perf: reuse semantic state for ToolService queries

[e54-bot]

Summary

• Reuse the semantic service, symbol index, and findings for the lifetime of a loaded ToolService. A non-default session lint configuration gets its own cached findings while sharing the same index.
• Build inspect references directly from the index in symbol order, avoiding per-symbol service requests and JSON ID extraction.
• Extend wright-bench with a recorded repeated-query workload over representative Workshop inputs.

Measurement

The same benchmark harness measured the original no-reuse path before the cache change and the updated path afterward. Each measurement averages 5 trials; each trial repeats the listed 10 operations 5 times after CompilerSession::load(). loadedService includes ToolService::new plus the repeated queries.

Fixture	Before	After	Reduction
synthetic/control-flow	4.52 ms	3.28 ms	27%
synthetic/declarations-rules	3.81 ms	2.67 ms	30%
real-world/overpy-cake	37.49 ms	13.48 ms	64%

The measured sequence is rules, symbols, references(symbol:0), usage(symbol:0), cfg(rule:0), findings, persistentObjects, lint, lintRules, and inspect. The report records this workload in the semantic_queries.workload field.

Validation

• cargo test --locked --workspace --all-targets --all-features — 260 passed.
• cargo clippy --locked --workspace --all-targets --all-features -- -D warnings — passed.
• cargo fmt --all -- --check — passed.
• cargo run --locked -p wright-bench — passed.

Closes #356