Omit native thinking blocks from peek history reads

[shaneconner]

Peek currently replays native assistant thinking blocks and opaque signatures as ordinary text inside a tool result. This change verifies the original messages first, removes thinking and redacted_thinking blocks from the history view, then serializes and paginates it. Visible text, tool calls and results stay verbatim. Pi's saved messages, Fold's source references and exact expand/recovery remain unchanged.

The result identifies the view and counts omitted blocks. README and tool guidance explain the behavior and Anthropic compatibility limits. This addresses a confirmed exposure in #10; the reported Claude Opus 5.5 refusal through 9router has not been reproduced against a provider, so this PR does not claim to resolve the classifier behavior or close the issue.

Validation: lint passed; all 82 runtime gates passed with 9,794 assertions; the full experiment verifier passed all 118 gates. New gate 187 fails on the previous implementation and covers nested reads, pagination, thinking-only messages, exact restoration, unchanged originals and rejection of source drift. Six disposable offline SDK scenarios on Pi 1.0.0 and 1.0.4 cover simulated refusals, durable reads and reload. No paid provider calls or private sessions were used.

The benchmark digest was genuinely re-extracted from the authenticated campaign. Only sourceHashes.runtimeTreeSha256 and extractSha256 changed; all findings and all 102 evidence rows are identical. Package dry-run includes only the expected 25 files.