D206 — D163 was a copied off-by-one, not a missing record

Status: resolved (2026-09-14).

Context. The gate added by D204 — every D<n> cited in Go source must resolve to a file — found exactly one violation on its first run: D163, referenced eight times across internal/provisioning/provisioning.go, internal/provisioning/provisioning_instructions_test.go, docs/sync.md and D183, always as an established decision and always for the same idea — "D163's metasyntax trap": that a syntax Cartographer recognises inside a file the operator is also expected to document must not fire on text that merely looks like it.

163 is a gap in the numbering, alongside 4, 6, 7 and 11. Nothing was ever filed there. (This paragraph originally counted 130 among them, which was wrong: three records depend on that number, so it got a record of its own — see D130 and D208.)

Decision. The eight references are renumbered to D162, whose second defect and matching alternative are that decision: {{repo:<name>}} written to show the generic form was indistinguishable from a real reference and produced eleven warnings per sync, resolved with the authoritative escape {{\repo:...}} plus a metasyntax heuristic that silences the warning for a key of the form <name> or ... while leaving the text verbatim. internal/provisioning/expand_test.go already labels that work "D162 WP3". No D163 record is written, and the number stays a gap.

Alternatives rejected.

• Write the record from the citations. This was the plan until D162 was read in full. The eight comments describe the trap consistently enough that a plausible D163 could have been written — and it would have duplicated a decision that already exists, leaving two records for one choice and no way for a reader to tell which one the code meant. It is worth stating plainly: the citations were coherent, which is exactly what made the wrong fix look correct.
• Retire the number the way D188 was retired. D188 was retired because nothing depended on it and nothing could be reconstructed. Here the dependency is real and the target exists, so retiring would have left eight comments pointing at a deliberate hole.
• Leave the references and allow-list the number. That is what the gate did for one commit, as the honest interim state. Kept, it would have normalised a reference that resolves to nothing.

Consequences. KnownDanglingDecisions is now empty, and it is checked in both directions: a stale entry whose record exists fails the test, because an allow-list nobody prunes is how a gate quietly stops meaning anything. (D208 replaces that map with two typed ones, GapDecisions and ReservedDecisions, for the reason this paragraph gives: an untyped allow-list cannot distinguish "never existed" from "not yet".) D163 must not be reused — the same reasoning as D188, weaker only in degree: released CHANGELOG.md entries and the git history contain comments that discussed "D163" as the metasyntax decision, and a future D163 about something else would make them read as a description of it.

This is also the first thing the new gates caught, and it is the argument for them in one line: the reference had been wrong for weeks, was copied seven more times, survived a documentation refactor, and nothing anywhere reported it.