fix(core): retire instruction notices after revert resync - #62028

One thing this leaves in place: after the undo, the baseline is rebuilt from the current values, so the system prompt changes and the next request writes the whole conversation to the prompt cache again.

An alternative that fixes the same contradiction and keeps the cache: when the undo removes an update but keeps the epoch start, reset the fold to its baseline instead of deleting it. The surviving notices stay valid, the next step restates what changed since, and the cached prefix is still read. The same idea covers forks (#61550). The change is here: dev...swf-cmd:opencode-upstream:fork-instruction-baseline

I can't open it as a PR right now because I'm at the repository's open-PR limit, but I'm happy to once maintainers say which direction they prefer.

@swf-cmd Thanks, I agree that preserving the baseline can improve prefix reuse. I tested your alternative at 44fff3b and found the same retained-notice mismatch on undo:

• A → B → A, then undo that keeps B but removes the later A update. Resetting current_values to baseline A means the next observation of A emits no correction for the surviving B notice.
• A → B → C, then a native checkpoint, followed by undo past the checkpoint while B survives. The fresh C baseline is accompanied by B without a corrective C notice.

I added the A → B → A case and the checkpoint case in b42e774a. Both pass here; the four related test files pass (313 tests), each new case passes 10 repetitions, and bun run check passes.

To retain the notices safely, the tracked state needs to reflect the surviving history and its instruction epoch before calculating the current update. Resetting it to the baseline alone leaves these cases inconsistent.

AI-assisted testing and analysis.

Right on both, thanks for testing it. Resetting to the baseline forgets what the surviving B notice told the model. And once the revert goes past a checkpoint, my version just clears the state the way dev does, which is the bug this PR fixes.

Keeping the prefix would need the state rebuilt from the surviving notices on their epoch's baseline, which instruction_state doesn't keep for earlier epochs. That's more than this PR should take on, so please consider my alternative withdrawn here too. Any prefix-keeping change would come later, as a follow-up on top of this.