[BUG] Ensure context truncation always makes measurable progress #1254

Problem

Fallback context truncation can report success without reducing context. With a short history such as an assistant tool_use followed by one oversized user tool_result, the current 50% message calculation rounds down to zero removable messages. Zoo Code still emits a truncation event, leaves the oversized history unchanged, and retries into the same failure indefinitely.

The reproduced signal is Context Truncated with 0 messages removed while the recalculated context remains above the allowed input budget.

Scope

Make context recovery monotonic and bounded.

• Treat recovery as successful only when the recalculated model-facing token count decreases.
• When message-level truncation cannot remove a valid turn/tool pair, shrink the largest eligible textual tool-result block while preserving its tool identifier and result shape.
• Recount after degradation and verify the result fits the target or makes measurable progress.
• If protected content makes the target impossible, return a distinct actionable error instead of emitting another successful truncation event or retrying indefinitely.
• Keep retries bounded.

Acceptance criteria

• One-to-four-message histories cannot emit a successful zero-progress truncation result.
• An oversized tool_use/tool_result pair is reduced without orphaning either side.
• newContextTokensAfterTruncation is lower than the pre-recovery count whenever recovery reports success.
• An impossible budget produces a controlled error and does not loop.
• Focused tests reproduce the 4 MB tool-result case and verify bounded recovery.

Out of scope

• Preventing individual tools from producing oversized results
• Condensation-input sanitization
• Full-output spill files or resource references
• Aggregate budgeting across parallel tool calls

Estimate

1 story point. Independently mergeable and non-stacked.

Parent tracking issue: #648