fix(media): name a station that could not be reached, and one that withheld its key

[lenoxys]

Closes the last link for #1089: the SDK now raises both refusals as their own types, and this is what makes them arrive named in an archive.

Why

sourceFailure put both in source-error — the bucket for a source that failed. So a base that never connected, and a base that would not seal what it was sent, both read as the camera a viewer happened to open. On #1089 every camera behind one base failed identically and the reports named the cameras, which is where the investigation went.

Neither is the camera:

• station-unreachable — every camera behind an unreachable base fails together, so the next step is that base or the network to it.
• station-key-unavailable — an attached camera's media start has no unencrypted form, so a missing session key is a refusal rather than a fault, and retrying the same camera does not resolve it.

Why the key was missing is the SDK's own trace vocabulary — level2-unavailable with no-cipher-key, derivation-failed, not-negotiating, session-closed or grace-elapsed, plus cipher-fallback and station-resolved — which this plugin already retains. This adds the outcome the session ended on, which is the part a reason vocabulary owns.

What changed

• Both reasons on LiveSessionFailure and RecordingFailure, mapped in both sourceFailure functions, and admitted to the diagnostics reason allowlist so they reach an archive.
• The stage stays sdk-source-acquisition, because that is where the session ended.
• SDK pin moves to the published 0.2.0-beta.10, which is what exports the two types.

Verified

• npm run verify green: prettier, ECS guard, tsc, 893 tests / 50 files.
• npm run qualify:release clean against the exact pin — 273 packed files, no dist-tag, no symlinked install.
• Contracts pin both by SDK type, on live and on recording, and keep the existing case that an error merely calling itself StationBusyError stays a source failure. A name is a string any thrower can write.