Skip to content

feat(cell-runtime): implement canonical LTX scaling plan - #216

Open
forhappy wants to merge 102 commits into
mainfrom
design/canonical-cell-ltx-scaling
Open

forhappy wants to merge 102 commits into
mainfrom
design/canonical-cell-ltx-scaling

Conversation

@forhappy

@forhappy forhappy commented Sep 18, 2026

Copy link
Copy Markdown
Contributor

Summary

Implements the canonical Cell LTX scaling track from crates/crab-cell-runtime/docs/canonical-ltx-scaling.md, with plans 004–017 kept self-contained and evidence-driven.

Included

  • Architecture guard for crab-http-server -> crab-cell-runtime -> crab-ltx, including a pure coordination-kernel gate.
  • Deterministic adversarial coordination simulation, replay corpus, broken-variant checks, and pinned TLC fast/negative/broad model checks.
  • Resident local routing before catalog/control reads and bounded sparse hydration on the existing SQL worker path.
  • Shared bounded native/bundle CellReplica publication paths, persistent verified directory cache, and fault-injection/cancellation cleanup coverage.
  • Actor-owned idle eviction, hysteretic pressure shedding, paced movement budgets, persisted Queue/Workflow inventory protection, and one runtime/SQL/hydration/primitive ledger.
  • Signed placement observations and deterministic weighted planning clamped by live runtime reservations.
  • OS-process movement probes for one authority winner, exact-root preservation, owner fencing/recovery, lost-release reconciliation, receiver rollback, and membership-loss ordering.
  • Schema-v3 qualification receipts, exact artifact binding, ten-row matrix manifests, path/symlink rejection, and fresh-process matrix verification.
  • Plan 016/017 hard removal of the standalone Replica, ReplicaHead, standalone paged/VFS, and CompactionSchedule surfaces. Cell-scoped mechanics remain; old prefixes are never reinterpreted.
  • Blob/Cron absolute-time windows anchored to mutation-issued time, plus Queue/Workflow durable recovery controls.

Verification

  • crab-cell-runtime: full locked suite passes (214 library tests, actor/integration suites, with only explicitly ignored provider cases).
  • crab-ltx: locked no-default and replica feature suites pass, including the canonical Cell root inventory qualification.
  • Process-support actor suite passes once without retry masking; the membership-loss simulator regression passes.
  • make -C crab architecture-check passes, including the standalone-LTX hard-cut guard.
  • Matrix receipt CLI smoke passes in a fresh process; Clippy, formatting, docs validation, and git diff --check pass.
  • PR checks: PR #216.

Remaining protected gates

The implementation does not fabricate release evidence. Protected provider/Kubernetes three-Pod fault receipts, matched Crab/Celld latency measurements, fleet-convergence/pressure receipts, and release-scale matrix rows still require their configured provider environments and exact tagged image. Local RustFS and multi-GiB runs are recorded as iteration evidence, not as substitutes for those protected gates.

@forhappy
forhappy force-pushed the design/canonical-cell-ltx-scaling branch from 16c4608 to aad779c Compare September 18, 2026 16:27
@forhappy
forhappy force-pushed the design/canonical-cell-ltx-scaling branch from 276d5d5 to ea897af Compare September 18, 2026 16:50
@forhappy

Copy link
Copy Markdown
Contributor Author

Follow-up pushed in d6684352d23 / cf8642a37fb / bda23353d24: active-Cell file descriptors are now part of the shared RAII ledger and eviction cost, surfaced via runtime stats and Prometheus gauges, with lifecycle/metrics assertions and a four-way concurrent reservation proof. Focused and full local gates remain green; external provider, multi-GiB, multi-process/Kubernetes, and standalone-contract decision gates remain intentionally open.

@forhappy

Copy link
Copy Markdown
Contributor Author

CI follow-up 6850d4c94c6 fixes actionlint failures by moving runner.temp Cargo target paths from job-level env (invalid context) to step-level env. Local actionlint now passes; the workflow will re-run the model and contract gates.

@forhappy

Copy link
Copy Markdown
Contributor Author

Follow-up f1d4f546727: coordination simulation movement release and release invariants now use the same ready_to_deactivate adapter-observation gate as production. Coordination simulator corpus (10 tests), docs validation, formatting, and diff checks pass.

Test added 30 commits September 18, 2026 17:08
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant