Add external runtime evidence ingestion

[200lz]

Summary

• add versioned external runtime-observation control and evidence schemas
• add runtime-compat import-external and runtime-compat verify-external
• add bounded, streaming manifest ingestion and deterministic offline reconstruction
• keep process completion separate from reviewed output predicates and OMIV stage status
• enforce retry, ownership, artifact-grammar, limit, and privacy invariants
• document the external-evidence workflow and add adversarial regression coverage

Why

This lets OMIV ingest independently captured llama.cpp GPU/runtime evidence without trusting runner-authored PASS text, copying raw operational material into Git, or upgrading unsupported claims. It turns the Muse Glimmer A5 run into portable, verifiable evidence while preserving fail-closed semantics.

Muse Glimmer A5 result

The imported observation remains honestly PARTIAL:

• five process attempts are INCOMPLETE
• all five runtime stages remain UNKNOWN
• the fixed PNG predicate is OBSERVED, not PASS
• three artifacts are MATCHED_PLAN_OBSERVATION with zero payloads verified
• DFlash records 23/27 accepted draft tokens
• DFlash PNG and Ollama remain NOT_RUN

No source-to-GGUF binding, publisher/origin authenticity, fidelity, performance, safety, production-readiness, or OMIV stage-PASS claim is added.

Validation

• full hermetic suite: 2,244 passed, 11 expected skips
• external runtime-observation suite: 276 passed, 1 fixture skip
• final independent review: no actionable P0/P1/P2 findings
• online public-release-readiness audit: PASS, 22/22
• Ruff, strict mypy, compileall, and git diff --check: passed
• reviewed seven-file diff fingerprint: e2be95ccee60735a7075b88aea16f23428b4fe2c6983d5ffaef5ee8ceb174899

Stack

This PR intentionally targets codex/p7a-muse-glimmer-preflight-20260815 so the runtime-ingestion change is independently reviewable. Retarget it to main after the foundation PR merges.