Cache TemplateCaps.detect results with a bounded LRU

[leehack]

Context

PR #446 measured repeated Jinja capability detection and found that parsing dominates the cost: one explicit Lexer/Parser pass plus a second parse inside dinja Template construction, followed by four distinct render probes. The render probes cannot be collapsed because they exercise different payloads.

The redundant second parse cannot currently be removed without an upstream dinja API change; #351 tracks the broader private-source/upstream coupling. A bounded cache at the TemplateCaps.detect boundary would avoid both parses and all four renders when the same template is detected repeatedly.

Proposed work

• Add a bounded LRU keyed by exact template source, with an explicit conservative capacity.
• Preserve all current capability results and failure diagnostics from fix: label and log Jinja capability probe render failures #446.
• Define cache ownership and invalidation behavior; do not introduce unbounded process-wide retention.
• Add hit, miss, eviction, and failure-path regression tests.
• Benchmark representative real templates before and after.

Boundary

This is a non-blocking performance improvement discovered during #446. It is not a prerequisite for the diagnostics fix or the next release.