Finer CI assertion gates: per-phase pruning, stats coverage, max files read

[cdelmonte-zg]

The current gates are --min-pruning (total pruning percentage) and --assert-stats (table-wide stats presence). They cannot express which phase must do the work, nor bound the absolute cost of a scan, nor catch a predicate column whose statistics are missing on the files that matter.

Proposed additions:

• Per-phase minimum pruning thresholds, one for partition pruning and one for data skipping (exact flag names to be settled in the design; the phase vocabulary is the one already in the report).
• A stats-coverage gate over the per-column coverage shipped in Per-column stats coverage for stats-safe predicate columns #132 (analysis.stats_coverage): fail unless every predicate column's coverage over the candidate files meets a threshold. 100% recovers the original "listed columns must carry stats" case; 0% is the structural no-op, which by design does not degrade confidence (see the resolution in Per-column stats coverage for stats-safe predicate columns #132) - the gate, not the label, is the CI channel for it. Requirement-aware for free, since the block already distinguishes min_max from null-count coverage. Exact flag shape to be settled in the design (a global --min-stats-coverage <pct>, a per-column form, or both).
• --max-files-read <N>: absolute bound on surviving files, complementing the percentage-based gates on tables where percentages hide real cost.

All of these follow the existing gates model: pure assertion records in gates.rs, deterministic failure messages, additive JSON.

The stats-coverage gate asserts on data the report already computes: #132 shipped analysis.stats_coverage with schema_version 0.5.0.