[AdaWorldAPI]

Summary

This removes slice C's 64-position limit and follows your rulings: wider call chains, presence bits for thin provisioning, sparse writes merged on read, and versioned replayable immutability kept in the DR index itself. Since the architecture review, it also makes refresh_crate atomic, closes codex's finding about binds to unmapped ids, and runs the codegraph tests in CI.

dt64 (dr)

• MapTable::locate maps a global id to a (tile, local) pair, which closes D-DYNBAG-4's open map item.
• LocalMap::extends checks that a map only appends ids. LocalMap::view lets a caller check a map before handing it to the table.
• Axis::for_survivors picks the tile width from the survivor count (D-DYNBAG §4.1).
• Versioned index. Each tile's mapping is an append-only log keyed by RouteVersion:
  • an extension logs a longer length, with no copy;
  • a replacement adds a new generation;
  • a removal logs an event.
• MapTable and RouteLog share one sequence through a VersionClock.
• map(frame, tile) reads the layout the frame was sealed at, and map_at(version, tile) reads any version by number.
• A frame from another table, or a future version, is refused.
• Lookups return a borrowed TileMap and allocate nothing.

xlog

• closure_wide handles any number of positions. Each row stores only its non-empty 256-position blocks, with a multi-word presence mask and no tile cap. The narrow path for 64 positions or fewer is unchanged.
• Cube::new_tiled spans several tiles.
• Per-relation delta log, kept as a multiset: an add counts +1, a retire −1.
  • Every read sees the base plus the log, through one accessor.
  • refresh_closure recomputes only the rows that reached a retired edge's source.
  • refold runs only when called.
• rebase accepts a new frame only if every tile extends its old map, and checks everything before it changes anything. Its private snapshot is gone; it reads the old layout through the index.
• Cube::stage / Cube::stage_closure / Cube::commit.
  • stage checks a batch of edge operations without changing the cube or the table: the grown map must extend the current one, and every id must be placed.
  • stage_closure computes the closure as it will be after the batch.
  • commit applies the batch and cannot refuse.
• Binds keep their global id (codex, resolved). An extending rebase or commit places a bind that was unmapped when the view was built. A mapped bind keeps its coordinate.

codegraph

• Stores of any size now open on the multi-tile cube. Pos is u32.
• refresh_crate is atomic: prepare, then validate, then commit.
  • Prepare and validate run against staged state:
    • the interner (SymbolStage);
    • positions and last_owner (PosStage);
    • the grown tile map, not yet given to the table;
    • the edge batch (Cube::stage);
    • the closure refresh.
  • A refusal at any point changes nothing: no interner entry, position, version, frame, log entry, lane, mask or digest.
  • The commit (table advance and seal, Cube::commit, absorbing the stages, lanes) has no error path.
  • The rest is unchanged:
    • positions never renumber;
    • growth past the reserved tiles returns NeedsReopen;
    • writes are grouped by relation and sorted ascending by (source tile, source, target).
• Rows are interned id lanes, and an unchanged crate is skipped by its digest.
• references compares ids instead of strings.

Bugs found by measurement and fixed

• impact copied only the first 64 bits of each closure row, which became wrong past 64 positions.
• callers panicked on a multi-tile store, because the cube's mask covers every reserved position while codegraph's masks cover only the symbols.
• A refresh lost a duplicate call site. A call graph is a bag, so the fold is now a multiset.
• Found by the review: a refused refresh could leave the graph half-updated.
  • Before the change:
    • interner entries stayed behind after NeedsReopen;
    • after a cube refusal, the positions and cube were already changed while the lanes were not;
    • a retry counted ADDs twice.
  • Fixed by the atomic refresh above.

Test Plan

• Tests: dt64, xlog, xlog-relcomp-probe, codegraph and egglog-r2il-probe (at the egglog pin) all pass. cargo clippy --all-targets -- -D warnings and fmt are clean on each touched crate.

• CI: rust-test.yml now runs crates/lance-graph-codegraph. It is excluded from the workspace and ran in no workflow before.

• Atomic refresh falsifiers (graph.rs mod atomic). Each compares a full-state snapshot: interner, positions, last_owner, lanes, digests, table version, revisions and tile history, cube frame and every delta log, masks, cached closures.

  • NeedsReopen, an injected stage refusal and an injected closure refusal each leave the snapshot unchanged, and no version is issued.
  • A retry after both refusals equals the refresh that never failed.
  • m::b keeps 4 call sites, as in a clean open.
  • The snapshot does change on a successful refresh.

• Bind and staging tests (tests/versioned_map.rs):

  • codex's example: the old view answers like a fresh one after a rebase, and also after stage/commit;
  • a mapped bind is unchanged;
  • staging issues no version and refuses a non-extending map.

• Disable runs. Each guard added since the review fails exactly its own test:

  • absorbing before the reservation check;
  • advancing the table before staging;
  • absorbing before the closure block;
  • deduplicated keys;
  • no rebind in rebase;
  • no rebind in commit;
  • no extends check in stage.

• Earlier falsifiers, unchanged:

  • narrow and wide closures are bit-identical up to 64 positions;
  • a 694-position, 3-tile graph matches a DFS written in the test;
  • absent blocks allocate nothing;
  • refresh equals a rebuild, with rows_recomputed == |A|;
  • refold preserves the closure;
  • rebase refuses a reorder;
  • dt64 F1–F6 and xlog F7;
  • a refreshed codegraph store matches a clean open, with one dangling call when a dependent is skipped and no renumbering.

  Every earlier guard was disable-verified. Re-pinned tests keep their old assertion as a comment.

• Measured on mask-risc's store (695 symbols, 3 tiles), release build:

  measure	result
  impact	286 ns per call
  callers	29 µs
  references	45 µs
  closures (one-time)	257 µs and 337 µs
  blocks kept	RCONF 614 of 2,304; RPOSS 951 of 2,304
  refresh of an unchanged crate	1.5–1.7 ms (fails the gate)

  The refresh fails because it reads and hashes the 2.5 MB text rows file. The fix is D-CODEGRAPH-0-BIN: binary lanes with a header digest.

Decisions recorded (not implemented here)

• P2: d-codegraph-0-binary-lanes-v1 Amendment A (§8):
  • a provenance section (charon version and commit, LLBC schema version, LLBC digest, toolchain, target and build config, source revision);
  • per-row identity (artifact, FunDeclId);
  • INVOKES_WORD to a registered DT64 reference;
  • falsifiers F9–F12.
  • Open: where "this function realises word W" comes from.
• P3: d-dt64-method-authority-census-v1 (D-DT64-AUTH). DT64 is the one method authority.
  • Parallel identities stay as listed LAB adapters.
  • A census with a two-way ratchet: an undeclared production word fails, and so does a stale exception.
  • egglog is pinned to commit 5deaa5637a306c48c998f88d5d08792d2dbfd037.

Board

• Architecture review entry and decisions entry.
• Plans: d-codegraph-0-wide-closure-v1, d-dynbag-4-dr-versioned-mapping-v1, d-codegraph-0-binary-lanes-v1 (amended) and d-dt64-method-authority-census-v1.
• STATUS_BOARD rows: D-CODEGRAPH-0-C2, D-CODEGRAPH-0-BIN, D-DT64-AUTH and D-DYNBAG-4.
• Checks: entries_index.py --check, append_only_gate.py, the supersession index and plan_dids.py all pass.

Not to be merged automatically.

🤖 Generated with Claude Code

https://claude.ai/code/session_01X6y3drwKSE2zSgoexheLFX

[coderabbitai]

Important

• Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

Run configuration

• Source mode: Auto
• Repository: Unavailable
• Branch: Unavailable
• File: No configuration file from this repository applied
• Commit: Unavailable
• Layers: Organization UI
• Review profile: CHILL
• Plan: Advanced
• Run ID: 1e339545-0d52-49d5-9134-dc6e0e07f740

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

Share

• X
• Mastodon
• Reddit
• LinkedIn

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-11-02T08:51:09.185216Z	dbe4505	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: dbe4505767

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".