P0: Add output-token inflation guard to benchmark

[jhonatanjunio]

Context

Research on competitor rtk-ai/rtk#582 revealed a critical failure mode of bash-output compressors: aggressive compression of input can increase total cost by 18% because the LLM generates 50% more output tokens trying to compensate for missing structural context.

Concrete numbers from that issue:

• Input savings: ~80%
• Output tokens: 1.6k → 2.4k (+50%)
• Total cost: $0.68 → $0.80 (+18%)
• Duration: +26%

We currently measure compression ratio but not information preservation — we are flying blind on whether our compression is causing Claude to re-investigate.

Goal

Add an information-preservation metric to cargo run --release -- benchmark that:

1. Counts preserved signal terms in compressed output vs baseline: file paths (path/to/file.rs:NN), line numbers, error markers (error/warning/FAIL/ERROR), test verdicts (passed/failed).
2. Reports preservation_pct per scenario alongside reduction_pct.
3. Emits a compression-risk warning when reduction_pct > 90% AND preservation_pct < 70% — the regime where rtk regression occurred.
4. Aggregates risk score across all scenarios.

Acceptance criteria

• src/economy/preservation.rs exposes fn info_preservation(baseline: &str, compressed: &str) -> f64 returning value in [0.0, 1.0].
• Detects: file paths, line numbers (:NN and (NN,NN)), error keywords, test verdicts.
• ScenarioResult gains info_preservation: f64 field.
• Benchmark output prints the new column.
• Risk warning emitted when threshold crossed.
• Unit tests cover: high preservation, low preservation (over-compression), edge cases.

Out of scope

Live Claude API calls — follow-up PR (requires API key). This PR delivers the offline proxy metric.

[jhonatanjunio]

Moved to upstream claudioemmanuel/squeez. Superseded.