Improve bounded RSI startup observability

[w4ffl35]

Summary

Closes #122.

• Reuse verified snapshot bytes within one package-load invocation, while checking the package file set, index, blob hashes, Git identities, and materialized tree before use.
• Emit bounded phase receipts for proof preparation and campaign observation, including terminal preparation failures. A receipt remains preparing and cannot admit inference; existing durable monitor admission stays authoritative.
• Add offline tests for cache tampering, slow preparation and observation, failure receipts, and zero inference enqueue before admission.
• In one offline profile using the same pinned fixture package and mocked pricing/queue, time to mocked campaign acquisition measured 40.071 seconds at the base commit and 23.156 seconds on this branch (416 schedule cells in each run). This is one before/after observation, not a general performance guarantee.

Verification

• npx tsc --noEmit — passed.
• env -i PATH="$PATH" npm test — passed: run-tests: all 248 file(s) passed. One separately opt-in PostgreSQL monitor test was skipped because no explicit isolated test database was provisioned.
• git diff --check — passed.
• npm pack — not applicable; package files were unchanged.
• E2E suites — not applicable; no orchestrator, watcher, spawn, or scripts/*.sh files changed.

Security impact

The change adds no permissions or inference path. Preparation receipts are diagnostic only and explicitly cannot authorize inference. The durable monitor remains the admission authority. Cached bytes are scoped to one package load and continue to be checked against the current package and tree identities.

Test plan

Run npx tsc --noEmit and the full npm test suite. The startup/profile tests use mocked pricing, queue, and inference boundaries; they do not dispatch inference. The optional PostgreSQL monitor test requires its own explicit isolated database configuration.