plamentsv/write-error-divergence

Injectable Skill: Write Error Divergence

L1 trigger: STORAGE or DATABASE_TX flag OR file/DB write APIs detected (write_all, fs::write, rename, flush, commit, put, insert, batch, transaction) Inject Into: depth-state-trace, depth-edge-case Language: Go and Rust Finding prefix: [WED-N]

Purpose

Node clients often pair durable writes with in-memory metadata, indexes, caches, or database transactions. If metadata advances before the durable write succeeds, restart or reorg behavior can diverge from the actual persisted state.

1. Write Unit Enumeration

For each storage path, enumerate the logical write unit:

| Write Unit | Data Write | Metadata/Index Update | Error Check | Rollback/Cleanup | Verdict | |------------|------------|-----------------------|-------------|------------------|---------|

Include:

• block/header/transaction persistence;
• cache metadata, expiry timestamps, indices, checkpoints, migration markers;
• database transactions and batches;
• file writes that use temp files, rename, flush, fsync, or parent-directory fsync.

2. Required Checks

For each write unit:

• Metadata-before-data: metadata, cache state, or checkpoint markers must not advance before the durable write succeeds.
• Success-after-side-effect: Ok(()), HTTP success, or actor success messages must be returned only after the write and required flush/commit complete.
• Commit-before-check: database transactions must not commit when the closure or inner operation returned an error.
• Partial write: short writes, interrupted writes, rename failures, flush/fsync failures, and parent-directory fsync failures must be handled consistently.
• Rollback/cleanup: failed writes must either roll back metadata or leave an explicit repair/retry marker.
• Restart trace: trace what a node observes after restart if the write fails at each boundary.

Tag evidence as [WED-METADATA-BEFORE-DATA:{file}:{line}], [WED-COMMIT-BEFORE-CHECK:{file}:{line}], [WED-SUCCESS-BEFORE-WRITE:{file}:{line}], or [WED-RESTART-DIVERGE:{file}:{line}].

3. Non-Finding Rules

Do not report internal best-practice issues unless a failed write can produce stale reads, corrupted state, replay, permanent data loss, fork-choice divergence, or a liveness failure after restart.

4. Output

Use normal finding format. If no finding exists, still emit the write-unit table with file:line evidence for each safe rollback or commit path.