D202 — One file per decision, with a generated flat index

Decision. A decision record is one file, docs/decisions/D<n>-<slug>.md, carrying a topic: field in its frontmatter. The list in docs/decisions.md is generated from those files into a marker-delimited block and verified by internal/repodocs. The ten thematic registers are gone. In code and in prose a decision is referenced by the bare D<n>, resolved with ls docs/decisions/D<n>-*; in markdown it is a real link, and every one of them is checked.

Why. The registers made reading one decision cost the whole topic. 197 records lived in ten files totalling 483 KB; the largest, sync-provisioning.md, was 101 KB, and the median register was around 55 KB. Answering "why is it like this?" meant loading up to 101 KB to reach one paragraph. After the split the median record is 2.1 KB and the largest is 7.3 KB — roughly 26× less for the common case. It also removes a second level of routing: previously docs/decisions.md → topic register → #dNN anchor, where the evidence on progressive disclosure is that one routing level helps and a second one never does and sometimes degrades accuracy.

The bare D<n> reference is deliberate. A path in a code comment breaks when a title is reworded; D47 does not, costs fewer tokens, and ls docs/decisions/D47-* resolves it in one command.

Alternatives rejected.

• Keep the registers and rely on anchors. Anchors do not reduce what a reader loads: reaching #d115 in a 101 KB file still pulls the file.
• Keep the registers and split only the largest. Two conventions in one directory, and the threshold would need defending at every commit.
• Maintain the index by hand. This is the failure mode observed in backstage/backstage, where adding an ADR means editing two configuration files and the two have already diverged. A hand-maintained index does not warn when it is wrong; it just stops being accurate, and then stops being read.
• List every decision in the mkdocs nav. Same divergence, plus ~200 sidebar entries. The nav points at the generated index and the records are marked not_in_nav.
• One directory per topic. That reintroduces the second routing level the split just removed, and topics change more often than decisions do.

Consequences. make decisions-index regenerates the block and CI fails when it is stale, so it is not optional. make decisions-next gives the next number on disk, but an open plan issue reserves its number before any file exists, so the plan list still has to be consulted — that has not changed. Two plans adding two decisions no longer conflict, because they no longer write to the same file; the generated index is the one shared artefact, and it is regenerated rather than merged. The thirteen original ADs stay in a single file (AD1-AD13-original-architecture.md): they were written as one historical table, never had individual bodies, and are all superseded — thirteen files each holding one table row would add navigation without adding anything to read. The old #dNN anchors are dead, and a test rejects any link that still uses one, because an anchor-shaped reference is exactly the kind that survives a split silently. CHANGELOG.md still names old register paths in released entries and is left alone: a changelog is a record of what was said at the time.