[wangtsiao]

Stack and review scope

Depends on #270 (Factory DAG scheduler) and #272 (RLM child lifecycle/collection). Neither prerequisite is merged by this PR. The stack base combines their published commits d2a2ca8 and 05495fc; the main-targeted diff therefore includes prerequisite changes until they land.

Review only this slice: wangtsiao/devo@stack/factory-base-270-272-20261108...feat/factory-durable-guards-20261108 . The incremental comparison is 18 files, two commits. Retarget/rebase after the prerequisites merge.

Added behavior

• Opt-in durable runs: host-owned, session-scoped save/load/list with 1 MiB snapshots, 128-run cap, CAS revisions, ownership checks, atomic replacement, private files and Unix directory fsync.
• Persist admission intent before spawn; checkpoint child binding after acknowledgement. Lost acknowledgements, missing/ambiguous child identities and failed cleanup remain explicit, never silently respawned or reported successful.
• Explicit restore first claims the revision, then reconciles exact direct children. Plan Mode denies writes before child materialization; read-only saved-list/load never creates files or hydrates children.
• Version 2 scalar guards and all/any joins, with deterministic all-predecessor settlement and distinct skipped branches. Guard numbers use a lossless range; no eval or authorization from output text.
• Stop/restore concurrency and malformed-checkpoint checks, including stop while saving intent, interrupted recovery, duplicate child bindings, inconsistent terminal flags and missing outputs.

This remains a foreground-driven DAG workflow. It does not implement automatic background scheduling, loops, foreach, residents, retry budgets, UI, or a machine library. See docs/factory-durable-workflows.md for the API and recovery limits.

Verification

• Factory Python tests: 42 passed, including an independent review pass.
• Entire Python runtime suite: 125 run, 122 passed, 3 platform-specific skips.
• Rust Factory store/dispatch tests: 11 passed, including real CPython kernel replacement, Plan gating, concurrent CAS, corruption, owner isolation and path-swap boundaries.
• Real CPython → production host → child-session integration: 2 passed. One executes a durable guarded diamond; the other rebuilds the runtime, reconciles a retained completed child, and dispatches only the pending successor. Providers are scripted/offline, not paid models.
• cargo clippy -p devo-server --lib --tests -- -D warnings: passed.
• cargo fmt --all -- --check and git diff --check: passed.
• Full cross-platform CI for this exact head: pending.

No PR merge, auto-merge or deployment is performed.