Every pass with a thinking part deep-clones the whole message array for a retry that almost never runs #798

What happens (main a5c029921): transform.ts:821-829 checks whether any message has a reasoning, thinking or redacted_thinking part. If one does, it calls cloneKeepingCoveredInstructionMarks(messages), which is structuredClone of the whole message array (host-served-rows.ts:159). On a Claude session with thinking enabled, that is every pass. The copy is used only by the signed-thinking fit retry at transform.ts:3295-3317. That retry needs all three:

• an active signed-thinking prefix hold (activeThinkingPrefixHold, set at :1830);
• a final-wire estimate above the boundary limit;
• no retry already running for the session.

Otherwise the copy is dropped.

Measurements:

• Host profile: one OpenCode 2 server process serving about 20 sessions, 0.49.22, 120 s CPU profile. The main thread was about 66% busy, with 81 event-loop lags over 250 ms (max 998 ms). structuredClone under cloneKeepingCoveredInstructionMarks was 8.0% of the JS thread, the largest single Magic Context entry.

• Cost per pass: Bun structuredClone of real message tails (the rows after the newest compaction row, parsed from the store), 7 runs each:

  Tail rows	JSON size	Median clone
  808	3.6 MB	13 ms
  1,384	4.8 MB	18 ms
  1,537	7.6 MB	31 ms

  This runs synchronously on every LLM step of every such session.

• Benefit: about 10,400 transform passes over four days of logs on that host contain 0 fit_retry lines; the retry's rebuild logs origin=fit_retry. So in that window every copy was thrown away.

Possible cheaper paths:

1. Take the copy only when the retry is possible. The hold is known at :1830. If no stage mutates message content between :829 and :1830, the snapshot can move there and be gated on activeThinkingPrefixHold. It could also be gated on pre-pass usage being near the limit. (We haven't checked whether anything mutates in that range.)
2. Copy on write. Shallow-copy the array and the message objects, and clone a part only when a stage changes it in place.
3. No snapshot on OpenCode 2. Rebuild the retry input from the host store instead.

Same class as #703 (a deep clone on every pass), at a different site.