fix(recovery): surgically remove unrecoverable reasoning instead of n…

[zuhairthabit]

…uking history

The recover strategy previously dropped all but the leading system messages and the latest user message whenever any thinking-mode tool-call assistant was missing cached reasoning_content. In long sessions this silently discarded almost the entire conversation (observed in production: 307 of 309 messages dropped), causing the model to "forget" the ongoing task.

Replace the nuclear latest_user strategy with surgical removal: drop only the assistant message(s) whose reasoning_content cannot be restored, together with their associated tool results, while preserving all surrounding context. The assistant<->tool pairing invariant is maintained (no orphaned tool messages), so upstream requests stay valid.

Also:

• Only consult the recovery-notice boundary when the first repair pass finds genuine cache misses, so a stale notice no longer permanently truncates later turns.
• Suppress the user-facing recovery notice for surgical removals: context is preserved, so there is no need to plant a boundary marker that could trigger truncation on subsequent turns.
• Add direct regression coverage for single/multiple misses in long conversations and for the no-tool-calls case; update existing tests that encoded the old nuclear contract.