docs(claims): count the AMI surface by type, and correct every published figure

[Harol-Reina]

The problem

README.md:45 published 148 actions, 278 events, 18 typed responses. Those three numbers come
from two different ways of counting: 148 counts types, 278 and 18 count files. No single definition
makes all three true, which is why docs/claim-registry.md has carried them as an unresolved ruling
— a guard cannot be written against a figure that does not come from a countable rule.

Meanwhile the same surface was published three more times, each differently:

Where	Said	Correct
README.md:45	148 / 278 / 18	148 / 269 / 17
README.md:65	278 events	269 events
docs/README-technical.md:543-544	111 actions, 215 events	148 actions, 269 events
src/Verbara.Sdk.Ami/README.md:7	111 / 261 / 17	148 / 269 / 17

The last one is the package's PackageReadmeFile — published verbatim on nuget.org.

The ruling

Count by type, using the definition the shipping source generator already implements: a
class-level [VerbaraMapping] on a non-abstract type (EventRegistryGenerator.cs:51,60). Measured
against the tree: 148 / 269 / 17.

Two alternatives were rejected on what each would have to count as an AMI action or event:

• By file (149 / 278 / 18) counts Actions/IEventGeneratingAction.cs — an interface — as an
  action, and Responses/ConfigCategory.cs — a helper record — as a typed response.
• By [VerbaraMapping] occurrence (149 / 270 / 17) counts two property-level mappings as
  types: Async in OriginateAction.cs:19 and 100rel in EndpointDetail.cs:74. Both map an AMI
  field name that is not a valid C# identifier, which is why the attribute is on them at all.

The nine event base types (Events/ResponseEvent.cs plus the eight under Events/Base/) carry no
[VerbaraMapping], so no definition reaches them. None is abstract either, so a filter written on
IsAbstract alone would not have excluded them.

Also corrected

docs/README-technical.md:543-544 named the types AmiAction and AmiEvent, in
Verbara.Sdk.Ami.Actions and .Events. Neither type exists. The bases are ManagerAction and
ManagerEvent, both in Verbara.Sdk (src/Verbara.Sdk/IAmiConnection.cs:70,79). Same line, same
class of defect: a false claim about the API.

Left verbatim, deliberately

docs/decisions/0001 and 0015 state 278 event types, and dated files under docs/research/ state
278 and 111. Both folders are Out of the claim registry's scope as period-correct records, and an
Accepted ADR is superseded, never edited.

No guard ships here

A guard on src/*/README.md cannot run today. scripts/ci/classify-docs-only.sh:29 treats every
*/README.md as docs-only, so a PR whose only change is the package README skips Unit Tests — the
job such a guard would live on. Demonstrated rather than assumed: a throwaway commit putting
999 actions, 999 events in src/Verbara.Sdk.Ami/README.md and nothing else classifies
docs_only=true.

The guard would therefore fire on the PRs that do not need it and stay silent on exactly the ones
that break it. The script's own comment records this as an accepted residual
(verbara-meta/ADR-0016 §6.1 risk 3), so turning it around is a CI decision of its own, not a test to
add — it moves with the registry's remaining open ruling. These rows stay GAP: the figures are
right and nothing yet stops them drifting again.

This PR itself classifies docs_only=false (it touches README.md, which the fast path excludes),
so the unit lane runs.

Verification

• openspec validate --all --strict — 13 passed, 0 failed
• Tests/Verbara.Sdk.OpenTelemetry.Tests — 31 passed (the claim-guard suites: MarketingClaimsTests, StatusBlockCoherenceTests, PerformanceTableCoherenceTests)
• Tests/Verbara.Sdk.Governance.Tests — 129 passed
• No test asserts any of the old figures.

🤖 Generated with Claude Code