Define semantic artifact lifecycle and persistence gates

[DavidBakerEffendi]

Description

Closes #817.

Defines and validates Bifrost's lifecycle policy for structural facts, semantic artifacts, CFG/ICFG views, oracle projections, and solver state. The change standardizes the evidence gates required before an artifact may move from request- or generation-local ownership into persistent storage.

The measured recommendation is to keep bounded exploded data-flow state ephemeral. Future reusable procedure, taint, and protocol summaries remain owned by #823 and should begin in memory, then earn SQLite persistence only after an equivalent serialized candidate demonstrates meaningful reuse and safe invalidation.

Key Changes:

• Add a lifecycle matrix covering artifact ownership, identity, invalidation, representation, completeness admission, observability, and current storage decision.
• Add a typed, overflow-safe promotion evaluator and migrate the existing semantic/CFG persistence experiment to the shared six-gate policy.
• Add a reproducible fresh-process data-flow benchmark across generated TypeScript, inline TypeScript/Java, and pinned VS Code/Spring PetClinic datasets.
• Bind retained solver evidence to canonical typed ICFG topology rather than snapshot-local IDs or graph counts.
• Harden benchmark provenance, interruption cleanup, locked execution, and untracked-file fingerprinting after specialist review.
• Keep the cross-adapter reference fixture external-owner-based after the base branch standardized same-owner usage exclusion.

Measured Result:

• 72 fresh release processes across eight datasets; 56 retained samples reproduced identical canonical topology, solver work, completeness, and checksums.
• The largest bounded result reached 98,313 states and retained an estimated 1,179,940 shallow bytes, with 31.526 ms first-solve and 28.001 ms repeat-solve medians.
• VS Code workspace construction had a 657.3 MiB median process peak while the finite solver result retained only 5,136 shallow bytes, showing that workspace cost is not evidence for persisting query-local state.
• No production persistence, packed data-flow DTO, cache insertion, or durable dense FactId is introduced.

Tradeoffs and Suggested Application:

• The shared evaluator makes promotion decisions consistent and auditable, but intentionally requires candidate-specific benchmarks rather than automatically selecting storage.
• The benchmark and evidence documents add maintenance surface without providing a runtime speedup by themselves.
• Keep raw worklists, truncations, witnesses, and exploded results request-local. Apply the promotion gates only to complete reusable summaries with stable identity, demonstrated cross-query reuse, and targeted invalidation.

Validation:

• cargo test --features nlp,python — all enabled library, binary, integration, and doc-test targets passed; the library target reported 1,870 passed, zero failed, and five ignored.
• cargo clippy --locked --all-targets --all-features -- -D warnings
• Focused artifact-lifecycle and semantic/data-flow benchmark tests
• cargo fmt --all -- --check
• 72-process retained lifecycle benchmark matrix
• Security, duplication, intent, DevOps, and architecture specialist reviews

Touch Points:

• .agents/docs/semantic-artifact-lifecycle-matrix.md
• .agents/docs/dataflow-lifecycle-benchmark-2026-07-24.md
• .agents/plans/issue-817-artifact-lifecycle-foundation.md
• src/benchmark/artifact_lifecycle.rs
• src/benchmark/mod.rs
• tests/measure_semantic_cfg_persistence.rs
• tests/measure_dataflow_lifecycle.rs
• tests/code_query_pipelines.rs
• scripts/run-dataflow-lifecycle-benchmarks.sh