Track A memo cut: per-module normalize+ownership memoization — resolve 371s→319s, both stages −95%, fingerprints byte-identical

[gunbai-bot]

Track A memo cut: normalize + ownership per-module memoization (−52s of the receipted 371s)

The resolve-split receipt (#6535) priced two per-entry recomputations at ~15% of whole-corpus resolve: normalize (28.5s, 8%) and ownership (24.3s, 7%) — both pure per-module facts recomputed per entry at zero marginal information. This PR collapses them to first-touch-per-module within a worker.

Model first, then host cache (the typed_module_cache pattern)

• Authority (src/v1/03_normalize.dag, src/v1/compile.dag): the per-module row of each stage's fold is factored into a named fn — normalize_module_diagnostics(m, source_indices) and module_ownership_proofs(m) — with normalize_graph / extract_ownership_proofs now folding over exactly those rows. Pure factoring; seed regenerated (regen_divergence_count=0).
• Host (cli_run.rs): two MultiEntryIndex RefCell caches keyed by declaring-file path (the same in-process source-stability contract parse_cache/typed_module_cache ride; module_source_identity walls name↔file divergence loud). resolve_entry_with_parse_cache assembles diagnostics per module in authority order — a module with no bodied items contributes the same empty row the authority's filter skips, so the assembled list is identical by construction.

Receipts (all by execution)

• Equivalence, controlled pair (same tree, only src/v1 differs): whole-tree probe PRE (main binary) vs POST — corpus_fingerprint=95da9a3c56e6bcaf, emit_graph_fingerprint=97a9b47b34c387e3, 1,104 modules — byte-identical. (The probe path bypasses the memo, so this isolates the authority refactor; the memo path's equivalence rides the full corpus below.)
• Whole-corpus falsifier, memo live: EXIT=0 green, 1,776 witnesses, 0 skipped, governor healthy (width 9, forced_serial=0, hard_backoffs=0):

Stage	Pre-memo (#6535 receipt)	This PR
normalize	28,518ms	1,346ms (−95%)
ownership	24,263ms	1,168ms (−95%)
resolve total	371,096ms	318,956ms (−52.1s)
typecheck / reconcile_assembly	194.7s / 107.7s	197.4s / 104.2s (noise)

• Battery: cargo test --release -p v1-compiler --lib 214 passed / 0 failed; regen_stage0 --verify divergence 0.

Note on an overlapping hunk

The branch carries the one-row nfr roster backfill also shipped standalone as #6541 (main is whole-corpus red without it — #6533 landed frontier_blocker_class_matches unrostered; my falsifier receipt above needs the row to run green). Whichever merges first, the other's hunk falls out cleanly.

Remaining Track A mass after this cut, per the receipt: typecheck cold computes (197s — cross-worker share, Inc B territory) and reconcile_assembly (104s — resolver-graph-major lane), both handed off with numbers.

🤖 Generated with Claude Code