BACKLOG.md

Backlog

Five-slice upstream merge trial — 2026-10-03

• Validate and integrate the first five upstream PR milestones through 4d8ac9e1cea0d645a02f876f47536c54a5bfa406. Isolated merge source is reconciled: native trace representation, numeric daemon admission logging, EINTR-safe backoff, native help and registry selector filtering. Three branch-sync merges accompany five PR merges/five non-merges (13 ancestry commits total). Full Jenkins #54 passed at d3e2ac40 (7,528 tests, zero failures; TSan 671 passed). Main was fast-forwarded after validation. Final origin publication and branch receipts are recorded in run evidence final-promotion.json. The ancestry gap is 665, reduced by 13; the global semantic audit remains open. Canonical full-hash dispositions and acceptance are in the 308 audit's “Five-slice upstream ancestry trial” section; evidence is in the existing CBM-UPSTREAM-SOL-20261002-R2 run's five-slice-merge-trial/ directory.

Resolved

• Installer fork defaults and AST/watcher/macro/COUNT integration landed via PR #12; Windows/macOS repairs via PR #13 (bbabc073). PR #14 is merged on current main; historical release claims below remain revision-specific.

• Release-tooling implements deferred npm/PyPI/public publication, removes Claude automation, enables CodeRabbit, adds Linux-only prerelease selection. Package migration is deferred until exact fork assets and checksums exist; the previously merged package revisions are retained. PR #14 is merged; real artifact qualification remains open; these are implementation-complete tasks only.

• Reconciled completed upstream verticals: COUNT/status, persisted LSP decode (PR #4), Cypher semantics (PR #5), XML admission (PR #6), TypeScript await (PR #7), properties scanner cleanup, macro coverage, watcher backoff and all AST walker stacks are represented on main. See the canonical ledger for exact commits and remaining scheduler/allocation-failure evidence gaps.

• Removed stale upstream candidates: 77a0c7d9 empty-graph sort and 04ba2fa3/98899ba5 Cypher capacity checks already exist in current source.

• One-shot index metrics are implemented (0f5a84bc). Neutral operation/ daemon test source lists supersede the historical MCP link-time blocker.

• Resolve application-teardown ownership omission for neutral session state. cbm_daemon_application_free() now calls cbm_operation_session_state_free() before freeing each session. Valgrind identified the omission through test_daemon_application_free_releases_live_watch_once as 80 direct + 107 indirect lost bytes.

• Validated local release-tooling through Jenkins job #34 (full gate passed at 37dd63ba). PR #12 merged with installer fix c9841610 (install.sh and install.ps1 default to ngallodev/codebase-memory-cli). Resolved thread PRRT_kwDORjrg5c6n5KE9 2026-10-01.

• Validated complete AST slice 77ab4bea / 7707c9d7 in shared release-tooling Jenkins job #34 (full gate passed at 37dd63ba); all definition and auxiliary stacks ported to main.

• Added opt-in, single-run index metrics for performance diagnosis via index --metrics-out flag (0f5a84bc). Coverage in src/operations/index.c and tests/test_index_resilience.c. Merged via PR #12.

• Fixed the PR #12 Windows/macOS CI failures in bbabc073 (PR #13). The Windows hook-lifecycle test now uses the fork's installed binary name (tests/test_cli.c), and the macOS connection-cap test gets a 30 s server hello deadline (tests/test_daemon_runtime.c). Jenkins #35 and full GitHub CI passed. The pre-commit hook also clang-formatted both files.

Open

• Complete semantic reconciliation of all frozen upstream merges and constituent commits. Sol R2 inventories all 923 merges / 2,429 non-merges; 311 merges / 367 non-merges differ by ancestry. It reconciles the older 87 canonical rows, but 356 non-merge whole-commit remainders and all 311 merge dispositions remain unresolved. Census, patch ID and titles are not coverage. See the canonical source audit.

• Finish independent reconciliation of the Luna team evidence. All R3 handoffs were partial; all 40 declared checksums and exact 111/170/86 assignments were verified and archived. R4 successors continue interactively in managed terminal managed terminal / managed terminal / managed terminal. The 2026-10-02T23:53Z checkpoint contains 54 non-unresolved worker claims and 313 unresolved claims; no whole-commit coverage or completion is inferred. Parent retains 356 whole non-merge / 311 merge remainders pending exact source/test review. Full hashes are in the canonical ledger; the consolidated report records provenance, source corrections, scope and validation limits. Parallel split followup: CBM-RECON-ADAPTER-LOW-20261003 ran Luna low in managed terminal on 16 explicitly released, previously unreviewed runtime hashes; all 16 were source-reviewed, 8 closing as adapter-only and 8 retaining shared-semantic remainders. The consolidation auditor verified the run's 67-file manifest, confirmed every changed path is src/mcp/** or tests/test_mcp.c (that subtree exists upstream and is absent from production and both QA trees), and confirmed the 16 are upstream ancestors absent from those three CLI trees. After the Phase1 transfer, original runtime retains 39 hashes; all 367 hashes still have one effective primary owner. The transfer, acknowledgement, index and launch receipts are in the new run's evidence. Completed work was not reassigned. Phase1 successor plan now transfers 115 still-unreviewed runtime hashes to three prepared DeepSeek Flash assignments (39/26/50 disjoint groups), preserving the exact releases and completed work. Unstarted low2 was retired; corrected low1 finished partial with detailed handoff and its pane closed. See the canonical Phase1 plan. Native assignment delivery failed: all three returned predecessor reports, including after persisted steering and direct follow-ups. Sessions were stopped; zero assigned comparisons or correlated worker acknowledgements were established by that failed attempt. Interactive recovery now runs Codex --model deepseek-flash in managed terminal / managed terminal / managed terminal, with correct worker-issued run/cwd/count launch acknowledgements and actual process Jev-token inheritance verified. The same 115 hashes remain the comparison scope; see the canonical interactive recovery section. No completion/acceptance is inferred.

• Review additional bounded missing behaviors: malformed confidence parsing 62b44519c68b403ea44d9a85ba8ee74cddee66e6 (first additional localized candidate), route handler position 592894a4387a50e1d128ff6a2695ff48f8cc32a5, late annotation arguments c36b4fbc446f9085704a6073b81ec743fd4180fc, and client-registry selection 1e25d962d69999438229b0562d44ea60e97f5456. Group route followups with slice 5; review full tests/dependencies before broadening implementation scope.

• Resolve daemon memory policy before selecting active-job division 21591d51168d26f931f64d62edc36013cd1f6249. Fork uses fixed worker-capacity division; upstream active divisor is not represented. Review aggregate reservation invariants, separate Windows job-memory capability hunks from MCP adapter policy in merge 2b8bd0c065c1d2728994fddc1683a66f70745748.

• Daemon/cache/ReScript candidates and ADR alias/outline are integrated in local QA 559214af28ce2d6f5d6ec86662f6871316d28b45, per acknowledged parent steering and source comparison. Candidate implementation is separate from runtime validation, acceptance and production-main integration.

• Finish parent review and shared Jenkins validation for that QA revision and its later repairs. No current Jenkins success or acceptance is established by Sol R2. The pre-commit candidate retains formatting-only behavior; candidate ancestry does not establish production Git-child environment isolation.

• NEW upstream slice 1: Cypher OPTIONAL relationship fallback growth, 2fabaa2b877986bdededb4f4f78e9a31014c6fdf; existing append helper avoids the per-binding overflow. Jenkins cypher.

• NEW upstream slice 2: SQLite zero-cell trailing interior-page packing, e7324bb6904c6d0a295b956e83c8585219198358. Jenkins sqlite_writer,graph_buffer,pipeline.

• NEW upstream slice 3: Swift scanner shift widths, regression and integrity, 3f831936b5f3012d13175e7c97326165efecefad -> a2ea711c29560511eb8467bea7ad309ebf500d6e -> e8204092333ee1c38cd681a81abb36f125288894. Jenkins language,extraction,grammar_regression and vendored integrity; preserve QA's ReScript manifest/checksum row.

• NEW upstream slice 4: ancestor ignore chain with correct anchoring, c8184ed94887a6b5e130726885b44a07c63a47b7 -> 282fbfcdabfec79c19c36f5bec5f75c22eb36df5. Jenkins discover,gitignore,git_context,pipeline.

• NEW upstream slice 5: deterministic route selection and >64 proto services, 6b48097471e8d1a3e5cf16447304899a21e12d7a; adapt route hunks without importing upstream memory core or unrelated tests. Jenkins pipeline,parallel. Ranking is bounded by inspected source; unresolved work can change priorities.

• Evaluate GitHub ancestry reconciliation after semantic review. Adapted commits have different IDs; documentation cannot alter GitHub's behind count. An upstream ancestry merge requires explicit review of missing/excluded changes and merge-resolution behavior. Do not use an ours merge to conceal unresolved work, rewrite published history, or declare the fork globally caught up.

• Clarify CLI coverage array syntax and missing-path results. At dev CLI / project tmp-cbm-upstream-sol-20261002, generation 2026-10-02T17:30:32Z, JSON/comma --paths values became literal nonexistent paths with no_recorded_issue / freshness=missing. Repeated --paths / --scopes flags work. Document this spelling and make missing paths prominent.

• Align trace truncation with documented pagination. extract_func_def both-direction trace at limit 100 returned 127 total callees and truncation without a next cursor; limit 300 returned all 127 callees / four callers. Emit a cursor or document the limit restriction. Exact receipts are in Sol R2 evidence; these are observed CLI issues, with no source fix in this audit.

• Retain validation gaps: live watcher scheduler integration and AST allocation-failure injection are not established by completed helper tests.

• PR #14 at 1ce4e164 has failing GitHub diagnostic, macOS LSan, Unix/Windows test and shard-completeness checks (2026-10-02T04:14Z snapshot). Diagnose and pass required checks before merge/release; no release is qualified.

• Qualify native Windows/WSL and real fork package installations. Fork endpoints and a Linux-only prerelease route are implemented on release-tooling (1ce4e164), but implementation is not artifact qualification or publication.

• Selectively evaluate remaining upstream fixes from ../codebase-memory-mcp before merging into the CLI fork. Do not cherry-pick MCP-only changes or overwrite CP89/release-tooling work; apply compatible patches into a temporary worktree, run focused suites, then review the resulting diff before integration.

  • 9846c3f1 and c1b9c451 Windows daemon/session-root separator handling; port daemon-only hunks and omit MCP hunks.
  • 92725a5e Windows stdio parent-death cleanup; verify relevance to the CLI-only process model before porting.
  • e5c17de6 and b04f5450 coverage correctness/performance; manually reconcile against the fork's changed coverage implementation.
  • Exclude direct MCP-only adapter imports at the source seam; shared neutral capability parity still requires source verification. Historical exclusions are superseded by the canonical audit where mixed hunks were found.

• Assess a lightweight path for store-only test execution. Run just the store test binaries after shared prerequisites are built, while keeping the existing incremental suite selection and full-suite gates intact. First inspect the runner's configuration and targets to determine whether this is already supported. The cbm-upstream-count-20260929 focused run passed 96 store tests, but a cold worktree spent several minutes compiling shared prerequisites including lsp_all.c.

• Classify residual Heaptrack allocations in the neutral daemon test path.

  • Baseline evidence: Heaptrack reported 91 leaked allocations across the daemon application and IPC suites. Valgrind reported 538 bytes from cbm_daemon_ipc_endpoint_new() in the forked crash-simulation child of test_daemon_ipc_posix_current_generation_crash_cleanup_requires_startup_lock (src/daemon/ipc.c:889, test at tests/test_daemon_ipc.c:3735).
  • Remediation applied: the child now frees its inherited endpoint object before _exit(); it still bypasses listener_close, so kernel-released descriptors/locks and crash artifacts remain part of the test.
  • Current evidence: fresh scripts/analyze-memory.sh reports no static findings and Valgrind reports 0 definite, 0 indirect, and 0 possible lost bytes. Heaptrack reports 87 leaked allocations, which require attribution as still-reachable allocator/process-lifetime state before any further cleanup is attempted.
  • Next path: inspect the saved Heaptrack allocation call trees; only add teardown code for allocations proven to be owned by the neutral daemon rather than allocator or test-process lifetime state.

• Finish exact-revision Jenkins evaluation and independent review of the 2026-10-02 daemon/cache/ReScript integration candidate. Parent corrected the premature maintenance signal and ReScript child assertion; runtime claims, remote push and main integration remain pending. See the canonical 308 ledger.

• Reconcile the GPT-6.1-Sol full upstream merge/constituent audit CBM-UPSTREAM-SOL-20261002-R2, including ungrouped verticals and the next five recommended slices. Do not count running audit work as coverage.

• Complete DeepSeek Phase1 independent source review. All115 assignments and 342 evidence-manifest entries pass integrity checks; six bounded semantic questions were reviewed in three batched Jev requests. Documentary result is changes requested: narrow runtime 9724d903046626df0640ced218259cb58070e4e8 committed-counter/budget claim, preserve operations b04f54506ab42eddcbd89232a0d47d8355edd596 mechanism remainder and f6afd6138e605365a76d065d04521ec2037bed3d title-policy question, and classify integration 00411ce7bc41995783bfb16aada673515585446d changed Linux alias behavior as missing. Other109 hashes remain independently unreviewed; integration 29 partial rows plus1 unresolved row require function-level work. Exact hashes, requests, full distributions and source corrections are in the canonical308 review section and consolidated report; machine crosswalk stays handoff evidence. No completion/acceptance or semantic-closure counter advanced.

• Gather Luna low pilot and successor source evidence; independent review is recorded in canonical308. Remaining16 full-hash whole-commit reviews are tracked below; former launch-only20-unverified snapshot is superseded.

• Complete Luna medium runtime20 whole-commit source review: documentary changes requested; manifest31 passes, two bounded source dispositions verified, 18 exact full-hash remainders in parent evidence/medium20-independent-review/review.json. Group262ab012 telemetry with3507d955 Windows OS hard cap; no runtime/acceptance credit. Canonical308 finished-batch review owns corrections.

• Diagnose generic full-index pipeline failure for clean frozen worktree local run evidence; logs preserved in medium20 staging. Worker safely uses verified existing runtime graph; no daemon changes made.

• Review resumed DeepSeek pilot20 submission; four bounded source dispositions verified and16 retained unresolved (see following task). Worker completion is not acceptance.

• Resolve prior DeepSeek pilot20 independent review changes:16 exact full-hash whole-commit remainders in parent evidence/pilot20-independent-review/review.json. Correct false absent-alias/YAML/fd-guard/Codex claims, map missing CLI regressions, and preserve stale worker checksum failure; four source reviews credited, no runtime/acceptance claims. Canonical308 verification section owns corrections.

• Complete remaining17 DeepSeek ops20 whole-commit source reviews, exact hashes in parent evidence/ops20-followup-v4-review/review.json. Three source dispositions verified, including missing Scala route composition d4931241. Preserve56516eff recursive-grep fallback correction,26650255/f339aa62 dependencies, separate e5c17de6 unchanged cap context and canonical SQL/stale refusal policy obligations.

• Review followup-v2/v3 evidence and record their semantic limits; v3 hunk inventory is discovery only. Followup-v4 corrects wrong-function/test exclusion for d4931241; other17 full-hash semantic mappings remain pending above.

• Independently verify d4931241adaef732066afb5ebf1cb225e2474403 source comparison: all3 hunks missing across production and both frozen QA refs. Applicable shared Scala regression/harness verified; prerequisite041eb99c11692b93d43ea72c884cc50c51285db5 present. Future implementation/Jenkins/acceptance remain unperformed.

• Confirm ten historical merged/adapted tasks against current release-tooling 327dccc0 source and regression contracts: Go selector binding; daemon UI/Unicode; Cypher scope/star; registry receiver chains; cohort retry; Windows YAML guards; cache-root cohort handoff; complexity determinism; Razor; embedded sibling hosts. 10/10 source confirmed, selected remainder0. Canonical308 owns evidence/limits; no subtraction from678 ancestry or separate35 ops/runtime pending reviews. Native Windows qualification and fresh Jenkins51 result remain separate.

External Luna worker CLI cache resolution — 2026-10-05

• Investigate codebase-memory-cli: secure CLI coordination could not be created (cache-resolve) in prepared external managed terminal workers. Observed on cbm status --project ttm-jenkins-repair-20261005 --json / exact-project status/projects probes in Tax Machine runs ttm-jenkins-repair-20261005 and ttm-jenkins-verify-20261005; installed parent CLI reports dev. Both exact worktrees under local run evidence were fully indexed with persistence false, ready at 2026-10-05T14:27:42/43Z, source efa9b0a1971ec6d9a33457401ddeb337c1eeea3b, and unchanged empty Git porcelain. Parent status/projects succeed; worker calls fail before graph query despite launch args approval never / danger-full-access. Expected workers to access those ready graphs. Runtime/environment cause remains unproven. Workers preserve unchanged porcelain and use exact source/config fallback; issue remains open. Durable evidence: local run evidence and both runs' evidence/launch-preflight/. Do not reset the shared daemon or widen allow-roots as a workaround. Reproduced 2026-11-01 in AW-owned headless Luna runs cbm-sync-{security,ci,13a}-20261101 on clean isolated CLI worktrees at ff8b8906; all full persistence-disabled parent preflights were ready, with 63,168 nodes, 142 parse-partial files and no porcelain changes. Parent projects/status succeed; workers fail projects/status before query, despite danger-full-access in recorded launch. Exact runtime cause remains unproven. Receipts: each run owns evidence/launch-preflight/; worker failures are in output.log and executor-events.jsonl. Source fallback used; no shared daemon reset, cache-root switch or widened authorization.

• Isolate diff-scoped clang-tidy diagnostics outside selected hunks. In cbm-sync-13a-20261101, make -f Makefile.cbm lint-tidy-diff with clang-tidy-22 checked 26 staged production C files and failed with diagnostics including extract_defs.c:6518 outside the worker-reported staged ranges. Source checkpoint is upstream 5dc83ac9 over fork ff8b8906. Worker suspects relative line-filter paths versus absolute diagnostic paths; this cause is unverified and needs a minimal isolated reproduction. Do not reinterpret the failed command as green or suppress source findings. Raw evidence remains in that run's executor events.