Accelerate versioned graph history

[forhappy]

Summary

• hard-cut versioned history onto sealed immutable realizations and verified
  shared extraction caches
• build arbitrary revisions through bounded extraction with portable AST/program
  cache reuse
• add request-scoped realization readers so diff/query/view paths open a graph
  once
• cache canonical semantic diffs and bounded viewer projections for instant
  repeat access
• add history verify plus explicit history cache status/GC commands
• add reproducible CocoIndex and Podman release qualification

Real-repository results

CocoIndex (90571539 → 71f9cc9d):

• adjacent history build: 154.65s baseline → 14.37s
• first semantic diff: 0.39s; repeated: 0.13s
• historical viewer export: 0.06s
• no-op history build: 0.08s

Podman (d8380c9c → 7ac3e837):

• first semantic diff: 1.02s; repeated: 0.23s
• historical viewer export: 0.09s
• no-op history build: 0.08s

Semantic and viewer first/repeat artifacts are byte-identical on both
repositories. The original source checkouts remained clean.

The qualification report explicitly records remaining gaps: first-time history
construction is still above the current-extractor-relative time/memory targets,
and Podman's first diff peaks at ~649 MiB (the repeat path is ~169 MiB). Cached
graph access, diff, and viewing meet their latency targets.

Verification

• cargo fmt --all -- --check
• cargo check --workspace
• cargo test -p compass-files --test contracts (19 passed)
• cargo test -p compass-history --test diff (3 passed)
• cargo test -p compass-core --test history_materialize (7 passed)
• scripts/qualify_history_real_repo.sh against the approved CocoIndex and
  Podman revision pairs
• graphify update .

Full measurements and SHA-256 evidence are in
docs/superpowers/reviews/2026-07-26-versioned-history-performance-qualification.md.