Index admitted mixed typed source without losing bundle authority

2026-09-24 14:34:22 -07:00
parent aaec78fcc9
commit a18b7966e5
7 changed files with 964 additions and 99 deletions
@@ -0,0 +1,51 @@
+# Private mixed typed occurrence index
+
+Source checkpoint on `codex/analysis-mixed-index-956atf-20260924`, based on
+`aaec78fcc9b273d1485497126392afdd1fb2e54e`. This records implemented structure;
+compilation and runtime acceptance have not yet occurred at this checkpoint.
+The accepted contract/review are published at
+`7cfdd13d3a1472e63fd34b0080b2cfc4b921b23e:harness/analysis-mixed-index-contract-20260924.md`.
+Independent exhaustive fixture data and its raw-byte checker are at
+`ae6df4fa1f103ff9b14963731b2edd84e38738b1:harness/analysis-mixed-index-20260924/`.
+
+`ResolvedMixedProject::index` consumes the sealed complete-Project witness.
+`IndexedMixedProject` owns it and shared `IndexData`; read-only methods expose
+the Project, warnings, original bundle/canonical files, names and typed maps.
+It has no conversion to the legacy numeric-source Index or its updater. The
+constructor-only adapter keeps its file table private, so shared verification
+does not accept caller-fabricated mixed source/token ownership.
+
+Shared verification preserves original token, owner, structural child-path,
+qualified call-label and physical occurrence checks. Mixed catalog construction
+streams complete Project globals and scoped parameters once. It does not
+reparse, relower, duplicate model records or repeat admission. Ordinary globals
+are names/targets here; ordinary source occurrences are not indexed by this view.
+Complete non-expression occurrence emission remains the sealed lowerer's
+invariant, checked by an independent exhaustive output oracle.
+
+Mixed allocations reserve against the retained aggregate ledger before growth.
+New map/vector/set/Path entries charge entry slots; retained raw lowerer
+occurrences keep occurrence slots. New names charge copied UTF-8 bytes. Owner
+and physical-token scratch drop after occurrence checks; path/call scratch drop
+before syntax-edge traversal. Shared Path nodes remain retained while map keys
+own them. Legacy typed-only construction retains its original local policy.
+
+Precompile independent review found stale diagnostic file attribution in two
+late verification loops and unsafe fallback for invalid file/UTF-8 spans. The
+source now updates attribution for each checked record and maps invalid
+locations explicitly to a safe start span, retaining a valid file when known.
+Regression cases are authored; no passing runtime result is implied.
+
+Seven new tests cover complete OA/SA names/maps/model and Unicode/CRLF variants,
+qualified corruption cases, inherited-plus-new storage and copied-name byte
+boundaries, original provenance/admission metadata, and small-stack success and
+late rejection. Existing rich fixture/model construction is factored without
+removing its original full arena test. Actual test inventory, strict core lint,
+focused checks, full private-library dispositions and formatting still require
+current-source execution. Prior private source8 failures are not suppressed.
+
+The working agreement is synchronized to current main blob
+`e76647f49913b314e1543c4464f6199533909bae`. This source remains separate from the
+Windows prerelease. Public source8, ordinary occurrence composition, source
+updates/output admission, migration, checked save/history, GUI and API parity
+remain required subsequent work.