initialize_runtime(preprocess=True) runs full preprocess on every call

[u2135]

Summary

Developers report initialize_runtime() taking ~8–9s and suspect it may run once per user request.

Current behavior (SDK)

• ensure_bootstrap_cached() is idempotent per process (middleware/bootstrap work skipped after first success).
• With default preprocess=True, initialize_runtime() always runs ensure_preprocess_intercept_padding (if POSTGRES_URL is set) and run_preprocess(...) on every invocation — there is no “only first call” guard for preprocess.

Impact

• Correct usage (once at worker startup): one-time cost is expected.
• Incorrect usage (per HTTP request / per agent turn): repeated multi-second preprocess even when bootstrap is cached.

Acceptance criteria

• README + examples explicitly state: call initialize_runtime() once per process before enabling intercepts; do not call from per-request handlers.
• Consider API/docs clarification: e.g. split “bootstrap only” vs “bootstrap + preprocess”, or document initialize_runtime(preprocess=False) for advanced setups.
• Optional follow-up issue: skip redundant preprocess when cache indicates preprocess already completed recently (product decision).

References

• src/provably/handoff/client.py — initialize_runtime
• src/provably/handoff/_bootstrap.py — ensure_bootstrap_cached