Fix nightly inference, extension, VOPR, and CI cache regressions

[ajroetker]

Fix the code and test failures exposed by nightly run 35511643401, following the merge of #821.

• Preserve native inference prefill-plan ownership across isolated execution, materialize source-backed broadcast inputs, and reject mixed extraction inputs before model lookup.
• Avoid reading undefined stateless allocator context pointers when reusing PDF render lanes. Match Wasmtime v45's nested C handle padding.
• Bind extension members to durable catalog table identities in standalone and distributed mode, preserving bindings across version upgrades. Authenticate host reads/writes against public scopes and reject renamed or reused names that no longer resolve to the installed identity. Report binding failures at the MCP boundary even when a guest swallows the host error. Update lifecycle fixtures to project committed v2 mutations.
• Restore VOPR test imports and current storage fields; verify resource pressure through the actual admission response and rejection counter. Give the production baseline the existing 240k-transition witness budget: the pinned failing schedule stopped at 1.756 virtual seconds under 30k, then completed and exact-replayed with only its budget extended. All correctness, completion, and cleanup assertions remain required.
• Retain cached test executables during CI pruning, deleting only disposable linker inputs. Exercise cache hits and source rebuilds with LLVM, which emits disposable objects on Linux; separately verify that pruning preserves self-hosted executables without linker inputs.
• Download the diarization speaker dependency when needed and give Gemma thinking-mode fixtures sufficient output tokens.

Validation: targeted inference unit tests; PDF ReleaseSafe tests and OCR E2E; 38 extension/ABI tests and all five extension E2E tests with Wasmtime v45; 21 targeted inference E2E cases with CI memory budgets; clean speaker-model download; model-helper regressions and six cache-pruning tests; Linux cross-build verification of self-hosted and LLVM artifact layouts; storage/provider/full-cluster VOPR and resource-pressure replay; secret-backend tests. The final production VOPR ReleaseSafe smoke gate passed: 18/18 build steps and both 15-test suites, including baseline and bounded-lifecycle exact replay.

The build-cache job separately lost contact with its runner. Its logs were unavailable; this PR does not claim a code fix for that infrastructure failure.

The rename/name-reuse regression failed against the previous binary in both standalone and distributed modes, then passed with the fix. The production ReleaseSafe server build passed. CI requested for f811469b0547ae3560699ba20a78e532104f8ca7.

[ajroetker]

/ci run 4551c51

[ajroetker]

/ci run f811469