perf(analysis): reuse projections and analyze corpus cases once

[zhen8838]

Why

• Fixes perf(tests): cut the time the analysis corpus takes #197.
• Analysis corpus cases repeated analysis and scope construction, while each analysis repeatedly projected the same call into the same topology window.

What

• Cache projected access relations on the scope that records the call, keyed by call identity and topology level. Scope construction, memory, footprint, view folding, and allocation alias proofs share the cache.
• Partition the 72 corpus cases into 64 CLI cases and 8 API cases. Each case runs one analysis of all four families; both paths use the same report checks and retain every expected memory peak. API cases share one additional scope tree for internal checks.
• Full verification and benchmark on frozen bb4472aa: source 1332 passed, 10 skipped in 150.53s with 8 workers, compared with the recorded 315s baseline (runner wall time: 154s); installed wheel 162 passed in 273.79s (runner wall time: 287s). Both manifests have zero failures/errors; pre-commit passed.
• Focused verification: 484 passed across analysis, schedule, CLI, and inspection (rv1-analysis); pre-commit passed.
• Worst corpus case: 39.1s → 10.3s. The analysis corpus file occupies 1 of the slowest 6 source tests.
• cProfile: the old complete case accumulated 3106 projected calls across two analyses and three scope builds; a new CLI case runs one analysis with 570 = 190 calls × 3 windows. An additional standalone scope build needs 190 projections.

Contract

None.

Risk

• Timing comparisons use the recorded baseline and a shared development machine.