fix(paseo-omp): replay failed turns from native transcript

[omercnet]

Closes #78

Problem

OMP's get_messages RPC exposes model-safe context. OMP deliberately removes assistant turns ending in error or aborted and their paired tool results from that context. paseo-omp used the same response for user-visible history replay, so resume/import silently omitted failed assistant output and completed tool activity even though the native JSONL remained intact.

Tool-intensive valid assistant messages could also exceed the plugin's 64-content-part validation and projection limits.

Change

• Read the authorized native root transcript for initial resume/import replay.
• Reconstruct only the active root-to-leaf branch and preserve message entry identities.
• Validate the file handle, native session ID, workspace, size, entry count, message count, parent links, duplicates, cycles, and image blob hashes.
• Read a stable bounded transcript snapshot and hydrate bounded blob:sha256:* image payloads from the selected OMP store.
• Keep RPC get_messages for rewinds, where the in-memory branch can differ from file order.
• Fall back to RPC history with an explicit incomplete-history timeline warning when persisted replay is unavailable.
• Raise the shared assistant content-part limit from 64 to 4096 and project indexed content through the same bound.
• Replay tool calls alongside their results so histories with more than 64 calls preserve every call's arguments and specialized details.
• Forward persisted transcript replay through named-profile runtimes.

Verification

• npm test — 679 passed, 3 skipped
• npm run check
• npm run typecheck
• End-to-end smoke: 81 messages, 79 completed tool results with the final path preserved, 108 assistant parts including display content after index 63, reasoning after index 63, and a hydrated native image blob.
• GitHub CI passed on Ubuntu, macOS, Windows, WSL, and the real OMP compatibility matrix.