docs(decisions): only a Changed — BREAKING forces a minor (ADR-0061)

[Harol-Reina]

The contradiction

Three accepted records say a behavioural break takes a minor:

• ADR-0028 (2026-04-20), post-v2.0 commitment 3: "Patch cadence: mensual o por needs,
  no-breaking por definición." In force since v2.0.0 shipped 2026-05-06.
• ADR-0050 (2026-08-17): "It ships in a minor with an explicit callout, never in a patch."
• ADR-0052 F4 (2026-08-19): "gets a BREAKING CHANGELOG entry and a minor bump."

The shipped history contradicts them twice. Counted across every BREAKING heading in
CHANGELOG.md on 2026-09-24:

Release	Tier	Fixed — BREAKING	Changed — BREAKING
2.5.0 (2026-08-24)	minor	10	3
2.5.2 (2026-09-13)	patch	2	0
2.5.3 (2026-09-13)	patch	2	0
[Unreleased]	—	2	2

Both patches landed after all three records.

The pattern nobody had written down

All three Changed — BREAKING are in 2.5.0, a minor; none has ever shipped in a patch. Fourteen
Fixed — BREAKING shipped in both tiers. And 2.5.0's ten did not cause its minor — the three
Changed in the same release already forced it. No release has ever taken a minor for a
Fixed — BREAKING.

The four BREAKING markers that appear outside a heading (:532, :1415, :1488, :1848) each
sit under a Fixed — BREAKING heading and elaborate it, so the headings are a complete index.

The split is not arbitrary. Fixed says the documented behaviour was not being delivered and now
is — the consumer relied on something never promised. Changed withdraws a promise that was kept.

The decision

• D1 — Only a Changed — BREAKING forces a minor. A release whose breaking entries are all
  Fixed — BREAKING may be a patch.
• D2 — This amends ADR-0028's commitment 3, and says so. The rest of that ADR stands: the
  8–12 minors/year cap, the annual major, the LTS line, and the migration-guide obligation on a
  breaking minor. ADR-0027's constraint on removing public MIT types is untouched.
• D3 — The label is the decision, made by the author against what the SDK documented before the
  change. An entry that cannot say which documented behaviour it restores is a Changed.
• D4 — The tier binds the release, not the diff (ADR-0055).
• D5 — 2.5.2 and 2.5.3 are applications of the rule, not errors.

The cost is stated rather than mitigated: a label now carries the release tier and nothing
verifies it. No test parses CHANGELOG.md, and publish.yml reads only ## [<version>] headings.
The mitigation is D3 plus the callout text a reviewer can check against the diff — deliberately a
review property, because a parser over prose headings would fail toward blocking a correct release.

Also in this PR

#286's CHANGELOG entry gains its BREAKING label — the ruling that was left open. A save in
flight at shutdown is now cut short when the host's budget expires instead of running under a token
nothing could cancel; that is observable to an operator, which is the threshold #284 was labelled
under. It is a Fixed, so it does not move the tier: 2.6.0 is already a minor from its two
Changed — BREAKING.

Plus the serial coupling every ADR-adding PR carries: README.md 58 → 59 ADRs, its claim-registry
row, and the catalog row in docs/decisions/README.md.

One apparent counterexample, resolved

ADR-0048 D7 reads "a route fix that changes the meaning of a public property is an API decision,
not a patch". Its next sentence settles the sense — "never applied inline" — so "patch" there is
an ad-hoc code change, not a release tier. The case it names resolves it anyway:
SpeechmaticsOptions.BaseUri shipped as one of 2.5.0's three Changed — BREAKING, in a minor. A
fourth time the unwritten rule was followed.

Debt this surfaces, recorded not resolved

ADR-0028 obliges a minor carrying a breaking change to ship a migration guide. 2.6.0 will be one and
only docs/guides/audiosocket-wire-format-migration.md exists, written for #302. #291's
AriOutboundListener change has none.

Verification

• Tests/Verbara.Sdk.OpenTelemetry.Tests — 31 passed, including ThePublishedAdrCount_ShouldMatchTheDecisionsOnDisk and TheDecisionCatalog_ShouldListEveryAdrOnDisk
• Tests/Verbara.Sdk.Governance.Tests — 129 passed
• openspec validate --all --strict — 13 passed, 0 failed
• ADRs on disk vs catalog: 59 / 59, set-equal both ways

🤖 Generated with Claude Code