fix: relay tool-result media + harden persistent Claude SDK stream (batch)

[mradwankhalil]

Tool-result media fix + local hardening batch

Headline (this push): tool-result images now reach Claude

Problem - on the claude-code provider, images returned by tools (e.g. the read tool on a PNG) never reached the model; it only saw "Image read successfully". Directly attached images were unaffected.

Root cause - OpenCode cannot place media inside tool results for openai-compatible providers, so it promotes tool media into a synthetic user message: "Attached media from tool result:" (sst/opencode packages/opencode/src/session/message-v2.ts; supportsMediaInToolResult returns false for @ai-sdk/openai-compatible). In the parked-bridge (MCP) flow this plugin resumes the turn by resolving the parked tool call only - the synthetic message was ignored, so the media was dropped on the floor.

Fix -

• collectToolResults now captures tool-result image parts ({ text, attachments }) when a transport does include them.
• collectSyntheticToolMedia detects OpenCode’s promoted synthetic message and merges its images into the single resolved tool call.
• The parked MCP reply now returns image content blocks alongside text.

Evidence (claude-code/opus through this proxy)

• read on a generated PNG -> "The image says SQUEEZ-TEST-7-3-9. It’s orange monospace text on a dark background."
• read on a generated JPEG -> "JPEG-TEST-4-2-0 ... cyan monospace text on a dark navy background".
• Adapter debug log: [opencode-claude] attached promoted tool-result media {"count":1} immediately before resuming parked bridge {"resolved":1}.

Regressions checked - plain-text reply (PONG), bash tool call, text-file read: all pass. Repo suite: npm test (12 bun regression files) green.

Limitations - images only (PDFs in tool results stay text); remote http(s) image URLs are skipped; if a single request resolves multiple parked tools the promoted media cannot be mapped and is skipped with a warning; the history-transcript path (non-resumable sessions) is unchanged.

Also in this batch (previously local)

• de6c815 declare input window for 1M models so auto-compaction fires before the edge
• 55112db / 9584142 / 5f600b8 persistent Claude SDK Query + prompt input queue
• 8c24d64 serialize persistent response pumps; 0b1aa2f serialize SDK stream pulls
• 00b87c9 harden bridge teardown; d0cb807 defer input.close() off the response path
• 0c89d3a resolve Windows Claude CLI fallbacks
• f858a34 reset Claude session after compaction; 994db2b classify summary updates as metadata; 29ffae0 per-turn usage across compaction continuations
• f6fe917 / 0865ab1 / ae1a42e regression coverage for lifecycle, continuation cancellation, Query lifecycle

[mradwankhalil]

Follow-up fix added after live validation. /compact was lowering OpenCode's local context correctly, but the next normal request resumed the old sticky Claude session, whose hidden pre-compaction context reported ~806k cached input tokens. b4d999 adds the regression; 858a34 clears the base conversation binding at a classified summary boundary so the next turn transfers compacted OpenCode history into a fresh Claude session. Focused regressions, typecheck, build, and diagnostics pass. The full suite still stops at the unchanged Windows smoke assertion in test/smoke.ts:236.