fix(release): the four corrections 2.5.0 could not be tagged without

[Harol-Reina]

Four things stood between main and a v2.5.0 tag. None is a code-behaviour change; all four are claims the release makes about itself that were untrue, unstated, or stale. Plus one hole that the gate merged in #221 exposed on its first contact with a real CHANGELOG.

1 — DeepgramVoices.Helios: the CHANGELOG mis-stated the cost of the only new obsoletion

The entry read "callers who kept the old name keep compiling". Under TreatWarningsAsErrors — which this repo requires of itself and many consumers run — CS0618 is an error like any other, so they do not. This is the only new [Obsolete] in 2.5.0; the AMI ones all predate v2.4.0, so a consumer already living with those is unaffected.

The alias now carries DiagnosticId = "VSDK0001", giving that build a surgical <NoWarn>VSDK0001</NoWarn> instead of silencing every obsoletion in its tree — and the entry says so.

The trade-off is named, not hidden: a custom DiagnosticId invalidates any existing CS0618 suppression. No downstream consumer pays it — the alias is new in this release, so nobody has one yet. Exactly one tree did: DeepgramVoicesTests pins the alias on purpose, and the build caught it. Its pragma moves with it.

2 & 3 — Two telemetry breaks shipped without a BREAKING heading

• voiceai.sessions.failed stops counting cancelled sessions
• openai_realtime.sessions.failed starts reading non-zero for connect failures that previously had no accounting at all

Both move a number a dashboard or an alert is built on. The prose already called the first "the same family of telemetry break" as the second — the headings now match the prose, and match the eleven entries in this release that already carry BREAKING:.

4 — The package-validation baseline was three releases stale

2.1.0 → 2.4.0. This is the check that would catch a binary-breaking change, so running it against a three-version-old surface is the difference between the guard meaning something and meaning less.

It passes. 29 packages pack at 2.5.0 with the 2.4.0 baseline resolved and no CP**** — a third independent confirmation that 2.5.0 breaks no binary API, and the first against the immediately-preceding release. Two comments citing the old baseline by number now name the property instead, so the next bump does not strand them.

Also adds the ## [2.5.0] - 2026-08-24 heading the release notes are cut from. The date is a forecast, not a record — if the tag slips, it moves.

5 — A hole the new gate exposed on first contact

The notes gate (#221) checked that the CHANGELOG section was not empty. It did not check that it fits. GitHub caps a release body at 125,000 characters and gh release create runs after dotnet nuget push, so an oversized body would publish all 29 packages and then fail to announce them — the exact ordering that gate exists to prevent, reached from the other end.

2.5.0's own section is 117,443 bytes: 93% of the cap. It passes, with a warning that says the next release of this size will not. Measured in bytes rather than characters because wc -m needs a UTF-8 locale the runner does not guarantee, and for this file bytes over-count by 639 — erring toward blocking rather than toward a failed announcement.

Verification

• dotnet build -c Release -p:TreatWarningsAsErrors=true — 0 warnings, 0 errors
• dotnet pack — 29 packages at 2.5.0, ApiCompat resolved the 2.4.0 baseline, no CP****
• Verbara.Sdk.VoiceAi.Tts.Tests — 149/149
• The notes gate's own awk against the new heading — 1467 lines extracted, size guard warns at 93%

Still outside any commit

• A tag ruleset on v* — the actual boundary, since the workflow that runs on a tag push comes from the tagged commit itself
• A publish-liveness gate (Sdk.Pro release.yml:56-86) — 2.5.0 was bumped in fe666f66 and never tagged, which produces not a failing run but no run at all

🤖 Generated with Claude Code