is one indirection when a reader wants the current value, paid at the named authority.
• The guard catches only the checkable failures. It cannot check timelessness — a numeric claim can still rot — so the convention and review carry that half.
• Dated notes lose some point-in-time color; Git history retains it.

• Keep stacking one banner per amendment. Rejected: the stacks had already grown past readability on the most-amended records, and each banner repeated the boilerplate of its siblings.
• Edit records in place instead of annotating. Rejected: a revisited decision gets a new record, never a rewritten old one — the history of the decision must stay legible.
• Enforce timelessness mechanically. Rejected: a denylist of numbers would misfire on the many counts that are the decision itself. The guard checks structure; prose stays a writing discipline.