fix(prime-agent): adopt 0.7.3 and explain a rejected model selection

[rynfar]

Problem

Prime Agent shipped 0.7.3 and we were documented and verified against 0.7.2. Two behavior changes reach users:

• 0.7.3 regenerated its model catalog and dropped model ids (cerebras/zai-glm-4.7, claude-opus-4, claude-opus-4-1, several claude-3*). A thread pinned to a removed model fails its next turn with The daemon model switch failed., which does not tell the user what happened or what to do.
• 0.7.3 moved the supervisor ownership registry from $TMPDIR to ~/.prime/supervisor-owners (macOS cleanup was deleting long-running supervisors' authority records). Pylon now writes ownership records there even when a provider instance sets its own agent home, which makes the parity ledger's claim that the real-daemon test "never writes to the source Prime home" untrue.

How it was fixed

No adapter change was needed for the bump. I diffed the published 0.7.2 and 0.7.3 tarballs: daemon-agent-connection.d.ts, daemon-protocol.d.ts, and index.d.ts are byte-identical, protocol 7 / schema 16 / schema id are unchanged, and context-tree.js and daemon-socket.js — the code behind our two blocked parity entries (getContextTree, Windows daemon mode) — are unchanged too, so every recorded deferral still holds. 0.7.3's rewritten RLM spawn ledger is private to daemon mode, and its new authenticated host-request contracts belong to the Python kernel surface, not the extension API our approval gate uses. The one public removal in the release, listSavedSessionSiblings/siblings(), is unused here.

Model rejection now names the likely cause and the action, without copying Prime's native error text across the adapter boundary. Covered by a focused test that also asserts the native path in the rejection does not leak.

Ownership registry: keep the upstream behavior, document it. It is safe for us because ownerConflicts keys on socket path and worker descriptor directory — Pylon derives a unique socket per (state directory, provider instance) and the descriptor directory hashes that socket, so a Pylon daemon never claims ownership over a user's interactive prime-agent daemon even when both share the default ~/.prime/agent home. Prime exposes an internal env override for the location; we do not use it, because sanitizePrimeAgentDaemonEnvironment strips every PRIME_AGENT_INTERNAL_* variable by design and that policy is explicitly tested, including a forward-compatibility case. The ledger's test-isolation claim is corrected rather than papered over.

Verification

• vp test run apps/server/src/provider/prime/ — 243 passed, 1 skipped (was 242 + 1; the new test is the delta)
• Opt-in real-daemon test against the installed 0.7.3 binary — passed (full native turn, cancellation, catalog/model selection, manager restart, resume, socket cleanup)
• vp run -F t3 typecheck — 0 errors, nothing reported in the touched files
• vp lint on both touched source files — clean

No UI changes, so no before/after images.

Model: Claude Opus 5 in Pylon (Claude Code harness)

---

^{Need help on this PR? Tag @codesmith with what you need. Autofix is disabled.}

[github-actions]

Thread transfer impact

⚠️ The latest CI run did not produce a thread transfer result for 64e1710.

This comment will update automatically after the next completed run.