selfuse-batch

selfuse-batch (experimental, opt-in)

go run ./benchmarks/selfuse-batch runs one caller-frozen investigation plan (3..5 query/context/path-impact views) through a single corvint batch at one snapshot, or one standalone process per view (--mode standalone). It consumes only the accepted ordinary contract SBQ-V0-001..006 (decision 0052): it never sends possessed, refuses a plan that carries it, and defers AT-07 consumption. It never replaces the dogfood range impact --base phase. Output goes to a new directory outside the inspected repository, with fixed raw file names (batch.*, <route>-<ordinal>.*); the inspected repository is expected clean (--allow-dirty labels working bytes NOT_OBSERVED). Raw receipts are retained as private development evidence outside the public tree, not as a wire contract; corvint_processes, bytes and wall time are engineering measurements of Corvint processes only (Git children are uncounted), never token or cost evidence. Tests: go test ./benchmarks/selfuse-batch. The native tests use controlled Go executables and native process supervision.

Every usable packet is checked against the verb's emitted envelope (internal/contextindex/receipt.go for query/impact, taskcontext.go for context) and the tree the consumer captured itself; a batch answers at one snapshot (exact tree/commit), a standalone or fallback answer carries its own revision, and a drifted packet is unusable rather than silently accepted. On timeout, overflow or a signal the consumer stops corvint's descendants (Git children live in their own process groups) before the shared supervisor stops corvint. Audited cooperative AfterStart and BeforeStop hooks retain observed PID/start identities and join the bounded observer before returning. Each hook may add up to one second, separately from the two-second group cleanup budget. An exited leader or failed snapshot leaves descendant completeness unverified; the receipt never claims universal containment. Native RSS is recorded as max_rss_bytes. On a dirty tree (--allow-dirty) the mutation aggregate is null: working bytes are NOT_OBSERVED, HEAD and the .corvint listing (metadata, not content) and the ledger are reported separately.

The content-addressed artifact bundle retains the original raw/ paths, complete streams, failed/confounded runs, clean comparisons, historical source versions and independent reviews. Each artifact names a SHA-256 and byte length; blobs stores the corresponding base64 bytes. Every entry was decoded and hash-checked before packaging. The first intermediate repair's source snapshot was not captured; its output remains historical evidence, not a reproducible candidate qualification. Current source is the code alongside this document.

The historical Python-driver comparison returned four byte-equal packets: one Corvint process versus four, with 33,645 versus 33,385 raw output bytes. Its negative control retained the same failed operation on both routes. Recorded intervals exclude initial driver preparation/repository observation and final receipt serialization; setup, all workers, review, tokens and complete cost are not covered. The original dirty-tree comparison is confounded. No complete-task, token-cost or AT-07 claim.

Independent review required removing false possession, preserving failed/malformed responses, containing separate-group Git children on signal/timeout/overflow, and retaining exclusive wait4 ownership on raced exits. The historical focused run passed 47 tests, including 17 consumer tests and the existing measurement/first-CEM callers. Canonical verification and change binding are recorded separately; passing this experiment does not complete the owning roadmap evaluation.