perf(test): consolidate agent runtime contract targets

[limityan]

Summary

• consolidate 27 cross-platform bitfun-agent-runtime integration-test crates into four responsibility-aligned targets
• keep the Unix-only native process/timeout suite as a fifth independent target so platform and process isolation is preserved
• disable Cargo autotest discovery and extend the existing core-boundary checker to reject unregistered roots, orphan leaves, duplicate references, and non-flat grouped roots
• document the target ownership and focused-test command without changing production code, dependencies, features, Cargo.lock, workflows, or CI coverage

Why

The package previously linked 28 integration-test executables against the same feature-free runtime dependency closure. Most of those crates represented source-file granularity rather than an architectural or execution boundary. The new layout preserves the one real platform/process boundary while reducing redundant test compilation and debug artifacts.

Behavior and architecture

• all 27 moved test sources are byte-identical renames; native_hook_execution_contracts.rs remains at its original path and unchanged
• Windows preserves 224 integration tests and 280 lib tests; the normalized pre/post integration-test list is identical (a03b977a31db21c247db636d85f60250666a979a57c2b186a0a0ef8129eabde2)
• Unix keeps the additional 22 native process tests in their own executable (246 integration tests total)
• grouped roots are intentionally flat indexes; the existing boundary check makes autotests = false fail closed when a future test file is not registered

Local performance observation

Windows 10.0.19045, i7-10700, rustc/cargo 1.97.1, dependencies warmed, A/B/A/B/A/B order. Each sample ran:

cargo clean -p bitfun-agent-runtime

cargo test -p bitfun-agent-runtime --no-run --locked --quiet

Layout	Samples	Median	Integration PDBs
28 targets	12.48s / 15.01s / 13.63s	13.63s	312.3 MiB
5 targets	11.90s / 11.04s / 11.31s	11.31s	77.4 MiB

This is a local observation: median package test compile/link time decreased by about 17.0%, and integration PDB size by about 75.2%. It does not claim fewer third-party dependencies or features.

Validation

• pnpm run fmt:rs
• cargo test -p bitfun-agent-runtime --locked --tests (504/504 on Windows; repeated three times before rebase and rerun after rebase)
• each grouped target and the normalized test-list equivalence check
• cargo check -p bitfun-agent-runtime --tests --target x86_64-unknown-linux-gnu --locked
• pnpm run check:core-boundaries:test (40/40)
• pnpm run check:core-boundaries
• cargo check -p bitfun-core --features product-full --locked
• two independent adversarial review rounds plus a final parser-bypass review: no remaining P0-P3 findings

The Windows host cannot execute the 22 Unix shell/process tests; existing Linux CI remains the runtime proof for that path.