docs(openspec): archive externalmedia-returns-a-channel-id-that-finds-its-stream

[Harol-Reina]

Close-out for #305, merged as 9f2cbde7.

Harvested: F1 (WebSocketAudioServer's URL-segment key), F2 (AudioStreamMetrics' ten uncalled instruments) and F3 (ten dialed-but-undefined dialplan extensions) → openspec/changes/a-published-surface-is-one-something-measures, opened during the change itself rather than at close-out. Its dependency line is updated here to read as satisfied and to name where the archived record and the live capability are.

Cited: (#305) backfilled into the CHANGELOG.md [Unreleased] entry.

Referrer sweep: two references, both by change name rather than by path, so the archive move breaks neither — ExternalMediaChannelIdFunctionalTests.cs:11 and the new change's proposal. The latter's tense was stale and is corrected; nothing else pointed at the directory.

One ruling, not a tidy-up

The entry is relabelled from ### Fixed — BREAKING to ### Changed — BREAKING.

ADR-0061 landed as #307 after this entry was written: only a Changed — BREAKING forces a minor, and D3 says an entry that cannot name the documented behaviour it restores is a Changed.

This entry has two halves:

• The behavioural half restores what the docs promised — GetStream said it returned a stream by channel id, and could not. That is a Fixed.
• The signature half changes a shipped public member. A caller that passed the cancellation token positionally stops compiling. That restores nothing.

Left as Fixed, D1 would have permitted the next release to ship that compile break as a patch. D3 puts the call on the author at the moment the entry is written; this is that call, made late because the ADR arrived in between.

The spec's Purpose is written, not left at TBD

It records that Asterisk decides this, not the SDK: data becomes the identification UUID the stream table is keyed by, channelId becomes the ARI Channel.Id, and supplying neither — the shipped behaviour — leaves a channel id that appears in no table, so the lookup cannot succeed by construction.

And it records why the answer had to be measured with two byte-order-asymmetric UUIDs per request: reading the code yields a plausible answer, and the review that tried it produced a wrong one.

Verification

openspec validate --all --strict : 15 passed, 0 failed
Governance tests                 : 129 passed, 0 failed

🤖 Generated with Claude Code