Compaction fails with thinking-block API error on Bedrock, clears context on failure #267

Acceptance Criteria

• Compaction succeeds without API error when the conversation contains thinking/reasoning blocks (Bedrock Claude Opus 4.6, same-model compaction)
• Reasoning blocks are unconditionally stripped (downgraded to plain text or dropped) from all messages sent to the compaction agent, regardless of model match
• A failed or incomplete compaction (orphaned marker with no valid summary) does not truncate visible conversation history — the user retains their full pre-compaction context
• Existing test suite passes; new test cases cover both the stripping behavior and the orphaned-marker guard

Testing Decisions

• Extend packages/opencode/test/session/message-v2.test.ts (already has a "thinking-bedrock" fixture) with a case verifying reasoning blocks are stripped when building compaction messages
• Add a compaction test that creates an orphaned compaction marker (user-compaction message without a valid summary-assistant) and verifies filterCompacted / history loading does not treat it as a boundary

Key Decisions

• Strip, not preserve: reasoning blocks are dropped from compaction input unconditionally (not just for cross-model). The summarizer benefits from the reasoning content only as labeled plaintext ([Reasoning]: ...), never as provider-native thinking blocks
• Guard is passive, not rollback: orphaned compaction markers are left in the DB but ignored by the boundary detector, rather than actively deleted. This avoids a delete path in the message store and handles any future failure mode (timeout, network, provider quirk) without additional code
• Option-based signaling: the V1 path signals compaction context via a forCompaction: true option in the existing options bag passed to toModelMessagesEffect, consistent with the existing stripMedia and toolOutputMaxChars options

Constraints & Invariants

• specs/v2/session.md invariant: reasoning blocks do not survive compaction boundaries — this fix is enforcement, not a policy change
• The Anthropic API constraint: thinking/redacted_thinking blocks in the latest assistant message of a request cannot be modified from their original form
• Compaction must remain idempotent: a second compaction attempt after a failed one must work correctly (no stacked orphaned markers blocking progress)

Prior Art

• specs/v2/session.md — compaction spec, reasoning survival rules
• CONTEXT.md (root) — "Native Continuation Metadata" definition, model-switch lowering rule
• AMICODE-PATCHES.md — Patch Uniform, slightly taller default tile widgets #26 (thinking block restructure), adaptive-thinking version-regex fix
• packages/opencode/test/session/message-v2.test.ts — existing "thinking-bedrock" fixture and "substitutes space for empty text between signed reasoning blocks" test

Notes

• Reproducible on Amazon Bedrock with Claude Opus 4.6 (model ID: anthropic.claude-opus-4-6-v1) when extended thinking is enabled
• The error message is: messages.7.content.8: 'thinking' or 'redacted_thinking' blocks in the latest assistant message cannot be modified. These blocks must remain as they were in the original response.
• The "context cleared" symptom suggests the compaction epoch advances (new context baseline) despite no summary being produced — the guard prevents this